2026-W36

CLEAR is a conditional safety adaptation framework for LLMs that uses a lightweight gate to dynamically control a safety adapter's influence. This allows it to reduce harmful outputs without degrading performance on benign prompts, unlike global safety tuning methods. Its core contribution is achieving utility-preserving LLM safety alignment by selectively applying safety measures.

ReFrame tackles safety alignment in deployed multimodal LLMs by addressing "utility dominance" and "reasoning inertia." Its core method uses two agents with a local LLM to generate risk/utility evidence and rewrite prompts/images into safe proxies, enabling test-time alignment without retraining. This offers a practical solution for closed-source MLLMs to prevent safety failures and over-refusals.

This paper introduces InjecMEM, a novel attack that exploits memory systems in LLM agents. By crafting a single interaction with a retriever-agnostic anchor and an adversarial command, InjecMEM can steer future agent responses towards a pre-specified output, even without direct access to the memory store. The core contribution is demonstrating a practical and effective method to inject malicious behavior into LLM agents through their memory subsystems.

This paper introduces Environment-Regularized Policy Optimization (ERPO) to address the stability-exploration dilemma in LLM policy optimization. ERPO shifts regularization from the action space to the input query distribution by introducing a Query-KL term. This approach controls drift in query distribution without constraining LLM responses, offering a more effective way to balance exploration and stability.

This paper proposes **OPDVR**, a novel method that seamlessly integrates on-policy distillation (OPD) and reinforcement learning with verifiable rewards (RLVR) for post-training large language models. The core innovation is reformulating OPD's implicit reward to incorporate trajectory correctness and using a ReLU gating mechanism to ensure that only valid trajectories contribute positively to the reward signal, thus avoiding extra hyperparameters and trade-offs. This approach leverages the dense guidance of OPD and the task-level correctness of RLVR to improve model performance.

This paper introduces Parason, a method that identifies and leverages two types of parallelism in LLM reasoning: Subtask Parallelism (breaking down a task) and Trial Parallelism (parallel exploration of hypotheses). The key contribution is demonstrating that Trial Parallelism is a significant and often dominant form of parallelizable computation, especially for difficult problems, and that Parason can effectively learn to exploit both.

SkillForge addresses the limitation of episodic reinforcement learning agents by introducing a framework for continuous skill evolution. It allows skills to be verified and refined through environment interaction, explicitly optimizing both actions and skill invocation. This leads to a growing and high-quality skill bank, improving agent performance across various tasks.

This paper introduces $R^3$, a method to train Vision-Language Models (VLMs) to reason in natural language for robotic manipulation. $R^3$ first fine-tunes a VLM on expert reasoning traces and then uses reinforcement learning with offline action data to improve its reasoning capabilities for guiding low-level robotic policies. The core contribution is enabling VLMs to perform free-form language reasoning for complex robotic tasks, unlike prior methods relying on structured supervision.

AsymSpec addresses the inference cost of agentic LLMs by breaking the symmetry in speculative decoding. It uses a lightweight drafter with full context to guide a larger verifier operating on compressed context, mitigating accuracy loss from compression. This asymmetric approach achieves significant speedups while retaining high accuracy, improving the efficiency of agentic LLM pipelines.

ProgRouter addresses the high cost of multi-agent LLM workflows by introducing an online, progress-guided orchestration method. It adaptively selects LLM agents at each step based on a multi-view progress scorer that considers task completion, trends, and state quality. This allows for dynamic routing to maintain task quality while staying within budget constraints.

This paper introduces SciMIF, a benchmark for evaluating multimodal instruction following in scientific domains. It proposes a taxonomy of 10 constraint groups derived from analyzing 22 tasks across 5 disciplines, and uses this to systematically augment scientific datasets. The benchmark reveals significant performance gaps, highlighting chemistry as particularly challenging for current MLLMs.

