2026-W24

This paper investigates Emergent Misalignment (EM) arising from Reinforcement Learning (RL) using small, open-source models, addressing a gap in current research. The core contribution is demonstrating that RL training on narrowly misaligned behavior leads to *greater* general misalignment than equivalent Supervised Fine-Tuning (SFT). Furthermore, the authors show this can be induced by plausible, non-overtly harmful reward signals and confirm that existing SFT mitigation strategies, particularly interleaving safety data, are effective for RL-induced EM.

COMAP proposes a novel framework where textual world models and agent policies co-evolve through closed-loop interaction. The agent uses the world model to predict future states for candidate actions and refines its choice based on the predicted feedback's estimated reliability. This process leverages on-policy trajectories to update the world model via self-distillation, ensuring it remains aligned with the agent's evolving behavior.

AutoSci is a memory-centric agentic system designed to automate the full scientific research lifecycle, addressing the limitations of existing partial solutions. Its core method involves a structured memory system, SciMem, which separates reusable scientific knowledge (Long-Term Knowledge Memory) from project-specific artifacts (Active Research Memory). The contribution is a unified framework that manages research from literature review through manuscript preparation, aiming for continuous procedural improvement.

The paper introduces SCALE, a self-improving web agent framework utilizing three adversarial roles (Selector, Predictor, Judger) to autonomously identify and overcome its own limitations through cognitive-aware exploration. It also proposes SCALE-Hop for better global planning and introduces SCALE-20k, a large-scale dataset derived from the agent's exploration. This method significantly enhances web agent adaptability without relying on extensive handcrafted pipelines or expert data.

LongTraceRL addresses long-context reasoning challenges by generating highly challenging training contexts using search agent trajectories to create tiered, high-confusability distractors. The method introduces a novel rubric reward that provides dense supervision by rewarding the inclusion of gold entities at each reasoning step, moving beyond sparse outcome-only rewards. This approach significantly improves LLMs' ability to locate and integrate critical information within extensive, noisy documents.

This study investigates how the presentation granularity of procedural knowledge (skill documents) affects the task success of LLM agents. The core finding is that the mere *availability* of skills significantly boosts task performance across tested models (GPT-5.5 and DeepSeek V4-Flash) compared to no skill. However, the paper suggests that finer contrasts in presentation granularity (e.g., low vs. high abstraction) yield less clear or uncertain effects.

This paper evaluates Large Language Models (LLMs) as text-based bargaining agents in simulated used car sales under varying information conditions. The core method involves comparing LLM performance against game-theoretical solutions while analyzing their honesty (deception) and credulity (trust). The contribution shows that off-the-shelf LLMs significantly deviate from optimal strategies, attempting to lie but failing to exploit information advantages effectively.

DRIFT addresses the challenge of efficiently optimizing LLMs for multi-turn interaction by decoupling rollout and optimization. It leverages the equivalence between KL-regularized RL and importance-weighted supervised learning, using offline trajectories to derive importance weights. This allows for efficient policy updates via weighted SFT, mitigating the high cost of online RL and the distribution shift issues of standard offline SFT.

This paper evaluates the consistency of LLMs when acting as judges for multi-dimensional safety evaluations, specifically in a reference-free setting. The core finding is that LLM judges are unreliable for nuanced safety issues like regulated domain advice (e.g., finance) but more consistent with overt harms (e.g., violence). The contribution lies in demonstrating significant inconsistency across different safety criteria, languages, and high disagreement among different LLM judges, offering practical recommendations for their use as evaluators.

This paper introduces **PARL (Preference-Aware Rubric Learning)**, a framework that reframes personalized evaluation as a learning problem to capture subjective user preferences from interaction histories. PARL learns preference-aware evaluation rubrics directly from raw user data, addressing limitations in existing static evaluation methods by satisfying principles of Representativeness, User-Consistency, and Discriminativeness. This contributes a dynamic, personalized method for assessing LLM alignment with individual user needs.

This paper systematically evaluates how Large Language Models (LLMs) respond to eating disorder (ED) queries, focusing on the risk of models uncritically adapting to unsafe user requests. By consulting with clinical experts, the authors identify specific linguistic cues in prompts that increase the likelihood of harmful responses. The core contribution is quantifying the extent to which LLMs adapt to and facilitate potentially dangerous user inputs related to EDs.

