Research

Reinforcement fine-tuning improves LLM social reading more than game wins

Zhang et al. let reinforcement fine-tuned (RFT) large language model (LLM) agents play social deduction games requiring hidden-state inference, social reading, and vote steering. They find that optimizing terminal win–loss outcomes does not reliably produce general social deduction ability, but RFT significantly improves social reading—inferring hidden roles from discussion and predicting others' decisions. Gains in social influence (steering votes, team approvals) require more behaviorally specific rewards.

Generative media

RL-trained cache scheduler adapts video diffusion to user-specified speedups

Authors propose MORCA, a cache scheduling framework that treats cache decisions as a Markov decision process and optimizes them via reinforcement learning. By focusing on terminal error — the quality loss in the final generated video — rather than per-step error, and by allowing users to specify an exact acceleration target, MORCA achieves better speed–quality trade-offs than existing dynamic caching methods on multiple video diffusion models.

Joint on-policy learning improves self-distillation by aligning teacher with student

The authors analyze why self-distillation in reinforcement learning can fail even when the teacher performs well, and derive a condition for the teacher's update to improve the student. They propose Joint On-Policy Learning and Teaching (JOLT), which jointly trains a single policy as a privileged teacher (with KL regularization toward the student) and an unprivileged student (via dense on-policy distillation). Across math reasoning, code generation, tool use, and terminal tasks, JOLT+ outperforms GRPO by 4.2 to 9.9 percentage points and uses substantially fewer tokens on MATH.

Safety & security

Open-MMUnlearning framework unifies MLLM unlearning evaluation across five benchmarks and eight models.

The authors introduce Open-MMUnlearning, an open-source framework that integrates target-model preparation, multimodal data processing, unlearning, and evaluation for multimodal large language models (MLLMs). Using a common evaluation protocol, they compare ten unlearning methods: GD and MIP-Editor tie for the highest overall score, with GD achieving the best forget quality and MIP-Editor preserving the most model utility. A meta-evaluation of thirteen metrics shows BLEU attains the highest aggregate reliability, while KS-Test has the best faithfulness AUC but lower robustness.

Generative media

Position-based self-conditioning improves single-stage 3D generation without extra stages

The authors propose Position Forcing, a self-conditioning framework that recovers token positions from predicted clean latents, quantizes them at progressively finer resolutions according to the noise level, and feeds the positional encodings back into a diffusion Transformer. This provides coarse-to-fine spatial guidance, improving generation quality in single-stage 3D generative models and outperforming several multi-stage approaches.

Local retrieval credit helps search agents learn multi-step evidence gathering

Huang and colleagues examine how intermediate retrieval feedback should be incorporated when training search agents with reinforcement learning. Across seven question-answering benchmarks, assigning evidence credit directly to the search action that produced it outperformed training based only on final answers, raising average F1 by 3.09 points and multi-hop F1 by 4.81 points.

Vision

Patch-level self-supervision scales to seven billion parameters with simpler objectives

Seitzer and colleagues derive a student-teacher learning objective from the idea that useful visual information should be shared across different views of an image. Their resulting method, JEM, trains stably from 300 million to 7 billion parameters and consistently beats a controlled DINOv2 baseline on dense prediction tasks, while remaining competitive on image classification.

Evaluation & benchmarks

Successful Agent Traces Predict Which Base Models Merit Post-Training

The authors propose evaluating base models at the decisive point in successful coding-agent trajectories, avoiding the tool-use failures that make end-to-end agent benchmarks nearly unusable before post-training. Across ten public base and post-trained model pairs, three such evaluations closely tracked eventual SWE-bench Verified pass@1 performance.

Generative media

Audio training teaches one model to edit scenes across modalities

Chen and colleagues developed CrossEdit, a unified model for editing images, audio and video from free-form instructions. By combining procedurally generated audio edits with self-supervised audiovisual reconstruction and selected cross-modal tasks, the model edits short movie scenes without direct training on that task, changing both sound and picture while seeking to preserve unrelated content.

