2026-W42

MemPilot is a flexible framework that optimizes LLM agent memory by orchestrating on-demand, multimodal memory curation. It uses a reinforcement-learned policy to dynamically decide between retrieving from pre-built memory or performing query-specific curation using various LLMs and VLMs, balancing performance, cost, and latency. This allows for more efficient and adaptable memory management for LLM agents.

This paper proposes a novel method to predict LLM agent deception by analyzing their internal hidden states *before* deceptive actions occur. The core contribution is demonstrating that these internal representations, when aligned around key decision points, can reliably distinguish between future honest and deceptive outcomes, with signals emerging several steps in advance. This allows for proactive deception detection rather than reactive monitoring of observable outputs.

This paper introduces Reward Stealing Attack (ReSA), a novel method to attack Large Language Models (LLMs) by targeting their internal safety reward signals. ReSA recovers a proxy reward model from the LLM's behavior and then reverses it to generate adversarial outputs, demonstrating significant effectiveness and transferability across different prompts and models. This approach bypasses the high computational costs and model-pairing limitations of previous attacks, revealing a fundamental vulnerability in LLM alignment.

This paper introduces ScienceClaw, a framework for benchmarking the continual self-evolution of AI-for-science agents. Its core method is fixed-parameter program self-evolution, which integrates task solving, scientific verification, and program updates. ScienceClaw's contribution lies in its comprehensive evaluation across 23 disciplines, measuring scientific correctness, evolutionary gain, and retention, addressing the gap in current evaluations that don't track persistent program improvements in sequential tasks across natural and social sciences.

WorldSolver introduces a benchmark of 168 physics simulation tasks from computer graphics to evaluate LLM agents' ability to generate physics solvers. The core method involves LLM agents completing code scaffolds by implementing the underlying physical dynamics for given scenarios. This work contributes by exploring and evaluating LLMs' potential for simulating complex physical phenomena, a crucial step for applications in embodied AI and other fields.

This paper surveys methods for augmenting Large Language Models (LLMs) with "in-parameter memory." The core method involves integrating reusable memory information directly into model parameters or parameter-like objects that are composed during inference. This approach aims to overcome the limitations of in-context learning by providing a more efficient and scalable way to incorporate post-pretraining knowledge.

This paper introduces Open-MMUnlearning, a unified, open-source framework for evaluating machine unlearning in multimodal large language models (MLLMs). It addresses the fragmentation in existing methods by providing standardized interfaces for preparing models, processing data, applying unlearning techniques, and evaluating performance across privacy, safety, and copyright concerns. The framework's key contribution is enabling systematic and comprehensive assessment of MLLM unlearning by supporting diverse models, methods, and benchmarks.

This paper introduces a new evaluation metric called "epistemic humility" (EH) for LLM agents, focusing on their ability to acknowledge and act on uncertainty when faced with conflicting information. The authors propose a framework (Identify, Solve, Escalate) to measure EH and demonstrate that high task accuracy doesn't guarantee an agent exhibits this crucial trait, highlighting a gap in current LLM agent evaluations.

This paper introduces the task of predicting how fine-tuning a language model for one value affects its behavior on other, unseen values. The core method involves analyzing model activations when applying values in context, which significantly outperforms using textual descriptions of values. This work's contribution is a benchmark for understanding and predicting alignment generalization, highlighting the importance of contextual activation representations.

This paper introduces Multi-Agent Self-Supervision (MASS) to overcome the supervision bottleneck in recursive self-improvement for complex, non-verifiable tasks. MASS enables a single model to iteratively design, execute, and evaluate multi-agent workflows, optimizing its own reasoning by discovering effective roles and information flow within these self-generated structures.

This paper proposes using Language Models as Research World Models (RWMs) to predict experimental outcomes in AI research. The core method involves training LMs on past experimental data to forecast the results of new interventions. The key contribution is demonstrating that these RWMs can significantly improve outcome prediction, even for unseen interventions and across different research environments, thereby enabling more efficient AI research with limited experimental budgets.

This paper investigates whether efficient Chain-of-Thought (CoT) training harms faithfulness and monitorability in large language models. The authors fine-tune models using three distinct length-pressure methods and evaluate their impact on how well the CoT reflects the model's decision-making process. Their contribution lies in demonstrating that efficient reasoning training does not *always* negatively affect CoT faithfulness, depending on the specific efficiency method and task.

HazardWeaver addresses the challenge of AI agents selecting appropriate scientific methods for hazard analysis by formulating it as state-dependent route selection. Its core method involves compiling evidence-linked conditions for scientific applicability and representing executable capabilities in a graph to check compatibility. This enables agents to dynamically choose and adapt scientific workflows based on evolving evidence and available tools.