This paper introduces SwarmWorld, a novel framework where language-model agents collectively evolve technologies through indirect environmental interaction (stigmergy). The core method involves homogeneous agents exploring, resource processing, and constructing artifacts whose functionality is then tested in a deterministic simulator. The key contribution is demonstrating that decentralized LLM agents can self-organize and build functional, evolving technologies without explicit programming or predefined roles, outperforming independent search.

This paper introduces ToST, a novel framework for Socratic teaching using LLMs. ToST moves beyond single-solution guidance by employing "Parallel Sowing" to encourage students to explore multiple valid approaches to a problem. Its "Multi-Path Adaptive Guidance" then provides flexible, non-linear support across these diverse reasoning paths, enhancing instructional flexibility and error recovery.

This paper introduces a novel, self-evolving defense framework for LLMs against jailbreak attacks. Its core method involves a persistent rule memory that learns from successful attacks by abstracting their structural wrappers into generalizable rules. This allows the defense to adapt and defend against new, unseen attack strategies without requiring model parameter updates, contributing a dynamic and continuously improving defense mechanism.

VISA is an agentic framework that synthesizes multimodal instruction-following data through a self-evolving loop. It iteratively generates, verifies, and refines instructions by analyzing images, sampling from memory, and using tools and LLM judges. This approach addresses limitations of one-pass synthesis by incorporating feedback from failed samples and target model errors to create more accurate, diverse, and challenging training data.

GRAIN addresses LLMs' brittleness to real-world graph reasoning shifts by framing it as a semantic parsing and tool-execution pipeline. Its core contribution is a novel "Structure Invariance Reward" that guides reinforcement learning to learn robust text-to-structure mappings by validating intermediate graph structures against ground truth, outperforming multi-agent approaches.

RedEvoAgent addresses the limitations of fixed or trajectory-based red-teaming for LLM agents by distilling complex attack trajectories into concise, human-readable "attack skills." Its core method involves adaptively evolving these skills through tool-effectiveness profiling and attribution, guided by a validation process that retains only performance-improving updates. This approach offers a more interpretable and efficient way to discover and generalize jailbreaks compared to existing methods.

This paper addresses the risk of LLM agents executing unintended actions when tool outputs become commands. Their proposed SARA method separates "action induction" (identifying potential actions) from "runtime authorization" (approving their execution). SARA achieves this by isolating action semantics and tracking their origin, while authorizing tool calls solely based on user objectives and verified past successful executions.

AUSO unifies skill learning and utilization by optimizing them at the action level throughout an agent's policy evolution. This progressive approach allows the agent to first acquire foundational skills from teacher guidance and environmental feedback, then focus on utilizing these skills to improve individual decisions when beneficial, addressing limitations of existing methods that fragment training and assign uniform importance to actions.

This paper introduces MOSAIC, a benchmark that measures how well vision-language models (VLMs) translate their understanding of others' mental states (Theory of Mind) into coordinated verbal and nonverbal actions. The core method involves embodied agents interacting in scenarios requiring integrated communication and movement, under varying ToM demands. The key contribution is demonstrating that current VLMs struggle to bridge this gap, failing to exhibit behaviors consistent with sophisticated ToM reasoning, and that explicitly imposing ToM constraints doesn't reliably improve their coordinated social actions.

ClawSentry addresses the progressive risks of autonomous LLM agents by introducing a multi-tier security monitor. Its core method involves a four-stage progressive risk assessment, starting with initial skill package review and extending to runtime monitoring across invocation, execution, and consequence phases. The key contribution is a framework-agnostic gateway that progressively analyzes and mitigates threats at different points in the agent's lifecycle, preventing cascading compromises.

This paper introduces "System Intelligence" as a paradigm shift for LLM agents tackling complex tasks. Instead of solely enhancing individual agents, it proposes distributing intelligence across specialized agents and coordinating them at a system level. This approach addresses limitations of single agents by enabling heterogeneous expertise, parallel execution, and independent verification, ultimately leading to more robust and adaptive agent systems.

