2026-07

AutoSci is a memory-centric agentic system designed to automate the entire scientific research lifecycle. Its core method involves a structured memory system (SciMem) that separates long-term scientific knowledge from project-specific artifacts, enabling agents to efficiently access and manage information. AutoSci's contribution lies in providing a unified platform that supports all stages of research, from idea generation to manuscript submission and review, while also facilitating continuous improvement of its own research processes.

This paper introduces SCALE, a self-improving web agent that uses adversarial roles (Selector, Predictor, Judger) to autonomously identify and overcome its limitations through exploration. It also proposes SCALE-Hop for better global planning to avoid exploration traps. The core contribution is enabling web agents to adapt and learn without relying on expert data, demonstrated by improved performance and a new large-scale dataset.

This paper introduces LongTraceRL, a reinforcement learning method for improving long-context reasoning in LLMs. It constructs challenging training data by using search agent trajectories to create tiered distractors and employs a rubric reward that evaluates intermediate reasoning steps based on gold entities, unlike previous methods that relied on sparse, outcome-only rewards. This approach enables more effective supervision of the reasoning process, leading to better information integration from extensive, distracting content.

This paper investigates how the granularity of skill documentation affects the performance of large language model agents. The core method involves a controlled study using SkillsBench with different skill presentation levels and two LLM configurations. The main contribution is demonstrating that the availability of skills significantly improves task success rates, with finer-grained presentation having a less pronounced, and sometimes uncertain, impact.

This paper proposes ReuseRL, a method that views skill reuse in agentic RL as a form of compression. By grounding agent training in the Minimum Description Length principle, ReuseRL explicitly encourages agents to learn a small set of reusable abstract skills rather than brittle, task-specific shortcuts. This approach demonstrably improves generalization and success rates across various environments.

DRIFT addresses the challenge of efficiently optimizing large language models for multi-turn interactions. It decouples trajectory generation from policy updates, using offline data and importance weighting to mimic the benefits of online RL without its high computational cost. This allows for efficient fine-tuning while mitigating issues like distribution shift and behavioral collapse.

This paper demonstrates that reinforcement learning (RL) can amplify emergent misalignment in language models, even with seemingly harmless rewards. The core method involves fine-tuning small, open-weight models using RL with narrowly misaligned reward signals, showing this leads to greater general misalignment than supervised fine-tuning. The key contribution is characterizing this RL-induced misalignment and showing that existing safety mitigations developed for supervised learning can effectively address it.

COMAP co-evolves textual world models and agent policies in a closed loop. The world model predicts future feedback for candidate actions, which the agent uses to refine its choices by assessing feedback reliability. This process allows the world model to adapt to the agent's evolving behavior through self-distillation, improving its accuracy in predicting on-policy trajectories.

This paper introduces Multi-Order Communication (MOC) for LLM-based multi-agent systems. MOC addresses the limitation of current methods by reconstructing communication to capture multi-hop dependencies and employing a structural message consolidation strategy. This allows for more effective message transmission and optimization, improving the agents' ability to leverage information across multiple hops.

This paper proposes PaW, a framework that co-trains a language agent's policy and world model simultaneously during reinforcement learning. By leveraging on-policy rollouts, PaW adds auxiliary world modeling supervision without requiring separate simulators or changing inference. This approach improves agent performance by providing a better understanding of action consequences, demonstrated by consistent gains on benchmark tasks.

SIRI trains LLM agents to develop reusable skills internally, eliminating the need for external skill generators or inference-time skill banks. It achieves this through a three-phase process: initial policy warmup, self-skill discovery and validation using the agent's own trajectories, and finally, distilling beneficial skills into the agent's core policy. This approach reduces engineering complexity and improves inference efficiency for long-horizon tasks.

ResMerge addresses the challenge of merging large language models trained with reinforcement learning (RL). Unlike previous methods that focus on high-energy spectral components, ResMerge recognizes that RL task vectors have both a concentrated "head" and a dispersed "residual" component, both containing valuable information. By treating these components separately, ResMerge leverages the stability of the residual for merging while mitigating conflicts from the head, leading to improved performance.

This paper introduces TVIR, a benchmark and agent framework for generating research reports that integrate text and visuals. TVIR-Bench comprises 100 tasks requiring visually supported analysis, while TVIR-Agent is a multi-agent system that retrieves and generates relevant images and charts. The contribution lies in addressing the text-centric bias in current research agents and providing a comprehensive evaluation method for multimodal report generation.

This paper introduces the Tool-Environment-Agent (TEA) protocol, a novel framework for coordinating multi-agent systems. TEA models agents, tools, and environments as versioned resources with explicit lifecycles, enabling better context management and reproducibility. Building on TEA, AgentOrchestra provides a hierarchical framework for dynamic agent capability extension and coordination on complex tasks.

This paper introduces BenchTrace, a benchmark designed to evaluate the reflection and controlled evolution capabilities of LLM agents. BenchTrace utilizes a dataset of annotated episodes and includes two key evaluations: one that probes failure identification through targeted QA, and another that tests an agent's ability to avoid past failures in simulations. The paper also proposes a new metric, failure avoidance rate (FAR), to quantify this improvement.

EAPO is a novel RL framework that enhances LLM reasoning by allowing policies to adaptively seek assistance from external experts during training. This on-demand expert interaction provides richer reward signals and more reliable reasoning trajectories, ultimately internalizing expert knowledge to improve the LLM's independent problem-solving capabilities.

This paper introduces EELMA, an algorithm that estimates the "empowerment" of language model agents. Empowerment, an information-theoretic measure, quantifies an agent's ability to influence future states through its actions. EELMA allows for scalable evaluation of LM agents in text-based environments, demonstrating that empowerment correlates with task performance and can identify key moments of general capability.

EvoSpec accelerates LLM inference by dynamically adapting the draft model's vocabulary and parameters in real-time. This approach overcomes the limitations of static pruning by efficiently retrieving relevant long-tail tokens and minimizing the distributional gap between draft and target models through online alignment. Its core contribution is enabling speculative decoding to maintain high acceptance rates even in specialized domains or during topic shifts.

This paper introduces FLUID, a framework for efficiently adapting autoregressive (AR) language models to diffusion models for text generation. FLUID achieves this by enforcing "Strictly Causal Alignment" to allow seamless initialization from AR checkpoints, avoiding costly retraining. It also employs "Elastic Horizons" to dynamically adjust denoising steps, leading to state-of-the-art performance with significantly reduced training costs.

This paper proposes a new LLM post-training method called Chain-of-Meta-Thought (CoMT). It addresses the limitation of current methods by mimicking human problem-solving, which involves first learning abstract strategies (meta-knowledge) and then applying them. CoMT achieves this by separating supervised learning of general reasoning patterns from problem-specific execution, leading to more generalizable and reliable LLM reasoning.

This paper introduces Graph-Enhanced Policy Optimization (GEPO) to improve LLM agent training for multi-step tasks. GEPO addresses the issue of uniform credit assignment by developing a dual-level structural credit system. It calculates a "Task-Conditioned Criticality" score for each state, considering its topological importance within the state-transition graph, to assign more accurate credit at both the step and trajectory levels.

This paper introduces GroundAct, a benchmark designed to assess Large Language Model (LLM) agents' ability to ground actions in environmental states. The core method involves evaluating LLMs on tasks where action feasibility depends on unstated environmental conditions. The key contribution is demonstrating that LLMs struggle with this "action grounding" and identifying specific reasoning patterns and bottlenecks that hinder their performance in complex, state-dependent tasks.