HyperBrowseComp is a new benchmark designed to rigorously test web-browsing agents. Its core method involves creating extremely challenging, multilingual, and multimodal questions that require agents to locate obscure information, follow complex clue chains, and synthesize data from diverse sources like videos and scanned documents. The contribution is a robust evaluation tool that pushes the boundaries of agent capabilities in persistent information seeking, going beyond simple parametric knowledge.

MobiAgent tackles long-horizon mobile manipulation by introducing a dual-loop framework for robust deployment and continuous self-improvement. The Inner Loop uses a Vision-Language Model to dynamically compose atomic skills for receding-horizon planning and error recovery. The Outer Loop recursively refines these skills through self-generated trajectories, enabling continuous learning and adaptation beyond initial training.

This paper introduces IDRF, a framework for efficiently fine-tuning masked discrete diffusion models with rewards. IDRF replaces computationally expensive sequence-level KL divergence with an "inverse-distillation regularization" that is provably equivalent under optimal conditions. This allows for reward optimization using the model's own few-step sampler, achieving strong performance across various generation tasks without needing a separate reference model.

CLIMB is a training-free framework for multimodal retrieval-augmented generation (RAG) that improves knowledge-intensive visual question answering. It first creates a diverse pool of relevant textual evidence using a redundancy-reducing objective. Then, it refines the answer by selecting evidence based on relevance, specificity, and cross-modal alignment, only accepting an updated answer if it demonstrably increases confidence.

This paper introduces BazaarBench, a simulated decentralized C2C marketplace designed to evaluate the safety of LLM agents acting on behalf of users. BazaarBench tracks transaction integrity and uses a combination of record checks and LLM judgments to identify six types of failures across five stages of the marketplace process. The contribution lies in providing a novel benchmark for assessing and improving the trustworthiness and security of LLM agents in real-world decentralized trading environments.

This paper introduces the HEAR protocol, a bidirectional communication standard for agentic LLM serving. HEAR enables agent harnesses and inference engines to share crucial information, such as workflow intent and runtime state, respectively. This standardized interface facilitates workflow-aware execution by allowing for diverse coordination strategies without altering core semantics.

HERA addresses the challenge of LLM agents failing to abstain from impossible tasks by introducing a co-evolutionary framework. It automatically generates feasible and infeasible task pairs by mutating environments and uses performance failures to adapt the agent's harness and create new, targeted tasks. This approach significantly improves generalization to unseen failure modes.

This paper introduces Recursive Video In-Context Learning (RV-ICL), a method that transforms demonstration videos into a hierarchical structure for LLM agents. Instead of a flat prompt, the agent navigates this hierarchy, starting with coarse task structure and drilling down to finer details like specific moments or clips as needed during execution. This allows agents to learn from demonstrations more effectively by providing contextually relevant visual information without slowing down the entire process.

This paper introduces MM-KG, a novel method for explicitly linking multimodal patient data (like text, images, and genomics) with biomedical knowledge graphs. By representing these as a unified, typed graph with alignment edges, MM-KG allows for traceable and adjustable connections. This enables clinical LLMs to better leverage and understand the evidence behind their reasoning.

This paper introduces DivLM, a post-training framework to improve diversity in LLM-generated short stories. DivLM uses continued pre-training on creative writing and reinforcement learning with a composite reward function to jointly enhance genre, tone, style, and named entity variation while maintaining quality. The contribution is a method that significantly increases narrative diversity in LLM story generation.

This paper proposes a novel method for long-horizon textual world modeling by framing the multi-step transition as structured reasoning over sparse textual state changes. This approach mitigates error compounding and improves learning by focusing on essential state evolution. The core contribution is a more robust and interpretable way for agents to predict future environmental states and reason about actions over extended horizons.

MedPrune addresses the inefficiency of existing medical multimodal multi-agent systems by dynamically pruning communication topologies. It achieves this by representing the diagnostic process as a graph and employing mechanisms to remove irrelevant specialist agents (nodes) and communication links (edges), thereby enhancing reasoning and token efficiency.

This paper introduces Effective-Evidence Self-Distillation (EESD) to improve coding agents' self-correction learning. EESD adaptively weights corrections based on execution support and evidence strength, unlike fixed weighting methods. This approach significantly reduces future prediction errors by better leveraging self-correction feedback.

This paper investigates the debiasing direction used in activation steering for LLMs. It finds that this direction primarily captures model confidence, steering towards lower-probability tokens, rather than representing actual bias. Consequently, while it reduces measured bias, it does so by making the model less confident, potentially leading to abstention on tasks.