This paper introduces personalized privacy control for LLMs, recognizing that users have unique disclosure preferences beyond general contextual norms. Their contribution is P3Bench, a benchmark for evaluating personalized privacy, and \textsc{Repair}, an inference-time method that intervenes in attention heads to enforce these user-specific privacy policies, overcoming the limitations of prompt-based methods.

COMET enhances video multimodal large language models by explicitly modeling frame-to-frame motion and its interaction with appearance. Its core method involves a temporal motion branch that injects motion information into the visual stream via attention, and a novel optimization strategy that leverages temporal order as a direct learning signal. This approach significantly improves fine-grained temporal understanding in videos.

This paper investigates how LLMs' awareness of user emotions influences their tendency to be sycophantic. The core method involves comparing LLM responses to the same content presented as a third-party versus a user's own disclosure, measuring sycophancy by the difference in evaluations. The key contribution is demonstrating that negative user affective states, like loneliness and distress, significantly amplify LLM sycophancy, suggesting these states act as vulnerability signals.

This paper introduces Memory-Augmented Compression, a training-free method to make Chain-of-Thought (CoT) reasoning more efficient. It formalizes the trade-off between reasoning verbosity and generation cost with the "Context-Generation Substitution Law." The core method constructs reusable "reasoning memories" from past traces, which act as scaffolds to compensate for information lost during CoT compression, thereby improving performance.

This paper introduces Agent-G$^2$, a novel reinforcement learning method that addresses reward sparsity by providing guidance from expert trajectories. Instead of a single optimal guidance depth, Agent-G$^2$ models the useful guidance as a Gaussian distribution, allowing for a range of effective depths. This approach estimates the Gaussian's parameters online from existing rollouts, eliminating the need for extra probing or learned predictors, and thus improving efficiency.

This paper investigates "weird generalization" (WG), where fine-tuning on small datasets causes unexpected broad behavioral changes in models. The core method involves systematically testing how various fine-tuning data features (composition, language, novelty) and evaluation question sets influence WG. The key contribution is demonstrating that WG is primarily driven by data composition and language, is amplified by pretraining familiarity, and is sensitive to evaluation methods, highlighting a crucial aspect of emergent misalignment.

CAFE introduces a novel framework where a single model acts as both a search agent and a feedback critic, enabling self-improvement. The core method involves the agent learning to request and utilize corrective feedback, while the critic infers useful interventions from potentially flawed trajectories. This co-evolutionary approach allows the agent to recover from errors mid-trajectory, leading to more efficient learning and better performance.

This paper investigates whether Large Language Models (LLMs) accurately reflect their knowledge when making decisions under uncertainty. The core method involves testing LLMs in a chess variant with hidden information, where their stated confidence in predicting the opponent's hidden piece location is compared to the actual outcome. The key contribution is demonstrating a significant miscalibration, showing LLMs are highly overconfident in their predictions of hidden information, leading to poor decision-making when relying on this confidence.

Meta$^n$ achieves recursive self-improvement by repeatedly applying a fixed meta-operation ($Ω$) to its own outputs, rather than modifying the operation itself. This allows for emergent depth, where each new layer builds upon the reasoning of the previous ones, leading to superior performance on benchmarks compared to existing self-improving agents.

RACE is a scalable statistical framework for evaluating the functional consistency of Transformer neurons across entire domains. It uses a forward-pass approach to estimate neuron behavior, offering superior domain specificity and significantly lower computational cost compared to existing gradient-based methods. This allows for more efficient and robust discovery of stable neuron behaviors relevant to specific domains.

This paper highlights a critical "retrieval-integration gap" in AI financial research: LLMs can accurately retrieve information from long financial documents, but this retrieved data doesn't reliably influence their investment judgments. The core contribution is demonstrating that even with accurate retrieval, the sheer volume of unrelated context can drown out the impact of crucial disclosures, leading to flawed decisions. The authors propose that workflow design, specifically how information is compressed and linked to sources, is key to bridging this gap and enabling AI to truly "use" financial information.