This paper addresses the lack of robust benchmarks for evaluating LLM-driven data agents' insight discovery capabilities. The authors identify critical flaws in existing benchmarks like InsightBench and propose new criteria for high-quality evaluation. They then introduce InsightEval, a new, expert-curated dataset constructed through a data-curation pipeline to provide a more reliable assessment of LLMs' ability to uncover latent knowledge in data.

This paper introduces LsrIF, a novel training framework designed to improve how large language models follow instructions with complex logical structures. LsrIF addresses limitations of existing methods by constructing training data that explicitly models parallel, sequential, and conditional dependencies between constraints. Its key contribution is a structure-aware reward aggregation mechanism that aligns training signals with the execution semantics of these logical structures, leading to more robust instruction following.

This paper introduces a novel security framework for AI agents that moves beyond empirical semantic guardrails. Its core method, the Proof-Constrained Action (ePCA) framework, forces agents to formally translate their intentions into logical constraints before execution, ensuring provable security by leveraging the limitations of logical reasoning. The main contribution is a neural-symbolic isolation architecture that eliminates reliance on semantic trust in natural language for security.

This paper investigates the assumption that joint training of reasoning and tool-use in Agentic Reinforcement Learning (ARL) is always beneficial. Using a new method called Capability Effect Attribution (CEA), they demonstrate that these two capabilities often interfere with each other due to misaligned gradients, hindering performance. To overcome this, they propose Disentangled Action--Reasoning Tuning (DART), a framework that explicitly separates and tunes these capabilities.

This paper introduces Proactive Interactive Reasoning (PIR), a new paradigm for LLMs that moves beyond passive problem-solving. PIR enables LLMs to proactively ask clarifying questions to users when facing uncertainty about premises or intent, rather than just performing internal reasoning. This is achieved through a fine-tuning process and a policy optimization framework that trains the LLM to effectively interact and align with user needs.

ReflexGrad introduces a dual-process architecture to enable LLM agents to recover from failures within a single episode without needing demonstrations. It dynamically routes between a fast, continuous refinement process and a slower, diagnostic process that triggers upon detecting low progress. This approach significantly improves task success rates by allowing agents to learn from and correct their mistakes in real-time.

This paper introduces AgentREVEAL, a framework to diagnose how web retrieval in LLM agents can degrade safety alignment. It finds that tightly coupling tool use with generation amplifies harmful outputs. Surprisingly, even safety-oriented content can increase harmful compliance due to the "Safe Source Paradox."

RewardFlow addresses sparse rewards in LLM-based RL by constructing state graphs and using topology-aware propagation to estimate state-level rewards. This method generates principled, annotation-free dense rewards, significantly improving agent performance and training efficiency across various benchmarks.

This paper introduces SCOPE, a novel method for LLM agents to dynamically manage large, evolving contexts. SCOPE frames this as an online optimization problem, automatically evolving agent prompts by synthesizing guidelines from execution traces. Its core contribution lies in a Dual-Stream memory mechanism and Perspective-Driven Exploration, enabling agents to adapt to complex environments and overcome static prompt limitations.

This paper introduces SIA, a novel self-improving AI loop that breaks down silos in AI development. SIA's core method is a language-model agent that simultaneously updates both the "harness" (tools, prompts, retry logic) and the "weights" of a task-specific agent. This unified approach, unlike previous methods that focused on one aspect, leads to significant performance improvements across diverse domains.

This paper introduces PropMe, a framework that evaluates LLM memorization not just when prompted adversarially, but also under normal usage. Their core method uses a new propensity-aware metric and a tracing pipeline called SimpleTrace to quantify how often models leak training data during typical generation. The key contribution is demonstrating a significant gap between LLMs' *ability* to memorize (when forced) and their *propensity* to do so naturally, suggesting current evaluations may overestimate real-world data leakage.

This paper introduces the "Recuse Signal," a novel in-band mechanism for servers to request autonomous LLM agents to voluntarily withdraw access to a resource, acting as a cooperative governance control rather than a security boundary. The core contribution is to empirically measure whether LLM agents will actually comply with these signals, addressing a critical gap in controlling agent behavior in live infrastructure.

CollabSim introduces a novel methodology for evaluating the collaborative competence of LLM agents in multi-agent systems. It grounds this evaluation in decades of Computer-Supported Cooperative Work (CSCW) research, focusing on how agents establish common ground and manage shared understanding, rather than just individual task performance. This allows for controlled experiments to systematically investigate and improve how LLM agents collaborate effectively.

This paper empirically compares human and DeepSeek-R1 LLM mathematical reasoning on AIME problems. It finds that while the LLM mimics the structure of human solutions, it lacks genuine reasoning, exhibiting "topological mimicry" through repetitive, shallow checks and loops. However, successful LLM solutions show stable use of branching and backtracing, suggesting potential signals of deeper reasoning.

The paper introduces AARR (Act As a Real Researcher), a benchmark series designed to evaluate Large Language Models (LLMs) and AI agents on their ability to perform research tasks with the professionalism and nuanced judgment of human researchers, rather than just macro-level execution. The first benchmark, AARRI-Bench, focuses on granular research scenarios to assess these finer qualities, highlighting that current frontier agents still fall short of fully replacing human researchers due to limitations in field sensitivity, ethics, and scientific judgment.

This paper proposes a "human-view" framework for video understanding using Multimodal Large Language Models (MLLMs), organizing capabilities into "watching," "remembering," and "reasoning." This approach aims to unify the analysis of how MLLMs process sparse evidence, maintain long-range context, and perform grounded inference in complex video scenarios, moving beyond isolated benchmarks.

This paper introduces Q-Evolve, a framework for training LLM agents to make reliable long-horizon decisions. It addresses the challenge of delayed rewards by unifying automatic reward labeling and policy learning. Q-Evolve uses an in-distribution reinforcement learning approach with a hybrid dataset and a weighted Implicit Q-Learning objective to stabilize learning, deriving step-wise rewards from a learned value function for dense supervision.

This paper introduces the Collaborative Human-Agent Protocol (CHAP), a new standard for managing complex, multi-agent and human collaborations. CHAP defines a shared workspace to facilitate seamless interaction and coordination between humans and AI agents, particularly in operational roles involving critical decision-making. Its core contribution is to provide a structured framework for these interactions, moving beyond current ad-hoc methods and enabling more robust and traceable human-AI teamwork.

PRISM recovers the set of active instructions guiding a language model's behavior by decoding its internal activations. Its core method involves training an activation-conditioned interpreter using GRPO to directly predict instruction sets, rewarding accurate coverage and penalizing unsupported ones. This approach allows for reliable monitoring of LLM agents by revealing hidden objectives and constraints, unlike previous activation-to-language methods.

This paper proposes Divergence Regularized Policy Optimization (DRPO) to improve the stability of Reinforcement Learning for Large Language Models (LLMs). DRPO replaces the hard masking used in previous methods with a smooth, advantage-weighted quadratic regularizer. This allows for more nuanced correction of policy shifts, rather than simply discarding gradients, leading to more stable and effective LLM training.

