Researchers present TurnCost, a tool for re-scoring trajectory corpora of coding agents to optimize caching, pricing, and routing policies. The study analyzes 600 publicly released coding-agent trajectories and finds that turn count explains 84% of the variance in log trajectory cost, making unnecessary turns the larger opportunity for optimization.

Researchers propose a user-product relevance metric for offline evaluation of Dynamic Product Advertising (DPA) retrieval systems using large language models (LLMs) as semantic relevance judges, enabling faster and more reliable iteration before A/B testing.

Researchers found that natural code transformations can be used to 'confuse' AI agents, such as Codex and GPT-6 Sol, in vulnerability analysis, reducing their success rates. This study explores the potential of using code transformations as a defense against malicious agents.

A new model, Elastic V-JEPA (EV-JEPA), is introduced for flexible and collaborative inference across heterogeneous devices. It can be reconfigured to optimize compute budgets and supports cross-device execution. EV-JEPA achieves high accuracy and improves upon previous elastic training approaches.

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.

A new paper introduces Multi-Ratio ConceptMoE, a model that supports multiple compute budgets across turns in multi-turn interactions, improving compute efficiency and accuracy.

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.

GF-Bench, a PyTorch-based framework, is introduced to generate diverse failures in distributed GPU workloads. It evaluates log analysis methods, revealing a trade-off between effectiveness and efficiency, and highlights the importance of preserving execution context for accurate failure detection.

Researchers developed a harness to learn interpretable models from data using LLMs, outperforming state-of-the-art rule-learning methods on tabular and text data.

A research paper proposes a method called Privileged Representation Isolation for On-Policy Self-Distillation (PreSO) to improve reinforcement learning with verifiable rewards. PreSO uses compact hints to limit solution-specific detail and estimates a leading low-rank subspace from pooled problem shifts to improve representation-level distillation.

VH-GPO is a new method for verified handoff in LLM agent reinforcement learning, which improves success rates by up to 35% compared to existing methods like GRPO. It uses a teacher to provide enough support to reach a state where the student can succeed, then returns control to the student to ensure successful experience is student-generated.

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.

Researchers introduce an Improvement Dynamics Model (IDM) that predicts task-level outcomes of proposed modifications for self-evolving agents, allowing for more efficient and effective learning.

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 propose an outcome-grounded confidence calibration method for LLM agents to predict their own failure. They introduce a proper scoring rule and a separate evaluation protocol to assess the effectiveness of the method. The method improves the agent's confidence in predicting its success and reduces expected calibration error.

A new method, Structured Recipe Optimization (SRO), is introduced to adapt pre-trained LLMs to diverse domains by constructing and optimizing a recipe of coupled decisions. This paradigm can adapt the recipe space to the target tasks and represents choices explicitly for systematic reuse and recombination.

Researchers propose ReTerm, a self-improving terminal agent that can recover from errors during long-running execution. ReTerm is trained to learn from its own failed tasks and uses a customized data pipeline and online RL algorithm to improve its recovery capabilities. The model achieves a 14.6 point improvement over the base model on Terminal-Bench 2.0.

WorkRover is a framework for training long-horizon cowork agents through recursive environment evolution, enabling iterative refinement of construction agents and their components while keeping them aligned.

SHELDON is a scientific reasoning harness that separately represents and jointly explores hypothesis and observation spaces, allowing it to discover new observations and generate new hypotheses beyond existing ones. It outperforms baselines by up to 26.7 points and generates up to 4.5x more novel solutions.