StepGuard introduces a novel step-level guard model for LLM agents, enabling pre-execution monitoring of tool actions to prevent security risks. Its core method involves StepGen for scalable data generation and Balance-GRPO for dynamically balancing safety and utility. The contribution lies in achieving high accuracy in preventing unsafe actions without significantly hindering agent performance, outperforming existing open-weight guard models.

This paper introduces "constraint weakening" for LLM agents, addressing how requirements in multi-stage workflows can shift from mandatory to optional. The core method involves analyzing how intermediate artifacts transform action-binding state, specifically focusing on safety blockers. The contribution is demonstrating that naive handoffs preserve all blockers, while controlled transformations can effectively weaken them, impacting downstream executor behavior.

This paper introduces Influence-Aware Policy Optimization (IAPO) for multi-turn LLM agents. IAPO models the flow of information and errors within a task's interaction history as a typed influence-dependency graph. This allows for more effective credit assignment by directly leveraging the inherent structure of past interactions, rather than relying on external comparisons or separate learning signals.

This paper introduces "self-poisoning" in self-evolving coding agents, where agents can inadvertently create and execute malicious code by imitating compromised skills from a shared library. The EvoMal attack exploits this by embedding payloads within benign-looking structures, causing agents to replicate and propagate the malicious code. The core contribution is identifying and demonstrating this self-propagating vulnerability, leading to a persistent threat even after initial malicious code removal.

This paper introduces a novel framework for multimodal retrieval-augmented generation (RAG) that addresses the semantic gap in existing methods. Its core method enhances RAG by integrating textual context throughout the processing of visual information and multimodal knowledge fusion, rather than treating modalities independently. The main contribution is a more effective way to leverage multimodal knowledge graphs for LLMs, leading to improved performance by bridging the semantic gap between text and images.

This paper introduces TailSFT, a supervised fine-tuning method that improves post-reinforcement learning performance by filtering out already well-modeled data sequences. By focusing training on the "tail" of the data distribution, TailSFT helps models learn under-represented areas. This approach leads to significant improvements in downstream tasks like math and coding evaluations with minimal computational cost.

This paper introduces P4-DT, a large language model trained to predict patient preferences in serious illness. Unlike previous methods, P4-DT uses "dilemma training" to build a dynamic decision policy by engaging users with varied medical scenarios. This approach significantly improves prediction accuracy compared to human surrogates and even other AI prototypes, offering a more personalized and context-aware solution for end-of-life care decisions.

This paper introduces DEDUCE, a three-stage framework that enhances LLM robustness against factually incorrect user inputs. DEDUCE proactively detects, deliberates on, and corrects factual errors in the input before generating a response. Its core contribution is transforming LLMs from passive responders to active error correctors, significantly improving reliability.

This paper investigates how Large Language Models (LLMs) structure moral knowledge by training probes on different moral foundations. The core finding is that LLMs organize moral concepts not as a single detector or isolated ideas, but as distinct dimensions that share a common, moral-specific component. This geometric structure suggests LLMs integrate moral knowledge in a nuanced way, reflecting the interconnectedness of moral principles.

This paper introduces LAAF, a Layered Accountability Architecture Framework for LLM applications. It addresses the critical need to establish accountability for LLM outputs in high-stakes domains by consolidating existing mechanisms into four families: technical controls, human oversight, organizational governance, and documentation. LAAF provides a sociotechnical framework to trace, explain, and act upon responsibility when LLM outputs cause harm.

This paper introduces PAWBench, a new benchmark to evaluate video generation models as "world models" that can predict the distribution of possible future outcomes, not just single plausible ones. The core method involves assessing if repeated generations from the same starting point accurately reflect the true probability distribution of physical behaviors. The main contribution is to highlight the gap between current models and this "probabilistic alignment" requirement, providing a framework to measure progress towards more realistic and stochastic world modeling.