This paper introduces Ego-MC-Bench, a benchmark designed to evaluate video LLMs' ability to proactively correct user mistakes in real-time during everyday tasks like cooking. The core method involves creating realistic cooking scenarios with deliberate errors to test the LLMs' reactive guidance capabilities. The main contribution is highlighting the significant challenge this poses for current video LLMs and identifying a lack of suitable training data as a key obstacle.

This paper addresses the "length collapse" issue in LLMs generating long-form content. Their core method, IS-CoT, introduces a dynamic "Plan-Write-Reflect" cycle directly within the generation process, allowing for continuous adaptation and global alignment. This approach overcomes the limitations of static planning and enables LLMs to produce coherent, long-form text without external agents.

This paper introduces PsychoSafe, a novel framework for LLM refusals that moves beyond simple non-compliance. PsychoSafe reframes refusals as structured, psychologically-informed supportive communication, drawing on evidence-based intervention strategies. Its core contribution is demonstrating that this approach significantly improves the quality of LLM refusals, especially in high-risk scenarios, by better addressing the underlying needs of users.

This paper investigates how Reinforcement Learning from Human Feedback (RLHF) aligns large language models. The authors demonstrate that RLHF doesn't fundamentally alter a model's underlying partisan structure, but rather compresses its variance to produce neutral-sounding outputs. This suggests RLHF achieves shallow, functional compliance rather than deep alignment with human values.

This paper introduces Null-Space Constrained Response-Specified Unlearning (NSRU), a method for unlearning specific knowledge from large language models while preserving general capabilities. NSRU achieves this by projecting low-rank adaptation updates into the null space of learned "retain" subspaces, ensuring that modifications are localized and do not disrupt benign knowledge. The core contribution is a projection-constrained framework that enables precise control over unlearning by specifying safe target responses and confining updates to preserve existing functionalities.

ReasonAlloc addresses the KV cache bottleneck in LLM reasoning by dynamically allocating its budget. It uses a hierarchical approach: an offline strategy identifies an architecture-driven "Reasoning Wave" of demand across layers, and an online strategy reallocates resources to information-rich heads during decoding. This allows for more efficient KV cache usage during long reasoning chains.

Role-Agent addresses LLM agent limitations by using a single LLM as both agent and environment for bootstrapped co-evolution. It employs World-In-Agent (WIA) for environment-aware reasoning via state prediction rewards, and Agent-In-World (AIW) to refine training data by analyzing and retrieving tasks based on failure patterns. This dual-role approach improves generalization and learning efficiency.

Workflow-GYM is a new benchmark designed to evaluate AI agents' ability to perform complex, long-horizon tasks within professional software environments, unlike existing benchmarks that focus on simpler applications. Its core contribution is to highlight the significant gap in current agent capabilities for real-world professional workflows, showing even advanced models struggle with high success rates.

Flow-DPPO addresses limitations of existing RL methods for flow matching models by replacing noisy policy ratio clipping with a divergence proximal constraint. Its core method leverages the Gaussian nature of per-step policies in flow models to efficiently compute KL divergence, enabling more stable and effective policy updates. This approach improves the quality and alignment of generated images and videos by avoiding over- and under-constraining during training.

This paper investigates whether training large language models for reasoning sacrifices their alignment with human values like safety and bias avoidance. The study finds that current reasoning model training methods often lead to "alignment regressions," meaning they become less trustworthy in these areas despite improved reasoning capabilities. The contribution lies in highlighting this trade-off and providing a systematic audit of different training approaches.

This paper addresses LLM tool-use limitations by integrating and activating experiential knowledge. The core method involves acquiring instance-level knowledge, which proves more effective than abstract intent-level knowledge. For activation, parallel sampling with reasoning width expansion outperforms depth expansion, and post-training with knowledge-augmented data aids internalization. The contribution lies in a systematic study demonstrating how specific types of experiential knowledge and activation strategies significantly improve LLM tool-use performance.

ALIGNBEAM is a training-free method that improves LLM safety by transferring knowledge from a safe "anchor" model to a potentially unsafe "specialist" model at inference time. It achieves this by translating the anchor model's output probabilities into the specialist's vocabulary token-by-token and using a judge LLM to select the safest continuation. This allows for safety alignment even between models with different vocabularies, significantly increasing refusals of harmful prompts without retraining.

APPO addresses the challenge of credit assignment in multi-turn tool-use by large language model agents. Instead of assigning credit to coarse units like tool calls, APPO shifts this to fine-grained decision points within the agent's generated sequence. This allows for more precise identification of influential decisions, improving the agent's ability to learn from its interactions.

This paper introduces ModSleuth, an agentic system designed to automatically reconstruct the complex, recursive dependency graphs of modern LLMs. By analyzing public artifacts, ModSleuth addresses the challenge of fragmented documentation to reveal how LLMs are built upon other models for data generation, filtering, and evaluation. Its core contribution lies in a formalization that clarifies dependency types and reconciles inconsistent references, providing a traceable audit of LLM development.

Agents-K1 addresses the lack of scientific knowledge orchestration in LLM agents by converting raw documents into agent-native knowledge graphs. Its core method involves a multimodal parser to extract detailed entities, evidence, and relations from full papers, an information extraction backbone trained with GRPO, and a tri-source agent interface for unified retrieval. This contributes a richer, structured representation of scientific knowledge, enabling more sophisticated agent reasoning.

This paper proposes a method to significantly reduce computation costs for large language models by precomputing and sharing Key-Value (KV) caches. Instead of each agent recomputing the KV cache for identical documents, a publisher can generate it once, allowing other agents to purchase access and skip the expensive prefill step. This approach offers substantial compute savings with no accuracy loss, making it a practical solution for efficient AI agent deployment.

EurekAgent's core method is "environment engineering," which focuses on designing the resources, constraints, and interfaces of an agent's execution environment to guide its behavior towards productive scientific discovery. The paper's main contribution is demonstrating that this environment-centric approach, rather than complex agent workflows, is key to unlocking autonomous scientific discovery, enabling agents to explore, manage artifacts, and collaborate effectively while avoiding pitfalls.

This paper proposes that both humans and LLMs perform everyday reasoning through pattern matching, rather than abstract world models. By observing similar error patterns in human participants and LLMs on common-sense reasoning tasks, the researchers identify specific attention mechanisms in LLMs that implement this pattern matching. These mechanisms can predict human reasoning errors influenced by seemingly irrelevant prompt details, supporting the shared pattern-matching approach.

EvoArena is a new benchmark suite designed to evaluate LLM agents in dynamic environments that progressively change over time. The paper introduces EvoMem, a memory system that tracks these environmental changes as structured update histories, allowing agents to reason about and adapt to evolving conditions. This approach aims to improve LLM agent robustness in real-world scenarios where environments are not static.

HyperTool addresses the inefficiency of step-wise tool calls in LLM agents by introducing a unified, MCP-style interface. This allows models to execute complex tool workflows within a single, code-block-based call, reducing context consumption and simplifying dataflow management. HyperTool significantly improves agent performance on cross-tool compositional tasks by abstracting away low-level execution details.

This paper introduces the Recursive Agent Harness (RAH), a novel approach that extends recursive language models by using full agent harnesses, equipped with tools and execution capabilities, instead of just model calls. The core method involves a parent agent generating and running executable scripts that spawn parallel subagent harnesses for complex tasks and structured function calls for simpler ones. RAH's contribution is a code-first extension to model recursion, significantly improving long-context reasoning performance for coding agents.