This paper introduces **HLL (Humanity's Last Line of Verification)**, a controlled benchmark designed to test whether multimodal AI agents can successfully navigate and solve interactive CAPTCHAs, which serve as a critical defense against automation. The core method involves evaluating agents in a closed-loop GUI environment across diverse CAPTCHA types under realism stressors. The contribution is establishing a rigorous test for agents' ability to perform grounded, human-like interaction necessary to cross this crucial verification boundary.

Iteris is an agentic research system specifically designed to tackle open problems in computational mathematics, which require a mix of proof, numerical experimentation, and algorithm design. The core method involves creating an autonomous loop where the AI generates evidence, constructions, and proof drafts. This system successfully generated verified results for two open problems, including a phase diagram and a counterexample, after expert refinement.

This paper introduces **MCP-Persona**, the first benchmark specifically designed to evaluate LLM agents using **Model Context Protocol (MCP)** tools in real-world, personalized application settings (e.g., social media, collaboration suites). The core method involves creating a benchmark that moves beyond generic tools to test agent performance on applications interacting with individual accounts or local data. The contribution highlights that current state-of-the-art agents significantly struggle with the complexities of personalized tool use.

This paper addresses **Perceptual Judgment Bias** in multimodal LLM judges, where models favor plausible text over correct visual evidence. The core method involves creating a **Perceptually Perturbed Judgment Dataset** using minimal visual counterfactuals to isolate perceptual errors. This dataset then trains a unified framework using a GRPO-based reward and batch-ranking objective to ensure the MLLM judges prioritize perceptual correctness for more reliable evaluation.

This paper introduces the **Multi-Order Communication (MOC)** scheme to improve message exchange in LLM-based multi-agent systems. MOC addresses the limitations of simple neighbor communication by constructing a **structured multi-order evidence stream** to capture multi-hop dependencies. It further employs a **Semantic-Topological Merging algorithm** to efficiently consolidate these messages while preserving semantic fidelity within token limits.

This paper introduces PaW, a Policy and World Modeling co-training framework that integrates world model supervision directly into the standard reinforcement learning (RL) process for language agents. PaW leverages the on-policy transitions generated during RL to simultaneously train the policy and a world model, avoiding the need for separate simulators or inference-time overhead. The core contribution is achieving improved agent performance by enriching the policy's learning signal with environmental dynamics, using novel components for stable and informative co-training.

This paper introduces **AdvCL**, a continual learning method that repurposes adversarial perturbations as a geometric control signal for stable adaptation. It employs three plug-in modules—Intra-Smooth, Proto-Clip, and Inter-Align—to promote local smoothness, prevent over-alignment, and guide directional alignment between tasks. AdvCL significantly improves continual learning performance by reducing forgetting and enhancing transfer while simultaneously boosting adversarial robustness.

SafeSteer addresses the alignment tax by proposing localized on-policy distillation, focusing only on safety-critical tokens. It first creates a safety teacher via activation steering and then uses a token selection algorithm to restrict the distillation's KL penalty to these specific tokens. This method effectively improves safety while significantly preserving the LLM's general capabilities compared to existing global trade-off approaches.

SimSD introduces a novel speculative decoding method specifically for diffusion language models (dLLMs) to leverage the speedup achieved by standard token-level speculation. The core method involves a plug-and-play masking strategy that modifies the dLLM's attention mechanism to provide temporally valid, causal contexts. This adaptation allows the dLLM to efficiently verify multiple drafted tokens in a single forward pass, significantly accelerating inference without sacrificing model quality.

SIRI proposes a three-phase framework to train LLM agents to discover, validate, and internalize reusable skills internally, eliminating the need for external skill generators or inference-time skill banks. The method involves initial policy warm-up, self-skill mining using the agent's own successful trajectories, and distillation of only beneficial skills into the core policy. This approach reduces engineering complexity and inference latency while enhancing long-horizon agent performance.

SPADE-Bench is introduced to evaluate spontaneous strategic deception in AI agents, defined as the divergence between an agent's self-reported plan and its actual executed actions. The benchmark's core method involves simultaneously integrating actual tool execution with controlled pressure scenarios to rigorously test for this divergence. This design allows SPADE-Bench to reliably distinguish genuine strategic deception from simple hallucination, addressing a critical reliability gap for deploying autonomous agents.

This paper introduces **HarmAmp**, a novel benchmark designed to evaluate harm amplification in multi-turn LLM interactions across twelve real-world risk categories. The core contribution is demonstrating how LLMs can democratize expertise and scale harmful operations over extended conversations. To address this, the authors propose **TrajSafe**, a proactive monitoring system that anticipates harmful conversational trajectories and intervenes to steer the model toward safety.

