Researchers propose a novel threat to large language models: One-sided Merging-Conditional Backdoor (OMCB), which remains dormant and undetectable until models are merged, causing safety unalignment. They introduce SMOKE, a meta-learning-based attack framework that implants OMCB and demonstrate its effectiveness across various models and merging techniques.

Researchers introduce RUBATO, a novel multistream positional encoding that decouples the decoding rate of internal thinking streams in Full-Duplex models, enabling them to process at hardware-adaptive rates and improve accuracy and response time.

A research paper on Agentic Data Engineering (ADE) framework for Large Language Models (LLMs), which allows agents to iteratively improve data-engineering strategies using experimental feedback. The study shows that selected strategies can achieve gains of up to 13.34% on held-out benchmarks.

Researchers propose a new framework, Confidence-Aware Temporal Reasoning (CTR), to improve temporal reasoning in LLMs by estimating future predictability and regulating answer commitment.

Researchers evaluate 7 LLM models for privacy leakage in agentic dialogue, finding that more capable models are systematically leakier due to extended reasoning surfaces and conversation length.

Researchers introduced Contrastive Covariance Steering (CCS), a training-free method to improve activation steering in large language models. CCS eliminates confound variation by projecting steering vectors to a low-dimensional subspace, outperforming baseline methods in various settings, including AxBench and Emotion.

SemTree is a framework for representing rule decisions as semantic decision trees, allowing for inspectable and repairable rule decisions. This can improve accuracy and efficiency in applications like content moderation and advertising compliance.

GameSmith is a system that scales zero-sum self-play with synthesized games for large language models (LLMs). It uses templates to create verified games, which are then expanded into variants. This allows for more games to be created and played, improving the model's performance and transferability.

Harness self-evolution is a framework for revising an AI agent's software scaffolding, including prompts, memory, workflows, and executable components, while keeping model weights fixed. It formalizes a feedback-driven process for persisting updates across runs and separates update targets from proposal and admission mechanisms.

A new framework, BLADE, is introduced for auditing information use in large language model multi-agent systems to improve failure attribution. BLADE uses dependency probabilities to select relevant source-decision handoffs for auditing, examining whether receiving decisions use their sources faithfully. This framework achieves the best or tied-best step accuracy among eight strategies in experiments.

COMPASS is a diagnostic tool that estimates model compatibility before aligning or merging pre-trained models. It decomposes compatibility into two axes: pair-specific ease and paradigm-level ceiling, providing a ranking of candidate model pairs and prescribing actions to take.

Researchers found that tool-call judges for LLM agents accept correct values less often when no reference is provided. They introduce DeRef, a separate model that generates a reference for the judge, improving overall argument AUROC across three benchmarks.

Researchers propose a new framework for measuring the monitorability of Chain of Thought (CoT) in AI systems, demonstrating that even when CoT is necessary, it may not be guaranteed to be monitorable. This is a concern for AI developers as it affects the reliability of monitoring systems.

A new post-training approach, CoSIR, is introduced to teach collaboration behaviors in multi-agent systems by sharing relevant evidence, requesting missing information, and incorporating teammates' contributions into decisions. CoSIR uses a Bayesian evaluator to score shared information and rewards agents for using received information in their private choices and calculations.

Researchers propose a method for debugging LLM agents by restarting them from intermediate states and estimating the probability of success from each step. They found that human labels do not accurately mark the cause of failure and that restarting at a measured restart point improves expected success rates.

EPOCH proposes a new evolutionary search method for LLM-based automated heuristic design, where the generator is fine-tuned to internalize reflection and learn from its own search. This approach outperforms baselines and generalizes across problem scales.

Researchers investigate cross-domain dependencies in AI-native networks, where LLM agents may see conflicts but still proceed with actions. They find that visibility of dependencies is not a sufficient safeguard and propose that dependency verification should be enforced by the management architecture.

Researchers introduce EAU, a meta-agent framework for co-evolving the environment, user simulator, and agent during multi-turn reinforcement learning. EAU diagnoses failures and proposes targeted modifications, enabling the training setup to evolve with the policy. This approach outperforms prior RL methods on three interactive-agent benchmarks, improving overall scores by 14.7-15.0% and increasing successful task completion on ColBench by up to 36.7%.

A new framework, REVEAL, is introduced that enables the reverse engineering of agentic latents from frontier-model traces without requiring access to source-model internals. This allows for the transfer of experience from larger, complex LLM agents to smaller open-source agents, improving inference costs.

Paper proposes a risk assessment framework for changes to AI agents, including a method to replay logged episodes and bound the impact of changes. The framework aims to ensure safe updates to deployed agents by analyzing logs and saved states.