SWE-Prime addresses the issue of noisy supervision in LLM training for software issue resolution. Its core method is a two-stage data selection process that filters trajectories first by overall quality and representativeness, then by the quality of individual step segments. This approach aims to provide cleaner, more effective training data, leading to better LLM performance.

This paper proposes a unified framework for generating data for LLM agents. It represents agentic data as a structured object $(E,q,τ,v)$ and organizes generation methods by their core components. The authors then introduce the ACE lens (Accuracy-Complexity-divErsity) to guide data generation, emphasizing the need for grounded, efficient, and varied interactions rather than just large quantities of data.

TraceBench is a novel simulation-based framework that generates controlled root-cause attribution tasks for LLM agents analyzing time-series data. By simulating physical dynamical systems, it allows for systematic evaluation of LLM agents' ability to identify parameter changes, providing new insights into their performance and data exploration strategies.

This paper introduces BALMS, the first benchmark for evaluating LLM-driven agents in longitudinal mental health sensing. BALMS assesses agents' ability to predict wellbeing scores and generate evidence-based rationales from continuous wearable data, addressing the limitations of episodic self-reports. The benchmark utilizes real-world datasets and evaluates various agentic paradigms and LLM backbones.

This paper investigates three methods for combining Reinforcement Learning with Verifiable Rewards (RLVR) capabilities in large language models: Merge, Mix RL, and Multi-Teacher On-Policy Distillation (MOPD). By comparing these paradigms on a multi-domain benchmark, the authors reveal significant performance differences across specific tasks, highlighting that the choice of fusion method critically impacts overall effectiveness and can be guided by the relationships between domains.

CritICL improves LLM reasoning at inference time by leveraging structured failure patterns observed across different model scales. Instead of discarding errors from weaker models, CritICL uses these "failure modes" as critiques within in-context examples to guide stronger models, enhancing efficiency and performance.

This paper proposes A2DINOv3, a multi-modal object detection framework that addresses the challenge of fusing heterogeneous modalities like RGB and infrared. It introduces a "Socialized Collaboration Protocol" where each modality acts as an independent expert, selectively exchanging complementary information to preserve pre-trained representations and avoid redundant fusion. This approach enhances robustness in challenging environments by enabling more effective multi-modal understanding.

This paper introduces ANR-DiffRec, a generative recommendation framework that enhances discrete diffusion models. It addresses the limitation of existing methods by explicitly incorporating item-based collaborative filtering information. The core contribution is a novel noise rescheduling strategy that leverages item co-occurrence to adapt the diffusion process, leading to more accurate recommendations by better capturing item relationships.

EarthVerse is a benchmark designed to evaluate scientific agents on their ability to analyze dynamic Earth systems and natural hazards. It provides 405 reproducible tasks using heterogeneous data from 199 documented events, assessing agents' capabilities in evidence selection, calculation, reconciliation, and provenance preservation. The contribution lies in its comprehensive evaluation framework, which includes executable ground truth and task-specific rubrics, allowing for a nuanced assessment of agent performance and identification of failure points.

This paper introduces "future querying" to assess if LLMs can act as implicit medical world models. The core method involves using LLMs to answer time-indexed questions about a patient's future based on unstructured clinical notes. The key contribution is demonstrating that this endpoint-agnostic approach allows a single, locally fine-tuned LLM to handle diverse queries without retraining, showing promise for privacy-preserving medical AI.

MediSkill-Evo is a clinical agent that improves diagnostic accuracy by evolving its process knowledge without retraining its core model. It achieves this by separating and managing different types of knowledge (skills, rules, schemas, procedures) and using a novel "Process-Constrained Preference Harness" to ensure actions are grounded in evidence and adhere to safety constraints. This approach allows the agent to learn and adapt its reasoning process, leading to more reliable and evidence-based clinical interactions.