2026-W24

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 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 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.