Language models

Adaptive token computation narrows diffusion language models’ quality gap

Fang et al. modify diffusion language models so that different masked tokens can receive different amounts of computation during each denoising step. At 1.7B and 8B parameters, their models achieved higher average scores across eleven reasoning, mathematics and coding benchmarks than all evaluated diffusion models and corresponding autoregressive baselines, while the 8B model retained a reported speed advantage on GSM8K.

Shared agent memory can leak information across enterprise users

Mudgal et al. study enterprise AI agents that store multiple users’ long-term memories in a shared vector database and rely on similarity search to retrieve relevant records. Across six controlled experiments, they show that ordinary queries can retrieve semantically related memories owned by another user, even without prompt injection or system compromise. Filtering retrieved records by ownership was the only tested mitigation that consistently returned response contamination to the clean baseline.

Reinforcement learning reshapes which reasoning strategies language models prefer

Cheng and colleagues studied reinforcement learning across Qwen and Gemma models, comparing performance and sampled solution traces before and after training. They find that RL changes which reasoning strategies a model tends to select, producing cross-domain transfer on some tasks but forgetting on others. This reweighting can also raise one-shot accuracy while reducing coverage over 1,024 attempts.

Preference-grouped reinforcement learning improves social adaptation in simulated conversations

Wang and colleagues introduce preference-batched group relative policy optimization, or PB-GRPO, for training conversational agents to infer and accommodate users’ unstated preferences over multiple turns. In a persona-driven simulation, the method outperformed vanilla GRPO and other reinforcement-learning baselines on an LLM-judged measure of user satisfaction, although the supplied paper excerpt does not report effect sizes.

System that builds agent teams per query outperforms all baselines

The authors introduce SHIFT, a framework that constructs a multi-agent harness—the combination of agent roles, instructions, and tools—for each query without executing candidate designs during search. Across 9,193 tasks on six benchmarks, SHIFT-search achieves 79.9% mean accuracy, outperforming 17 baselines by 7.2 percentage points, while a cheaper variant, SHIFT-value, exceeds all baselines with 32% fewer execution tokens.

Terminal-only control improves digital agents across web, desktop and CAD

Lu and colleagues tested whether computer-use agents benefit from replacing fixed, GUI-centered toolsets with a programmable terminal interface. Across six web, desktop and computer-aided design benchmarks, their roughly 3,000-line CUAWright harness generally improved task performance, robustness and estimated cost efficiency over GUI-only, hybrid and other command-line systems.

Forecast errors can improve event-aware time series predictions

Tan et al. present SEER, a forecasting framework that supplements numerical histories with external events and learns from its previous errors. Across six volatile benchmarks, the authors report that this closed-loop approach outperforms numerical time series foundation models, deep sequence models and language-model baselines, although the supplied material does not include effect sizes.

Theory

Anchor distributions define context-specific similarity in contrastive models

Kannan et al. demonstrate that contrastive representations (e.g. CLIP) contain a family of semantic geometries, each determined by a probability distribution over semantically relevant anchors. They derive 'Anchor Divergences' as Bregman divergences on fixed embeddings, enabling efficient context-specific similarity without changing the underlying model.

Simple credit shaping boosts exploration in language model reinforcement learning

The authors introduce ExPPO, a lightweight modification to GRPO that reshapes advantages using prompt-relative, length-normalized response surprisal and prompt pass rate. ExPPO preserves verifier polarity while assigning higher credit to rarer correct responses, leading to better exploration. Experiments across three language model backbones show improved pass@K, sustained entropy, and increased correct-mode diversity.

Safety & security

New method discovers language model circuits while accounting for component interactions

The authors introduce witness-integrated set effects (WISE), a family of causal estimands that average effects over component sets while blocking self-repair, and JuntaLearner, a gradient-based circuit discovery method that learns to rank components by their causal impact across varying set sizes. On five tasks across GPT-2 small, Qwen3-1.7B, and Llama-3.1-8B, JuntaLearner outperforms attribution baselines on measures of necessity and task specificity, and scales to large models without combinatorial cost.