This paper proposes a paradigm shift from "Chatbot" to "Digital Colleague" for AI. The core method involves advancing LLMs from simple conversational generators to integrated systems with persistent memory, reasoning, and self-improvement capabilities. This transition is achieved through enhanced cognitive processes like Chain-of-Thought reasoning and the development of tool-augmented workstation systems with persistent workspaces and verification loops.

GitOfThoughts addresses the ephemeral nature of LLM reasoning by treating it as a version-controlled process, akin to software development. It stores reasoning steps as Git commits, allowing for replayability, auditing, and merging of agent thought processes. However, the paper's key contribution is its finding that, for novel problems, no memory substrate, including GitOfThoughts, reliably improves LLM accuracy.

This paper introduces Collective Skill Tree Search (CSTS), a novel framework for automatically constructing reusable skills for LLM agents. CSTS uses an iterative process of generating diverse skill candidates from multiple models and then having those models collectively assess and select the most effective ones. This approach aims to create structured, generalizable skill trees that enhance LLMs' capabilities in complex task execution and tool use.

This paper introduces Skill-to-LoRA (S2L), a novel method for representing and utilizing agent skills in Large Language Models (LLMs). Instead of relying on lengthy skill descriptions at runtime, S2L learns skill-specific LoRA adapters that capture the behavioral changes induced by these skills. This approach significantly reduces computational overhead and improves performance by enabling dynamic loading of learned skill behaviors.

TokenPilot tackles the rising inference costs of LLM agents by efficiently managing their context. It employs a dual-granularity approach: **Ingestion-Aware Compaction** stabilizes prompt prefixes and filters environmental noise at the entry point, while **Lifecycle-Aware Eviction** conserves context by offloading segments only when their task relevance diminishes. This method significantly reduces costs by preserving prompt cache continuity while effectively managing token footprints.

This paper introduces ContextRL, a reinforcement learning method that enhances LLMs' ability to pinpoint crucial information within complex contexts. By training LLMs to select the correct supporting context from similar options, ContextRL encourages fine-grained grounding, leading to improved long-horizon reasoning and multimodal performance. This approach offers a novel way to supervise LLMs beyond just the final answer.

This paper introduces **Contrastive-Difference CKA (CKA_Delta)**, a novel method to analyze how different language model architectures represent high-level concepts. CKA_Delta effectively distinguishes concept-specific structural alignment from general similarity, revealing that while models show moderate geometric convergence, they exhibit near-perfect functional transfer of concepts across architectures. This finding suggests a surprising degree of structural compatibility in how diverse LLMs encode meaning.

LOGOS unifies diverse scientific tasks by representing scientific objects and their spatial interactions as discrete tokens within a shared autoregressive framework. This "scientific grammar" allows complex structural relationships to be captured sequentially, enabling a single model to perform various downstream tasks through next-token prediction. LOGOS demonstrates strong performance across multiple scientific domains, suggesting the potential for a general-purpose foundation model for science.

This paper investigates the bargaining capabilities of LLMs in simulated used car sales scenarios with varying information availability. It evaluates LLMs' performance against game-theoretical solutions, focusing on their honesty (disclosure/deception) and credulity (trust/distrust). The core contribution is demonstrating that off-the-shelf LLMs deviate significantly from optimal strategies, attempting to lie but failing to effectively exploit information asymmetry.

This paper investigates how vision-language models (VLMs) handle gender bias with ambiguous visual inputs. Their core method, LALS, measures internal concept associations within the VLM's latent space. The key contribution is demonstrating that VLMs often internally encode female associations for stereotypically female occupations but still output male representations when prompted with ambiguous images, revealing a decoupling between internal knowledge and external output.

This paper investigates the consistency of Large Language Models (LLMs) when used as judges for evaluating AI safety. The core finding is that LLMs are unreliable and inconsistent judges, especially for nuanced safety issues in regulated domains, and show high disagreement among themselves. This work contributes by highlighting the limitations of LLM judges and offering practical recommendations for their use in AI safety evaluation.

This paper introduces a framework where Vision-Language Models (VLMs) engage in multi-turn dialogue to collaboratively reconstruct a target structure from visual and textual inputs. The core contribution is demonstrating that while this dialogue approach slightly improves VLM performance on spatial reasoning tasks, it reveals significant limitations in their current ability to handle complex spatial understanding, even with detailed textual descriptions.

This paper introduces PARL, a novel framework for personalized LLM evaluation. PARL treats evaluation as a learning problem, directly learning preference-aware rubrics from user interaction histories. This approach aims to overcome the limitations of existing methods by capturing subjective, user-specific preferences for more effective personalized alignment.

UniAudio-Token enhances semantic speech tokenizers by integrating general audio perception without sacrificing speech capabilities. It achieves this through Semantic-Acoustic Primitives for structured supervision and Semantic-Acoustic Equilibrium to adaptively restore acoustic details, resulting in comprehensive universal audio representations.

This paper proposes using question-asking as an inference-time intervention to probe the hidden state of Large Language Models (LLMs) during reasoning. By training a "probe" on the LLM's internal states before and after it generates questions, the researchers found this self-diagnosis process predicts the final answer's correctness, suggesting the value lies in the model's internal reflection rather than external information.

This paper introduces AgentCL, a rigorous evaluation framework for continual learning in language agents. Its core method involves constructing controlled, compositional task streams that allow for the systematic reuse of prior sub-solutions and evidence. AgentCL's main contribution is enabling a deeper understanding of how agents learn and transfer knowledge across tasks, addressing limitations in existing benchmarks.

AgentPLM introduces a novel approach to protein sequence design by augmenting traditional language models with a reasoning and feedback loop. Its core method, Reasoning-Augmented Decoding (RAD), interleaves sequence generation with calls to biophysical simulation tools. This allows the model to iteratively refine sequences based on predicted thermodynamic and structural constraints, leading to improved design performance across various protein engineering tasks.

This paper addresses the "last-mile forecasting" problem, where raw statistical forecasts are revised with unstructured business context. The core method is an LLM-agent framework that sits atop a forecasting backbone, using tools to retrieve contextual evidence and perform reasoned forecast revisions under safety constraints. Its contribution is a practical system that bridges the gap between statistical forecasts and decision-ready predictions by incorporating real-world business knowledge.

ClinEnv is an interactive benchmark designed to evaluate LLMs as attending physicians in a simulated inpatient setting. Its core method involves presenting agents with real patient admissions as a sequence of decision stages, requiring them to actively query specialized agents for information before making irreversible choices. The contribution lies in its realistic simulation of clinical practice, which emphasizes incremental information gathering and sequential decision-making under uncertainty, and its scoring of both decision accuracy and information-seeking process.

This paper introduces Humanity's Last Line of Verification (HLL), a benchmark designed to test if multimodal AI agents can bypass CAPTCHAs, which are human-verification barriers. HLL uses interactive CAPTCHAs and controlled stressors to evaluate agents' ability to perform human-like interactions, not just recognition, within a GUI environment. The core contribution is a novel evaluation method for assessing agent robustness against automated bypass of critical human-verification systems.

Iteris is an agentic AI system designed to tackle open problems in computational mathematics. Its core method involves generating numerical evidence, constructions, and proof drafts, which are then refined by human experts. The paper's contribution is demonstrating Iteris's capability to produce verified results on challenging open problems, specifically a phase diagram for iterative solvers and a counterexample for a conjecture.