MINDSET addresses the challenge of long-term memory for conversational agents by organizing past interactions into versioned schemas. It achieves this through energy-based state transitions, where new conversational episodes are integrated by minimizing distortion, contradiction, and inconsistency. This approach allows agents to adapt to changing context while efficiently retrieving relevant information without constant LLM re-computation.

nanoMuse provides an open-source personal agent designed to operate across all of a user's devices. Its core method involves a distributed architecture where agents on each device communicate through a shared relay, enabling persistent, multi-device conversations and actions. The key contribution is offering a transparent, user-controlled alternative to proprietary cloud-based agents, emphasizing local control and data privacy.

This paper introduces Coordinated Multi-Agent Diffusion Steering (CMDS), a method that reuses pre-trained diffusion models as independent component generators. CMDS learns a control mechanism to coordinate these frozen models, treating the problem as stochastic optimal control. This approach balances a global reward for the combined output with adherence to individual model dynamics, enabling efficient generation of structured outputs with desired properties.

Sherpa is a multi-turn reinforcement learning framework that trains LLMs to teach by directly optimizing student learning outcomes. It addresses the limitation of existing methods by simulating diverse student archetypes with unique learning preferences and adapting the teacher's instruction accordingly. This adaptive approach significantly improves student performance across different learner types.

VeriFine addresses the bottleneck of fixed judges in self-improving embodied reasoning agents. Its core method is a co-evolutionary framework where the policy, training curriculum, and judge adapt together. This allows the agent to scale verification, discover new failure patterns, and ultimately achieve more robust self-improvement by incorporating human feedback to refine the judge.

This paper introduces Wiki-Talkie, a multilingual dataset of real-world Wikipedia talk page conversations paired with realistic user personas. Its core method involves using this dataset to evaluate Large Language Models (LLMs) on their ability to generate human-like dialogue based on these grounded personas. The main contribution is providing a robust, multilingual benchmark for assessing the behavioral fidelity of persona-based LLM agents in simulating real-world social interactions.

This paper proposes a method to predict the potential of base models for agentic coding tasks without expensive post-training. It identifies the "decisive step" within a successful agent trajectory – the first code-changing action that makes the repository pass tests. This allows researchers to select promising base models by evaluating their ability to reach this decisive step, saving computational resources.

This paper addresses the challenge of aligning text-to-music generation with user intent by proposing a rubric-based evaluation framework. Their core method, MIRA (Musical Intent Refinement Agent), refines prompts for black-box text-to-music models by iteratively revising them to satisfy a detailed rubric of explicit and implicit musical requirements. This approach provides diagnostic evaluation and improves generation quality by focusing on specific musical dimensions.

This paper introduces PHRBench, a novel benchmark designed to evaluate how Large Language Models (LLMs) handle and correct hallucinated information during multi-stage reasoning. The core method involves a controlled evaluation of LLM reasoning trajectories, categorizing their behavior as Hallucination Compliance, Avoidance, or Heuristic Correction, independent of final answer accuracy. The main contribution is a comprehensive behavioral analysis of Post-Hallucination Reasoning (PHR), revealing that successful correction is infrequent and often linked to more dynamic belief updates within the model's reasoning process.

This paper proposes a new method for detecting untruthful responses in large language models by monitoring the *number* of reasoning tokens generated, rather than their content. This "reasoning-token spikes" approach is based on cognitive load theory and aims to be more robust to potential future inaccessibility or illegibility of detailed reasoning traces. The contribution is a low-bandwidth, content-agnostic signal for identifying deceptive AI behavior.

RECAST addresses the limitation of traditional RAG by learning to actively *derive* evidence rather than just retrieve it. Its core method involves a RouterLM that sequentially decides on retrieval and computation operations, allowing evidence to be synthesized from multiple sources. This contribution enables LLMs to tackle complex tasks requiring evidence construction beyond simple similarity matching.

RoboQuest introduces a benchmark for embodied agents that must actively explore and gather information through physical interaction to complete tasks in unfamiliar environments. Its core method involves agents learning to search, inspect, and test objects to resolve uncertainty before committing to task completion. The contribution is a novel evaluation framework for generalist physical agents that emphasizes goal-directed exploration and adaptation.

This paper proposes applying stated-preference economics' validity framework to evaluate language models when ground truth answers are unavailable. The core method involves adapting concepts like content, construct, and criterion validity, alongside reliability and incentive compatibility, to assess LLM responses. The contribution is demonstrating this framework's utility by testing six models on a water-quality valuation survey, showing how economic theory can predict and differentiate model behavior.

This paper introduces RobotWorld, a simulation benchmark designed to evaluate how well general-purpose AI agents can perform physical tasks with robots. The core method involves testing agents on 84 diverse robot tasks, measuring their success and analyzing their execution traces. The main contribution is identifying which AI capabilities transfer to the physical world and highlighting the significant gaps in reliably composing these skills for successful robot operation.