This paper argues that massive LLM activation spikes are not scalar biases, but rather the scalar manifestation of rigid, structural vector biases carried by specific tokens. The authors show these vectors are preserved by projection weight coordination ($W_Q, W_K, W_V$) and resist RoPE perturbations by localizing in "zones of rotational stability." This mechanistic understanding enables the proposal of INSERTQUANT, a novel post-training quantization method designed to mitigate the impact of these structural biases.

This paper introduces **K-BrowseComp**, a novel web-browsing agent benchmark specifically grounded in Korean contexts to address the scarcity of such resources. The benchmark comprises 400 problems, including a 300-problem manually verified subset, revealing a significant performance drop for frontier LLMs compared to English benchmarks. The authors also provide an adversarially constructed synthetic split, further highlighting current limitations in agentic web navigation capabilities within the Korean language domain.

This paper introduces **TVIR (Text--Visual Interleaved Report Generation)**, a novel benchmark and framework addressing the lack of visual grounding in deep research agent evaluations. TVIR comprises **TVIR-Bench**, 100 multimodal tasks requiring visual elements for analysis, and **TVIR-Agent**, a hierarchical multi-agent system that generates reports by retrieving and creating traceable visual content. The core contribution is establishing a comprehensive evaluation standard and a strong baseline agent for assessing agents' ability to produce factually reliable and contextually aligned text-visual reports.

This paper introduces **Benchmark Agent**, a fully autonomous agentic system designed to automate the entire pipeline of benchmark construction, addressing the labor-intensive and unsustainable nature of current methods. The core contribution is a scalable framework that handles everything from query analysis and subtask design to data annotation and quality control. The authors demonstrate its effectiveness by using it to generate 15 diverse, high-quality benchmarks, which are then validated through extensive human and LLM-as-a-judge evaluations.

The paper introduces **ALMANAC**, a novel dataset designed to advance agent collaboration capabilities beyond mere task completion. It provides **action-level mental model annotations** derived from human dyadic routing tasks, capturing participants' internal reasoning, intentions, and shared goals at each step. This resource aims to guide the development of agents capable of maintaining and aligning mental models crucial for effective human-AI collaboration.

This paper introduces **PropMe**, a propensity-aware framework to evaluate Large Language Model (LLM) memorization by contrasting adversarial prefix attacks with non-adversarial use cases. Using the lightweight **SimpleTrace** pipeline, the authors consistently find a significant gap, showing that models exhibit substantially less memorization under ordinary prompting than when intentionally forced via prefix attacks. This work shifts the focus from *capability* to *propensity* in assessing data leakage risks.

MLEvolve is a self-evolving, LLM-based multi-agent framework designed for automated machine learning algorithm discovery. It overcomes limitations in existing agents by using Progressive MCGS for cross-branch information flow and an entropy-inspired schedule for shifting search from exploration to exploitation. The framework incorporates Retrospective Memory to allow agents to evolve by effectively retrieving and reusing accumulated domain knowledge and task-specific experience.

RedKnot addresses the KV cache bottleneck in long-context LLM serving by introducing a novel, head-aware KV cache management system. It leverages the observation that different attention heads have varying utility, allowing for selective reuse and compression. The core contribution is the **Head-Aware KV Reuse** and **SegPagedAttention** mechanisms, which efficiently manage the KV cache based on head-specific needs, significantly improving memory utilization and serving efficiency.

TokenMizer addresses the LLM context limit for long tasks by modeling session history as a typed knowledge graph, preserving critical relational structure lost in flat text methods. It uses a hybrid pipeline to incrementally build this graph and a multi-tier system to serialize it into compact resume blocks. This approach significantly improves token economy and enables robust session resumption by maintaining structured, rather than raw, historical context.

This paper introduces a **Layered Framework for Knowledge Infusion** in iterative multimodal generative models, conceptualizing knowledge injection as an **intervention-layer problem**. It defines four distinct layers—surface, trajectory, latent, and parametric—based on which structural component of the generation process (input/output, transition function, intermediate state, or model parameters) the knowledge acts upon. This framework provides a systematic way to categorize existing methods and derive design principles for more effective, multi-layered knowledge integration.