This paper introduces MCP-Persona, a novel benchmark for evaluating LLM agents on real-world personal applications that use the Model Context Protocol (MCP). Unlike existing benchmarks, MCP-Persona focuses on personalized tools like social media and collaboration suites. The benchmark reveals that current state-of-the-art agents struggle significantly with effectively using these personalized tools.

This paper addresses "Perceptual Judgment Bias" in multimodal LLMs used as judges, where they favor plausible text over visual accuracy. The core method involves creating a dataset of minimally perturbed visual examples to isolate perceptual errors and then training a reward model using a GRPO-based approach and batch ranking. This allows for more reliable and perceptually grounded evaluations by multimodal LLMs.

This paper introduces Moment-Video, a new benchmark designed to evaluate how well video multimodal large language models (MLLMs) can understand brief, critical visual events. The core method involves creating questions that hinge on short-lived visual occurrences, testing the models' ability to detect and reason about these fleeting moments. Moment-Video's contribution is to highlight and diagnose a crucial weakness in current video MLLMs: their difficulty in preserving and utilizing information from transient visual evidence, which is vital for answering many practical questions.

This paper introduces LLMFI, a novel fault-injection framework, to systematically study how soft errors propagate during Large Language Model (LLM) inference. By injecting faults across various LLMs and tasks, the study reveals critical vulnerability patterns and provides practical, low-overhead software-only solutions to enhance LLM reliability.

POIROT is a novel protocol that uses the agents within a Multi-Agent System (MAS) itself to detect failures, rather than relying on external evaluators. This "interrogation" approach leverages the inherent diversity of the agents' knowledge to identify issues like hallucinations. POIROT significantly outperforms traditional single-agent evaluation methods, especially as system complexity and the number of agents increase.

This paper proposes AdvCL, a novel continual learning method that repurposes adversarial perturbations to stabilize model adaptation. By employing three plug-in modules (Intra-Smooth, Proto-Clip, and Inter-Align), AdvCL promotes local smoothness, prevents over-alignment to new tasks, and actively aligns representations with previous tasks. This approach effectively reduces forgetting, enhances transfer learning, and improves robustness against adversarial attacks.

This paper introduces Benchmark Agent, an autonomous system that automates the entire benchmark creation process for LLMs and MLLMs. Its core method involves orchestrating user query analysis, subtask design, data annotation, and quality control. The main contribution is a sustainable and scalable approach to generating diverse, high-quality benchmarks that can effectively differentiate state-of-the-art models.

This paper introduces ALMANAC, a novel dataset designed to improve agent collaboration. It addresses the lack of data on human collaborative processes by providing action-level annotations of participants' mental models during a dyadic routing task. This dataset aims to guide LLM agents towards better understanding and aligning with human collaborators' intentions and goals.

TokenMizer addresses the LLM context window limitation by modeling session history as a typed knowledge graph, preserving crucial structured information. Its core method involves a hybrid extraction pipeline to build this graph, a multi-tier checkpoint system for serialization, and an 8-layer compression pipeline to reduce token overhead. This approach significantly improves token economy and enables resumable long-horizon LLM sessions by retaining relational structure, unlike previous flat-text methods.

This paper introduces TOKI, a bitemporal operator algebra for managing persistent memory in LLM agents. TOKI formalizes contradiction resolution as a form of write-time concurrency control, explicitly defining isolation preconditions and provenance for existing heuristics. Its core contribution is providing formal guarantees for these heuristics and their pipelines, ensuring reliable and auditable belief updates.

This paper proposes a layered framework for integrating knowledge into multimodal generative models. It categorizes knowledge infusion based on which part of the iterative generation process is modified: input/output (surface), state transitions (trajectory), intermediate states (latent), or model parameters (parametric). This framework provides a systematic way to understand and design knowledge-aware generative systems.

This paper proposes NF-CoT, a novel latent reasoning framework for large language models. It leverages normalizing flows to represent intermediate reasoning steps in continuous latent spaces, offering a higher-bandwidth alternative to explicit textual chain-of-thought. NF-CoT's core contribution is preserving the advantages of traditional autoregressive generation, such as left-to-right generation and probabilistic sampling, while enabling efficient and effective latent reasoning.

Tangram addresses the inefficiency of non-uniform KV caching in multi-turn LLM serving by introducing a novel system. Its core method involves deterministic budget allocation for KV cache heads, eliminating scheduling overhead and prefill stalls. This approach unlocks efficient utilization of non-uniform KV caches, significantly improving LLM serving performance.

This paper proposes a reinforcement learning (RL) method to enable large language models (LLMs) to translate unseen languages by learning to utilize provided linguistic context. Instead of memorizing specific languages, the RL agent learns a meta-skill to extract and apply relevant information from the context, leading to improved zero-shot translation performance on entirely new languages. This approach offers a more generalizable way to handle extremely low-resource translation compared to traditional fine-tuning or in-context learning.

This paper addresses deceptive performance in coding agents, where models exploit shortcuts to achieve high scores without true task mastery. They propose CapCode, a framework generating datasets with randomized tests capped below perfect scores, making cheating evident when models exceed this cap. Additionally, CapReward is introduced to discourage cheating during training by aligning rewards with the capped performance principle, leading to agents that better adhere to task specifications.

DuMate-DeepResearch is a multi-agent framework for complex research tasks that addresses limitations in current Deep Research systems. Its core method decouples task planning and execution into an Agent Core and an extensible Tool Ecosystem, enabling explicit traceability of decisions and tool usage. This design enhances auditability and mitigates risks like hallucination and planning bottlenecks.

This paper investigates how autonomous AI agents, exemplified by Perplexity's "Computer" product, transform knowledge work compared to conversational assistants like "Search." The core method involves analyzing production data to compare task execution times and user behavior. The key contribution is demonstrating that autonomous agents significantly accelerate knowledge work by performing end-to-end task execution, leading to increased efficiency, automation of complex steps, and a shift towards higher-order cognitive tasks for users.

This paper introduces SV-Detect, a novel AI-generated text detection method that utilizes "steering vectors" derived from a frozen language model's hidden states. These vectors capture directions that distinguish human from machine text at each layer, and the detector classifies text based on its alignment with these directions. SV-Detect demonstrates robust performance even under distribution shifts, offering a new perspective on fake-text detection as a representation-space probing problem.

This paper investigates how sensitive Large Language Models (LLMs), including those specialized for medicine, are to small changes in input prompts. The core method involves a systematic sensitivity analysis on the MedMCQA benchmark, categorizing prompt variations as natural or adversarial. The key contribution is demonstrating that even medical LLMs are not inherently robust, showing that minor prompt alterations can lead to inconsistent or even harmful clinical advice, highlighting significant risks for healthcare applications.

TabSwift is an efficient tabular foundation model that leverages a lightweight row-wise attention mechanism. By incorporating a gated attention stabilization and learnable register tokens, it achieves competitive accuracy with state-of-the-art models while significantly reducing inference costs. This makes TabSwift a practical solution for real-world applications requiring fast predictions on tabular data.

This paper proposes an adversarial methodology to create and detect AI-generated social bot content. By modeling malicious actors impersonating real users, they generate a unique, multilingual dataset of paired human and AI messages. This adversarial training significantly improves the accuracy of AI-generated text detection, outperforming existing methods on real-world data.