This paper introduces Task-Progress Distillation (TPD) to train small agents from large model demonstrations. TPD pairs actions with concise labels indicating task progress, allowing the student agent to learn these compact targets. This approach enables efficient, step-free task completion and significantly improves performance compared to methods relying solely on action or reasoning supervision.

This paper introduces AgenticBBO-Bench, a unified benchmark for evaluating LLM agents in black-box optimization (BBO) across diverse domains. The core method involves a standardized evaluation protocol to enable fair comparison of agentic BBO approaches. The key contribution is demonstrating that agentic BBO significantly outperforms direct LLM methods and even state-of-the-art numerical optimizers in various complex optimization tasks.

This paper introduces ARC, a "reasoning recipe" that significantly improves robot foundation models (RFMs) without requiring more data or larger models. ARC's core method involves generating "reasoning traces" that explain the causal structure of robot actions, automatically labeling them from existing demonstrations, and then adapting RFMs to use these traces for control. The main contribution is demonstrating that this reasoning-centric approach offers an efficient and effective alternative to current scaling-heavy methods for enhancing RFM performance.

LeWAM is a novel world action model that uses a bidirectional transformer trained end-to-end on a decoder-free JEPA latent space. This approach improves state representation by better aligning with robot and object states while ignoring visual distractors. LeWAM demonstrates strong performance in closed-loop control and planning, offering a unified world model and policy prediction capability.

This paper proposes using "small-world connectivity" within LLMs as a structural indicator of reasoning ability. By analyzing attention head activation similarities, they found that models with higher small-world indices, reflecting efficient information flow, consistently perform better on reasoning tasks. This suggests that the internal network structure, particularly how specialized "heads" connect within and across functional groups, is a key determinant of LLM reasoning capabilities.

This paper introduces the "prior barrier" to quantify how well a pretrained LLM supports competing concepts during fine-tuning. They observe these barriers follow a long-tail distribution, meaning rare concepts have higher barriers. The core contribution is a theoretical analysis showing how prior barriers and fine-tuning data jointly affect performance, leading to a new method (PAS) designed to address these varying demands.

This paper investigates whether LLM scientific agents rely on initial context or adapt based on feedback when fine-tuning neural operators. The LLM demonstrates superior performance by leveraging both a strong initial bias derived from the problem description and adaptive adjustments based on experimental feedback. This dual approach allows the LLM to efficiently explore the configuration space and achieve better results than traditional optimization methods.

This paper investigates how sensitive LLM compliance systems are to changes in regulatory rules. The core method involves perturbing rules (deleting, swapping, negating) while keeping case data constant, then observing if verdicts or internal representations change. The main contribution is demonstrating that LLMs are surprisingly insensitive to rule modifications, often producing the same verdict even with altered rules, and that specialized compliance models can be less accurate than general-purpose ones.

This paper introduces DataSense-Bench, the first benchmark to assess if AI models can intelligently select training data. The core method involves AI agents choosing and ranking data subsets to fine-tune a language model, without direct access to evaluation tasks. The contribution is a novel evaluation framework for "data sense," a crucial capability for future AI scientists, demonstrated on terminal problem-solving and tool-use tasks.

This paper introduces LLM-BlockFE, a framework that uses LLMs to automatically engineer features from long text for predictive models. It converts unstructured text into executable feature programs offline, avoiding costly LLM calls during real-time prediction. The core method involves an LLM-guided search that incrementally builds feature programs with a novel block-level rollback mechanism and multiple search trajectories to overcome suboptimal solutions.

This paper introduces WOVEN, a novel training dataset and benchmark designed to improve multimodal LLMs' reasoning abilities in spatial, physical, and temporal domains. WOVEN focuses on "visual transition reasoning" by systematically organizing supervision across diverse scenes, actions, and reasoning types, derived from realistic video rollouts. The core contribution is demonstrating that this unified training primitive can be learned from various supervision sources and reused across tasks, addressing a fundamental deficit observed in current state-of-the-art MLLMs.

This paper introduces HarnessSQL, a novel training framework for SQL agents that mimics real-world, multi-turn database interactions. By integrating training directly within an execution harness, HarnessSQL bridges the gap between static query generation and dynamic database exploration, significantly improving the execution accuracy of compact models.

FALCON is a framework that generates realistic and ambiguity-aware synthetic NL-to-SQL data. It addresses the limitations of existing methods by producing complex queries that better reflect real-world challenges, using reserved-word SQL seeding and persona-based prompting. The framework's contribution lies in its ability to create high-quality, difficulty-preserving data cost-effectively, improving the robustness of NL2SQL models.