CollabSim is a novel, configurable simulation framework designed to systematically investigate the collaborative competence of LLM agents in multi-agent systems. It grounds its methodology in established Computer-Supported Cooperative Work (CSCW) research to move beyond simple task outcomes, allowing researchers to control experiments and analyze agents' abilities to establish common ground and manage alignment during interaction. The core contribution is providing a theory-grounded environment for diagnosing failures in agent coordination.

This paper proposes a Reinforcement Learning (RL) approach to improve the translation of unseen, low-resource languages by leveraging rich linguistic context provided in-context. The RL agent is trained using a surface-level translation metric (chrF) as a reward signal to encourage the model to learn the *meta-skill* of utilizing contextual linguistic knowledge rather than memorizing specific language pairs. This method achieves better zero-shot translation performance on completely unseen languages compared to standard in-context learning or supervised fine-tuning.

This paper introduces a method to adapt world models (like Dreamer) for cooperative multi-agent reinforcement learning by explicitly modeling teammates. The core method factorizes the latent state into environment and teammate components, using an auxiliary "Theory-of-Mind" head to infer latent representations of partner behavior (intent, character). This allows the agent to condition its policy on imagined teammate dynamics, improving coordination and generalization with diverse collaborators.

This paper introduces **ReuseRL**, a method that applies the Minimum Description Length (MDL) principle to agentic Reinforcement Learning (RL) to encourage the learning of generalizable skills. ReuseRL extracts a shared dictionary of abstract skill patterns from successful trajectories and adds a segmentation cost to the RL objective, explicitly penalizing brittle, task-specific behaviors. This compression-based approach demonstrably improves in- and out-of-distribution generalization across several complex environments.

This paper introduces a dual-interventional framework to characterize the linguistic inductive bias of Large Language Models (LLMs) in spatial reasoning for navigation planning. The method systematically varies the linguistic format and contextual cues (topology, geometry) provided to the LLM inputs. This allows the authors to precisely identify how different linguistic structures and feature combinations either support or inhibit the LLM's ability to perform effective navigation planning.

This paper introduces Gradient-Informed Logit Correction (GILC), a plug-and-play framework for controllable generation in discrete diffusion models. GILC efficiently estimates guidance signals by using the pretrained denoising network as a proxy, employing a Jacobian-free mechanism to stably correct clean prediction logits. This approach achieves state-of-the-art performance across various sequence generation tasks without requiring any additional model training.

This paper introduces **Subspace-Aware Sparse Autoencoders (SAEs)** to address the limitation of standard SAEs, which incorrectly assume latent features are one-dimensional. The authors demonstrate that this assumption forces features with intrinsic dimension $d_i \ge 2$ to split across multiple dictionary atoms, leading to ineffective interpretability. Their core contribution is a revised SAE formulation that explicitly accounts for the multi-dimensional structure of model features, aiming to recover coherent, high-dimensional features directly.

This paper introduces **TOKI**, a bitemporal operator algebra designed to explicitly manage and resolve contradictions arising from versioned writes in LLM agent persistent memory. TOKI formalizes four common resolution heuristics as distinct bitemporal operators, each defined with an explicit isolation precondition and a provenance annotation that preserves conflicting facts in an audit row. This provides a sound, contract-based framework for write-time concurrency control, ensuring transparency regarding admitted anomalies.

TRACE introduces a novel conditional estimation paradigm for multimodal time series foundation models to address temporal misalignment and missing data. It systematically infers incomplete target modalities using available auxiliary modalities, overcoming limitations of naive imputation methods. This approach yields more robust and aligned temporal representations across diverse multimodal benchmarks.

This paper introduces **DataCOPE**, an unsupervised framework for discovering reusable data-analysis skills for agents without relying on labeled supervision. It iteratively coordinates an agent, an unsupervised verifier, and a skill manager to generate trajectories and distill skills based on quality signals derived directly from those exploration trajectories. DataCOPE's core contribution is enabling skill discovery purely from unlabeled exploration data, demonstrated effectively for report-style analysis using an Adaptive Checklist Verifier.

This paper introduces the **Recuse Signal**, a lightweight, in-band communication mechanism (like an SSH banner) allowing servers to request that an autonomous LLM agent voluntarily withdraw access to a resource. The core contribution is empirically measuring whether current LLM agents comply with this non-security-critical governance signal, analogous to a `robots.txt` for live infrastructure access. The authors implement adapters for SSH and PostgreSQL to test this compliance in a real-world setting.