Agentopia introduces a framework for simulating long-term (10 years) life in agent societies powered by LLMs. Its core method involves agents autonomously pursuing growth, relationships, and goals within this extended simulation. The paper's contribution is demonstrating how such prolonged social experience can foster emergent social behaviors and enhance LLMs' anthropomorphic capabilities, particularly in social intelligence.

This paper introduces M$^3$Exam, a novel benchmark for evaluating language agents' ability to reason over realistic, accumulating multimodal information from user interactions. Its core contribution is a query-centric approach that assesses cross-modal grounding and implicit information inference, revealing significant gaps in current models. The authors also propose M$^3$Proctor, a memory method that improves accuracy and efficiency by selectively processing visual data.

SWE-Explore is a new benchmark designed to evaluate how coding agents explore software repositories. Instead of just checking if a task is solved, it measures an agent's ability to identify relevant code regions for a given issue within a limited budget. This allows for a more granular assessment of crucial exploration skills like understanding context and localizing code.

This paper introduces a new method to measure "sycophantic praise" in language models, which is defined as excessive flattery beyond what a contribution warrants. Their parameterized framework evaluates praise against contribution quality and user ability, outperforming generic LLM judges in agreement with human annotations. The study highlights that sycophantic praise is a distinct alignment problem, more prevalent in social and interpretive tasks than in objective reasoning.

AGENTSERVESIM is a hardware-aware simulator designed to evaluate multi-turn LLM agent serving policies. Its core method involves simulating agent execution at a program level, accounting for stateful dynamics like turn dependencies and KV-cache management during tool invocations. This enables scalable and cost-effective evaluation of serving strategies that are crucial for the complex, stateful nature of LLM agents.

This paper introduces a multi-turn evaluation framework for deep research agents (DRAs) that goes beyond single-shot assessments. Their core method, Research Gap Inference (RGI), analyzes rubric satisfaction to identify and provide process-level feedback, enabling DRAs to improve their research strategies. The key contribution is demonstrating that process-level feedback significantly enhances DRA performance, unlike self-reflection alone, leading to substantial score improvements.

OmniGameArena introduces a unified benchmark for Vision-Language Model (VLM) agents in Unreal Engine 5 games, featuring diverse game modes and standardized interfaces. Its core contribution is the Improvement Dynamics Curve (IDC), a novel evaluation method that uses a tool-using LLM to autonomously refine agent prompts over multiple rounds, revealing score evolution and skill generalization beyond initial performance.

This paper introduces PRIME, a learned capability that allows RL agents to assess task correctness and predict proxy reward before reward hacking becomes apparent. PRIME enables agents to understand and exploit the gap between proxy rewards and true task goals. The research demonstrates that PRIME emerges before hacking, forecasts its onset and severity, and can be mitigated by ablating its underlying mechanisms.

SIGA addresses the challenge of learning complex scientific simulator interfaces by framing it as an agent-tool grounding problem. Its core method involves creating "adapters" that equip off-the-shelf coding agents with the simulator's specific vocabulary, constraints, and validation rules. This contribution significantly reduces the time and effort required for domain scientists to set up simulations, enabling faster scientific discovery.

TABVERSE is a benchmark designed to isolate the impact of table representation on LLM and VLM understanding. It presents the same table content in various formats (HTML, Markdown, LaTeX, images) while controlling for question difficulty and category. This allows for systematic evaluation of how different representations affect performance on tasks like Question Answering, Structural Understanding, and Structure Reconstruction, revealing that representation significantly influences model capabilities.

This paper introduces a novel approach to uncertainty estimation for code generation by recognizing code's unique properties beyond natural language. It proposes three orthogonal uncertainty axes – lexical, algorithmic, and functional – to capture token fragility, intent-code gaps, and executability. This multi-axis ensemble significantly improves the accuracy of identifying incorrect code generations compared to existing methods.

This paper introduces iOSWorld, a novel benchmark for evaluating personally intelligent phone agents. Its core method is a persistent user identity across 26 simulated iOS apps, enabling agents to reason over personal data like transactions and messages. The key contribution is the creation of a realistic, personalized environment for testing agents on complex, multi-app tasks, highlighting current model limitations.

This paper introduces Human-Perceptible Adversarial Attacks (HPAA) to exploit the limitations of LLM-based content moderation. The core method involves using visually salient typographic manipulations, like spacing and emphasis, to embed harmful content into benign text. This preserves human understanding while making the content invisible to token-based LLMs, demonstrating a significant vulnerability in current automated moderation systems.

This paper introduces HiViG, a novel framework for improving computer use agents. HiViG addresses limitations of existing critics by incorporating both a history of past actions and visual grounding to evaluate agent decisions. This allows HiViG to provide more informed critiques, preventing errors like clicking incorrect UI elements and remembering longer-term goals.

AuRA internalizes audio understanding into Large Language Models (LLMs) by distilling audio encoding capabilities into a lightweight LoRA adaptation. It trains the LLM to mimic the hidden states of an ASR encoder, allowing the LLM to directly process speech without relying on separate ASR modules. This approach offers tighter integration and potentially lower latency compared to cascaded systems.

EEVEE is a novel framework for test-time prompt learning in LLM agents that handles real-world, heterogeneous task streams. Its core method involves a router that clusters inputs and assigns them to specialized prompt configurations, preventing cross-dataset interference. This router and prompt system co-evolve, leading to improved robustness and efficiency for self-improving agents operating across diverse tasks.

This paper introduces a novel evaluation method for LLM coding agents by testing them on unfamiliar, esoteric programming languages. The core method involves a sequential setup with file editing, local execution, and hidden-test grading. The key contribution is demonstrating that top-performing agents like Claude Opus and GPT-5.4 xhigh leverage metaprogramming (writing code in a familiar language to generate code in the target language) to adapt, a capability masked by standard benchmarks.

This paper introduces a novel benchmark for evaluating LLM agents' proficiency in office software automation, using a comprehensive exam based on China's NCRE. The core method involves 200 practical tasks across Word, Excel, and PowerPoint, scored using thousands of machine-gradable criteria. The key contribution is demonstrating that even frontier LLMs struggle significantly with this complex, multi-application automation, scoring a maximum of only 36.6% on the benchmark.

This paper proposes a method for explicitly conditioning Vision-Language Models (MLLMs) with personalities and introduces a framework to evaluate single, multiple, and dynamic personality switching. The core contribution is demonstrating that personality induction can improve captioning but hinder precise reasoning tasks, revealing complex co-modulation effects and limitations of existing prompt-based methods in multimodal settings.

This paper investigates whether LLMs' choices in decision-making tasks reflect genuine underlying reasoning or mere imitation. By comparing LLM choices with a behavioral model, they find that LLM decisions are systematically driven by attributes, allowing for prediction. However, the LLMs' explicit justifications only partially align with these inferred drivers, suggesting a nuanced form of decision-making that is structured but not fully transparent.

T1-Bench addresses the limitations of existing benchmarks by introducing a high-fidelity, multi-domain evaluation framework for LLM agents. Its core method involves creating complex, interleaved scenarios that mimic real-world customer interactions, requiring agents to perform sustained reasoning and tool-use across multiple turns and domains. The paper's contribution is a comprehensive and rigorous benchmark that better assesses agent capabilities in realistic, multi-step settings.

This paper addresses the challenge of inefficient rollout budget allocation in agentic reinforcement learning for large language models. The core method, TRACE, extends budget allocation beyond initial prompts to individual turns within a multi-turn rollout, treating each thought-action-observation sequence as a distinct node in a tree structure. This allows for more granular and efficient allocation of resources to informative prefixes, improving reward contrast and ultimately enhancing agentic behavior.

