A new framework automates prompt refinement for LLM agents by splitting the observation-to-action pipeline into goal-conditioned and action selection modules. It uses an LLM-driven evolutionary loop to iteratively improve prompts based on environment feedback, achieving up to 72.5% success on PutNext where prior agents failed, without model fine-tuning.
Dynamic Rollout Editing (DRE) addresses overthinking in RL-trained reasoning models by modifying successful trajectories post-answer emergence. DRE preserves the correct reasoning prefix while editing unnecessary continuation, weakening the credit assigned to redundant thinking without penalizing valid reasoning. Experiments across diverse tasks demonstrate its effectiveness in reducing overthinking.
TokenPilot reduces inference costs by 61% to 87% in both isolated and continuous modes, outperforming prior systems in cost efficiency while maintaining competitive performance. It uses ingestion-aware compaction and lifecycle-aware eviction to preserve prompt cache continuity and minimize token footprints.
MetaSyn introduces a dataset of 442 expert-curated meta-analyses from Nature Portfolio. It evaluates twelve LLM agent configurations and reveals a critical bottleneck in study screening, where no system recovers more than 52.7% of ground-truth included literature despite high retrieval recall.
Qwen3-8B internally tracks the value of its current trajectory, defined as the likelihood of achieving its goals. This 'value' axis distinguishes confidence levels, backtracking behavior, and code correctness, and shows that preference optimization boosts confidence in rewarded behaviors. The model assigns low value to politically sensitive queries post-training, and fine-tuning increases confidence within specific domains.
ActiveSAM is a training-free, zero-shot framework that enhances SAM 3 for open-vocabulary semantic segmentation by identifying an image-conditioned active class set. It improves speed-accuracy tradeoff, outperforming SegEarth-OV3 by +1.4 mIoU on average and running up to 5.5x faster on large-vocabulary datasets, with strong robustness under image corruption.
ActiveSAM is a training-free, zero-shot framework that enhances SAM 3 for open-vocabulary semantic segmentation by identifying an image-conditioned active class set. It improves speed-accuracy tradeoff, outperforming SegEarth-OV3 by +1.4 mIoU on average and running up to 5.5x faster on large-vocabulary datasets, with strong robustness to image corruption.
ExpRL introduces a novel mid-training approach for LLMs using human-written question-answer data as reward scaffolds. Instead of imitating reference solutions, it constructs problem-specific grading rubrics to reward intermediate reasoning steps, enabling better initialization for sparse-reward RL and outperforming SFT, sparse-reward GRPO, and self-distillation on math reasoning tasks.
Hierarchical Advantage-Weighted Behavior Cloning (HABC) enhances online RL fine-tuning of vision-language agents by using separate critic heads for viability and efficiency. It combines their outputs via a state-adaptive gate and applies per-transition weights, while intervention-aware credit assignment prevents supervision leakage. In real-robot experiments, HABC boosts success rates to 92%, 88%, and 38% on three bimanual tasks, surpassing SFT baselines of 36%, 44%, and 12%.
HalBench evaluates 29 open-source LLMs on a custom benchmark for sycophancy and hallucination. Qwen 3.6 and Gemma 4 outperform larger models, with Qwen 3.6 achieving 36.6% pushback—higher than GPT-5.4 and Gemini 3.1 Pro. Model size does not correlate with honest responses, indicating that architecture and training data matter more than parameters.
LLMs can detect cultural cues and recall cultural conventions, but often fail to adapt responses accordingly. Their responses remain biased toward their native culture unless explicitly prompted to apply cultural context sequentially.
EComAgentBench introduces a benchmark of 662 real Amazon tasks that scatter shopper requirements across query, profile, and clarification. Agents must uncover hidden intent, verify candidates with evidence, and commit to a product within 100 tool calls, with typed rubrics attributing failures to specific requirement sources. Evaluation shows even top models achieve only 57.1% accuracy, and rubric satisfaction drops when intent is hidden.
Current coding benchmarks were designed before agentic software engineering and fail to capture the complexity of real-world systems. They conflate model performance with the entire harness, ignore valid alternative solutions, and lack feedback signals at individual component levels, making iterative improvement difficult.
DIFE evaluates backdoored CLIP checkpoints across different deployment interfaces, revealing that native success does not guarantee safety in reuse. The framework shows text-side poisoning enables adversarial exposure in retrieval, reranking, and selection tasks, while visual-only use remains largely unaffected. BadTextTower is introduced to generate strong text-conditioned exposure without compromising visual performance.
We present a framework for evaluating agentic skills by constructing realistic tasks and assessing skill utility through task execution. Applied to 500 real-world skills, it generates 1,000 tasks and scoring rubrics, evaluating 19 agent-model configurations across proprietary and open-source models. Results show significant variation in instruction adherence and performance gains, with skills substantially altering model behavior compared to no-skill setups.
GameCraft-Bench introduces a benchmark with 140 Godot tasks across 15 game families to assess coding agents' ability to generate playable games. Evaluations show the best agent achieves only 41.46% success, indicating significant challenges in producing complete, interactive games with coherent gameplay and visual feedback.
ChLogic evaluates how well large language models maintain logical reasoning when English logical structures are expressed in Chinese. It reveals a persistent English-Chinese performance gap, with back-translation improving results on general items but harming performance on difficult problems. The benchmark highlights the impact of surface realization, translation artifacts, and model-specific behaviors on multilingual reasoning.
NNN decoding selects documents as a joint set that jointly reconstructs the query embedding via a sparse non-negative linear combination. It strictly extends dense retrieval by handling queries that dense retrieval fails on, especially in corpora with correlated documents, and achieves superior performance through end-to-end training of embeddings.
VoidPadding introduces [VOID] as a padding token to separate semantic termination and response-length modeling. It improves performance on mathematical reasoning and code generation by 17.84 points over the original model and reduces decoding NFE by 55.7% on average.
Evalatro is an open benchmark that allows LLMs to play the actual game Balatro. Models receive game state as text, make decisions independently, and compete to reach Ante 12, with current results showing limited progress—mimo-v2.5-pro reached Ante 5, and deepseek-v4-pro failed to beat Ante 8.