Three voices,
and it becomes
news.

Retrieval quality depends heavily on how effectively index keys expose knowledge, but fixed optimization strategies struggle across diverse environments, leaving index evolution largely human-driven. We propose SELF-INDEX, a framework enabling an index to self-evolve without human intervention. Its Optimizer autonomously diagnoses retrieval shortfalls, selectively revises responsible index keys, and validates each revision prior to updating. Furthermore, a Query Simulator proactively explores additional demands beyond available queries. Across diverse corpora and retrievers, SELF-INDEX consistently improves retrieval performance and enhances the effectiveness and efficiency of search agents and memory systems.

Multi-turn agents trained with reinforcement learning often use on-policy distillation (OPD) with a privileged teacher for dense token-level supervision, but privileged info is not always reliable and teacher benefits are stage-dependent. We propose RetireOPD (Self-Retiring On-Policy Distillation), optimizing a decoupled teacher with environment rewards and training a student jointly with RL and OPD. Through Adaptive Retirement, the student drops the teacher once discrepancy stops shrinking and it reaches a target success rate, proceeding with RL alone. Across Qwen2.5 models from 1.5B to 7B, RetireOPD improves ALFWorld success rate over RL baseline by 14.1% to 18.8% and WebShop accuracy by 11.8% to 19.0%.

The authors propose EvoSkill-GUI, a training-free framework that enables GUI agents to dynamically update and refine their procedural skills through a reflect-revise-reuse loop during deployment. Each skill is structured as a multi-file package containing retrieval metadata, execution plans, and failure-recovery rules. An executor performs instant in-rollout revisions while an isolated critic diagnoses failures. Experiments across MobileWorld, AndroidWorld, and OSWorld benchmarks demonstrate consistent performance improvements over base models without any additional training.

The paper introduces Discovery Intelligence and Discovery Foundation Models (DFMs) as general-purpose systems for open-ended discovery. The authors instantiate this framework via Zetema—which couples research-state dynamics, verification, and skill evolution—and GALILEO, a closed physical discovery loop integrating Dry-Lab reasoning with robotic Wet-Lab experimentation. This approach enables discovery behavior to be trained, executed, and evaluated beyond final-answer performance.

The paper investigates the impact of candidate-generation schedules on LLM test-time scaling, demonstrating that candidate count N alone is insufficient to describe system costs. The authors compare four schedules (1x8, 2x4, 4x2, and 8x1) while keeping N = 8 fixed. Results on A100 GPUs show that eight serial calls consume 4.64-4.86x as much gross GPU-device energy and exhibit 5.77-6.12x the P95 latency of a single batched call. Consequently, when candidates are independent and memory allows, fewer generation calls with larger batch sizes are more efficient.

The paper examines the convention of applying loss only to agent-authored action tokens during supervised fine-tuning (SFT) and introduces ActObs to supervise observation tokens as well. By training the model to predict action consequences without adding data or parameters, this method prevents action and observation gradients from becoming orthogonal. On Qwen3-4B, GRPO from ActObs achieves higher pass@k on Terminal-Bench 2.0 than its action-only counterpart. On Qwen3-8B, it increases pass@16 by +3.4 pp and extends advantages to cross-domain code editing on aider-polyglot with +4.2 pp at pass@1 for 4B. Joint supervision retains more entropy during RL.

The authors introduce ScienceBuddy, an interactive scientific research workspace implementing recursive-in-recursive self-improvement to continuously adapt AI agents. The paradigm couples harness evolution with model reinforcement learning: an inner recursion optimizes the execution harness with fixed model weights, while an outer recursion trains the model using the improved harness. Case studies spanning four scientific task families demonstrate how this joint evolution drives discovery intelligence in sustained collaboration with human researchers.

The authors introduce HazardAuditor, an execution-grounded framework that supervises heterogeneous computer-use agents like Claude Code, Codex, Hermes, and OpenClaw by normalizing their runtime interactions into a canonical event representation. They observe that token-level post-training objectives create a structural mismatch for generative guards, causing longer rationales to dominate gradient updates. To address this, they propose Guard Policy Optimization (GuardPO), which converts deterministic safety outcomes into sequence-level advantages and normalizes rationale and verdict regions. Across benchmarks, HazardAuditor improves accuracy by up to 16.5 percentage points over the strongest prior guard.

The authors introduce Grouped Value Attention (GVA), a KV cache optimization technique that stores grouped values and reconstructs content keys using a learned linear map. At inference, this map can be absorbed into the query, eliminating the need to materialize content keys in the decode path while a small shared RoPE channel retains positional information. At the 350M-parameter scale with 30B FineWeb-Edu tokens, the 16-dimensional positional variant achieves 44.18 average accuracy across five tasks, compared to 44.36 for GQA and 43.88 for MLA, while reducing persistent cache scalars by approximately 45-47% relative to matched GQA.

This paper presents LynnReal-Omni, a native multimodal video generation framework built on a 32B shared multimodal diffusion transformer that unifies text-to-video, image conditioning, structural control, and long-video generation from heterogeneous visual inputs. It also features a dedicated 27B Flash transformer for real-time rendering and an MSAVP evaluation design with 100 prompts and 20 metrics. On one H100, warm generation and decoding of a 22-frame 540p video take 843 ms with LynnReal-Omni and 377 ms with Flash.

The authors present Video DeltaNet (VDN), a hybrid attention architecture designed to alleviate computational bottlenecks in video diffusion models. VDN's linear branch introduces Video Delta Attention to update memory per frame using spatial tokens, while keeping Softmax attention for text and audio interactions. Utilizing a staged teacher-alignment recipe, an eight-step distillation, and an optimized SGLang serving stack, VDN-H3 efficiently completes denoising for long high-resolution video streams on multiple NVIDIA B200 GPUs.

The authors propose EvolveTrade, a self-evolving framework for LLM trading agents that treats the system prompt as a text-parameterized policy. After each update interval, a Policy Agent revises this policy using accumulated decision traces and portfolio feedback while keeping the underlying LLM backbone fixed. Experiments across diverse market regimes demonstrate that EvolveTrade consistently improves the Sharpe Ratio and Cumulative Return compared to fixed-policy LLM baselines.

The paper addresses the inefficiency of Large Reasoning Models, which often overthink easy problems and underthink hard ones. The authors propose When2Think, a post-training framework that dynamically allocates computation based on problem difficulty. The method introduces Instance-level Difficulty-Aware Control (IDAC), a reward-shaping mechanism leveraging pre-computed reference statistics. Experimental results on AIME24 show that Pass@3 increases by 10.0% while token usage is reduced by 27.9% relative to the base model.

The authors introduce PACT (Pressure-Applied Compliance Testing), a benchmark evaluating rule-following in enterprise AI assistants across twelve regulated domains and forty-eight multi-turn scenarios. Testing twenty-two common LLM models under various pressure tactics, the study profiles compliance robustness and introduces PACTScore. Results show that even top assistants mis-apply rules on 6 to 10% of items, and ordinary user pressure increases violation rates by 65% on average.