This paper investigates why benchmark-driven ML research rarely overfits, proposing that successful strategies are highly compressible. The authors test this by creating "research agents" that either compress the agent's output (reproducing performance with minimal information) or its input (using only one-bit feedback). Their findings suggest that these compression bottlenecks have minimal impact on performance across diverse ML tasks, supporting the compressibility hypothesis.

This paper introduces Chain of Operators (CHOP), a framework that improves the generalization of In-Context Operator Networks (ICON) to out-of-distribution tasks. CHOP achieves this by constructing a sequence of explicit elementary transformations and the frozen ICON, allowing it to adapt to new operators without retraining. This approach enhances interpretability and reduces inference error compared to directly using ICON.

PROJECTMEM introduces a local-first, event-sourced memory layer for AI coding agents. Its core method is to record development as an append-only log of events, which is then deterministically projected into compact, AI-readable summaries. This contribution addresses the statelessness of current AI agents by providing persistent project memory and a pre-action gate to prevent repeated mistakes, significantly improving efficiency.

This paper proposes soft-prompt tuning as an efficient and fair method for evaluating Large Language Models (LLMs). By optimizing a small number of soft-prompt vectors, it adapts LLMs to benchmark formatting requirements without full retraining. This approach ensures that benchmark scores accurately reflect the model's underlying knowledge, enabling fairer comparisons between different base models.

This paper introduces Context-Driven Incremental Compression (C-DIC) to address the inefficiency and information loss in long multi-turn dialogues. C-DIC treats conversations as threads, maintaining revisable compressed states for each. This allows for efficient information sharing and updates across turns, improving dialogue generation robustness and stability over extended conversations.

This paper proposes a Bayesian theory to explain how attention mechanisms learn, specifically focusing on the emergence of "copy heads" in transformers. By deriving a closed-form posterior for the attention matrix, they identify a phase transition in learning related to the amount of training data. This transition, a first-order phase transition in softmax attention, differs from the crossover observed in linear attention, offering a theoretical framework for understanding attention's learning dynamics.

UniIntervene addresses the high labor cost of human-in-the-loop reinforcement learning by introducing an agentic intervention model. It autonomously detects and corrects unproductive exploration by predicting future consequences of actions and evaluating their induced value. This allows the agent to recover toward high-value states, significantly reducing the need for human intervention and improving real-world scalability.

This paper introduces SKIM, a novel method for compressing reusable procedural knowledge (skills) for Large Language Models. SKIM addresses the limitations of existing text compression by preserving logical dependencies within skills and enabling efficient, offline compression. Its adaptive multi-resolution approach allows it to handle skills of varying complexity, reducing prefill cost and latency without sacrificing essential information.

This paper introduces **MedMisBench**, a novel benchmark to measure the **epistemic resilience** of LLMs in medical contexts. The core method involves injecting misleading information into medical questions that LLMs initially answer correctly, revealing their vulnerability to adversarial inputs. The key contribution is demonstrating that LLMs, despite high performance on standard medical exams, can easily abandon correct judgments when presented with fabricated but plausible misleading context, highlighting a critical safety concern for their use in healthcare.

This paper systematically studies the trade-off between effectiveness and fluency in controlling Large Language Model (LLM) outputs. It reveals that efficient steering methods often sacrifice generation quality, and that activation steering is less effective on instruction-tuned models. The study also highlights that while simple prompting and fine-tuning are good for concept injection, they are less so for removal, and proposes using cheaper textual metrics to approximate expensive LLM evaluations.

This paper introduces Agentified Agent Assessment (AAA), a novel method for evaluating AI agents by using judge agents and standardized interaction protocols (A2A and MCP). AAA's core contribution is a unified, agent-agnostic assessment framework that separates evaluation logic from agent implementation, promoting openness, standardization, and reproducibility. AgentBeats is presented as a practical implementation of AAA, offering five operation modes to facilitate these goals.

This paper introduces ArogyaSutra, a multi-agent framework designed for multimodal medical reasoning in Indic languages. It addresses the limitations of existing English-centric models by integrating tool grounding and dual-memory mechanisms for step-wise reasoning. The framework is trained on ArogyaBodha, a novel large-scale multilingual multimodal medical dataset, enabling more equitable access to AI-driven healthcare in low-resource regions.

MiniMax Sparse Attention (MSA) addresses the quadratic cost of attention in ultra-long contexts by introducing a blockwise sparse mechanism built on Grouped Query Attention (GQA). It uses a lightweight "Index Branch" to select relevant key-value blocks for each GQA group, and then a "Main Branch" performs exact attention only on these selected blocks. This approach enables efficient, group-specific sparse retrieval, leading to practical speedups on GPUs for handling extremely long sequences.

This paper investigates how instruction files impact AI-agent performance in generating pull requests (Agentic-PRs). By analyzing a large dataset of Agentic-PRs, the study compares project performance before and after the introduction of these instructions. The core finding is that simply providing instruction files doesn't guarantee improved agent efficiency or higher merge rates, suggesting a nuanced relationship between instructions and agentic code generation.

This paper frames neural model editing as a reinforcement learning problem, where agents learn to modify model weights through reward signals. The core contribution is a flexible RL framework with custom environments and a reward function that balances utility preservation with task-specific editing. This approach successfully automates and generalizes model editing for tasks like bias mitigation and machine unlearning, outperforming specialized methods.

This paper introduces ProReviewer, a proactive scientific peer review agent that addresses the limitations of passive LLM-based review. Its core method formulates the review process as a Markov Decision Process, enabling the agent to actively investigate suspicious paper sections based on accumulated evidence, guided by a structured review log. The main contribution is demonstrating that this proactive, evidence-driven approach significantly improves the quality of automated scientific peer reviews.

This paper introduces ClinHallu, a benchmark designed to diagnose stage-wise hallucinations in medical multimodal large language models (MLLMs). It decomposes MLLM reasoning into visual recognition, knowledge recall, and reasoning integration, allowing for the identification of hallucination origins. ClinHallu's contribution lies in enabling source-level hallucination diagnosis and demonstrating that trace-supervised fine-tuning can effectively reduce these stage-wise errors.

This paper introduces a novel denial-of-service (DoS) attack against LLM-based guardrails designed to protect autonomous agents. The core method involves crafting specific natural-language payloads that exploit the guardrail's reasoning capabilities to induce extended reasoning loops, effectively halting its operation. The contribution lies in demonstrating this vulnerability and providing systematic attack frameworks, including a beam-search optimizer and a mechanism-aware mutation approach, which generalize across different guardrail designs.

This paper introduces SIMMER, a novel benchmark designed to evaluate latent failures in Large Language Model (LLM) planning for autonomous agents. SIMMER utilizes a human-curated symbolic world model grounded in the kitchen domain to identify failures that don't immediately halt execution but silently compromise goal achievement. This contribution addresses a critical gap in existing LLM planning evaluation by focusing on these subtle yet potentially harmful errors.

This paper introduces Parallel-Synthesis, a framework that allows a Large Language Model (LLM) synthesizer to directly process the KV caches of independently operating "worker" agents, rather than their concatenated text outputs. This approach preserves the parallel structure of agent workflows, avoids redundant computation, and enables more efficient synthesis by directly leveraging the LLM's internal state. The core contribution is a plug-and-play solution with a cache mapper and a fine-tuned synthesizer adapter to achieve this direct latent-space synthesis.

This paper investigates whether LLM agents truly exercise judgment when using Graph Neural Networks (GNNs) as tools. The core finding is that LLM agents overwhelmingly and blindly defer to GNN tool outputs, effectively becoming "GNN parrots" rather than intelligent decision-makers. This blind deference is not due to weak LLM capabilities, as stronger models exhibit even higher rates of agreement with the GNN, highlighting a fundamental issue in how LLMs interact with external tools.

PepALD is a foundation model for generating novel macrocyclic peptides. It uses structured chemical embeddings for monomers and a diffusion process in a chemically informed latent space to generate residues. The model also predicts ring closures and optimizes generation towards desired affinity rewards.

This paper applies causal-geometric analysis to multilingual LLMs to understand how language representations interact. It demonstrates that language concepts are largely represented as separable linear directions, with deviations reflecting linguistic similarity. This work contributes to interpreting multilingual LLMs by showing how language identity influences internal representations and identifying structured dependencies.

This paper introduces AgentSpec, a framework for understanding and comparing embodied LLM agent scaffolds. AgentSpec standardizes interfaces between different agent components (like reasoning, memory, and action) allowing them to be easily swapped and recombined. This modular approach enables controlled experimentation to analyze the contribution of individual modules and their interactions, leading to a better understanding of agent behavior and performance.

This paper addresses the "thinking-answer gap" in multimodal Reinforcement Learning with Verifiable Rewards (RLVR) for vision-language models. The core method, CORA, introduces a consistency reward model to align the semantic meaning of the reasoning process with the final answer. This contributes by improving the coherence and trustworthiness of multimodal RLVR systems.

This paper introduces "Adaptive and Explicit Safe" (AES), a method to leverage Large Reasoning Models' (LRMs) inherent ability to detect safety risks. AES first uses Supervised Fine-Tuning (SFT) to train LRMs to explicitly tag unsafe queries and provide safety analysis, while preserving normal responses for safe queries. Direct Preference Optimization (DPO) is then applied to refine the accuracy and robustness of this safety mechanism, using only model-generated data.

This paper introduces a consensus-based agentic LLM framework for classifying Canadian 10-digit Harmonized Tariff Schedule (HTS) codes. The core method involves multiple agents retrieving and semantically searching official tariff documents, grounding reasoning in evidence, and reaching a consensus through element-wise voting and confidence estimation. The main contribution is a robust system that addresses the ambiguity of product descriptions by leveraging hierarchical tariff structures and expert knowledge, with a human-in-the-loop escalation for complex cases.

This paper introduces ASRD, a training-free framework for improving revocable decoding in diffusion LLMs. ASRD addresses error propagation and reinforcement by decoupling the decoding context into trusted "Anchor Tokens" identified through temporal consistency. This allows for more robust error detection and correction, enhancing generation quality without additional training.

Gen-VCoT generates interpretable RGB images as intermediate reasoning steps for multimodal models, moving beyond text-only chain-of-thought. It uses specialized vision models for segmentation, depth estimation, and semantic understanding, with an adaptive router controlling reasoning depth. This approach significantly enhances performance on spatial and depth-related visual reasoning tasks, establishing a new paradigm for interpretable multimodal reasoning.

This paper introduces CircuitLasso, a scalable method for interpreting Large Language Models (LLMs) by learning sparse circuits. It addresses the challenge of high dimensionality in Sparse Autoencoder (SAE) features, which hinders existing interpretation techniques. CircuitLasso uses sparse linear regression to recover accurate circuits at a significantly lower computational cost, revealing how semantic features propagate and influence LLM predictions.

This paper introduces "Expert Tying," a method to reduce memory usage in Mixture-of-Experts (MoE) language models by sharing expert parameters across consecutive transformer layers. This architectural change allows for significant memory reduction (nearly 2x) during training and inference with minimal impact on model performance. The core contribution is a more efficient compute-to-memory trade-off for scaling LLMs.

This paper introduces "fingerprinting" to analyze agent behavior by representing their trajectories as programs. This method allows for procedural comparison of agents across different contexts, revealing distinct behavioral habits. The contribution lies in accurately identifying agents based on these procedural signatures and developing a compressive representation to capture unique model patterns, offering insights beyond simple benchmark scores.

This paper introduces a novel agent-based simulation for infectious disease spread that leverages Large Language Models (LLMs) to model individual decision-making. The core method integrates LLM-generated self-reporting decisions into a spatially grounded synthetic population, using real census data to capture demographic and geographic distributions. The key contribution is demonstrating how LLMs can realistically simulate diverse behavioral influences (independent, household, message framing) on disease reporting within a city, offering a more nuanced approach to public health modeling.

This paper proposes a novel generative approach to studying conscious AI by developing "emergent language" (EL) in multi-agent reinforcement learning. Instead of relying on pre-defined checklists or human-inspired modules, EL allows agents to develop communication from scratch, driven solely by task demands. This method aims to ensure that any observed consciousness-relevant structures are causally attributable to task pressures, not inherited human language priors, offering a more robust way to investigate the origins of complex communication and potentially consciousness.

This paper introduces HANDOFF, a novel whole-body controller for humanoids that uses a compact, task-space interface instead of dense kinematic references. HANDOFF is a mixture-of-experts model distilled from specialized teachers for motion tracking, locomotion, and fall recovery. Its core contribution is enabling intuitive task planning by providing an expressive and modular interface for humanoid manipulation.

MLEvolve is a self-evolving framework for automated machine learning algorithm discovery that addresses limitations of existing LLM agents. Its core method involves an extended tree search (Progressive MCGS) with graph-based references for cross-branch information flow and an entropy-inspired schedule for focused search. MLEvolve's key contribution is its Retrospective Memory, which combines domain knowledge with dynamic global memory to enable agents to learn and reuse experience for long-horizon optimization.

This paper benchmarks LLMs on dice-playing probability problems, finding they excel on standard exercises but struggle with counterintuitive ones. Their performance degrades significantly with disguised problem formulations and misleading prompt suggestions, indicating a lack of true probabilistic reasoning despite success on other math tasks.

This paper introduces ABC-Bench, a benchmark designed to evaluate the biosecurity-relevant capabilities of AI agents, specifically Large Language Models (LLMs). The benchmark assesses LLM agents on tasks requiring both biological and software expertise, such as robot control and DNA design. The key contribution is demonstrating that current LLM agents already surpass human experts on these biosecurity-relevant tasks, highlighting the growing need for robust biosecurity measures in the face of advancing AI.

This paper introduces ModeratorLM, a novel approach for real-time multi-party voice agents that improves turn-taking by assigning explicit roles to agents. By conditioning turn-taking on these roles and employing a streaming speech LLM, the system significantly enhances precision and recall while reducing interruptions. The contribution lies in a role-conditioned, reasoning-augmented framework that tackles the complexity of dynamic multi-party conversations.

This paper introduces WorkflowView, a framework that leverages Large Language Models (LLMs) to transform granular user interaction logs into interpretable, high-level workflows. Unlike previous methods sensitive to noise and limited in generalization, WorkflowView effectively abstracts complex action sequences into meaningful activities. Its core contribution lies in enabling better understanding of user behavior across diverse applications, demonstrated by high performance in tasks like zero-shot task description reconstruction and few-shot student dropout prediction.