Tag
This paper evaluates frozen video representations and finds that latent prediction errors in video world models reflect appearance and motion rather than reliably measuring physical surprise, presenting a framework that separates representation content from error signals across IntPhys2, Physion++, and Kubric benchmarks.
CEPO is a general framework for prompt optimization in multi-agent systems that separates shared system-level coordination from agent-specific execution prompts, consistently improving four existing prompt optimizers across 11 benchmarks in four domains.
This paper proposes two measurable preconditions—verifier strength and recoverability—that determine when an agent platform can safely run self-improvement steps without human oversight, validating them on a deployed platform and the τ²-bench benchmark.
OpenAI 发布新模型能力，显著提升在 COBOL、Fortran、Visual Basic 等低资源编程语言上的构建与迁移表现：GPT-6 Astra 在 MultiLangBench 20 种语言上平均通过率达 87.4%，GPT-6.1 Sol 的 COBOL 迁移准确率最高达 97.9%，且成本更低。
HarnessPro 提出一种无标签的递归自改进方法，通过轨迹诊断与加权 Pareto 选择生成 lite/medium/heavy 三档 harness，在 Terminal-Bench 2、SWE-Bench Pro 等基准上相比 RHO 平均提升 2.0 个百分点通过率并降低 26.1% 单任务成本。
Oracle AI 和宾大的研究者对数据智能体中的三种上下文工程技术（语义层、编排、压缩）进行系统研究，发现这些技术的效果高度依赖智能体的行为分组（调查者、过早作答者、早期放弃者），某些情况下甚至会损害准确率。
RSIFlow 提出一种递归自我改进框架，通过诊断任务执行反馈动态选择组件级改进工作流，并演化元智能体以优化编排能力，在六个基准上达到 SOTA 平均性能。
This paper investigates agentic retrieval that combines LLM reasoning with dense retrievers in a ReAct loop for complex search tasks, showing an 8.7-point nDCG@10 improvement over standard retrieval and strong generalization on ViDoRe v3 and BRIGHT leaderboards, but at a substantial latency and token cost of 107.4 seconds and 764K input tokens per query.
The paper identifies why self-evolving reasoning models suffer performance collapse — invalid questions and mathematically-equivalent question repetition — and proposes R-Quest, which uses validity and novelty feedback to sustain self-evolution, outperforming R-Zero by 17.32 points across 12 benchmarks over ten rounds.
Kepler is an open-source harness for ARC-AGI-3 that represents hypotheses as executable world models and validates them via retrospective transition and prediction checks, achieving a server-verified 100.00 RHAE on all 25 public games under a frozen Claude Opus 5 configuration. The paper also reports three evaluation failures — source-code leakage, harness reconstruction, and autonomous repair masking a broken planner — and argues that public-set scores alone have limited discriminative value, motivating first-attempt, cost-conditioned, and verification-aware reporting.
JevSpawn introduces a compositional policy that maps natural language task specifications to finite probabilistic exploration, using parallel action spawning and feedback-driven branch selection to speed up and improve LLM agent inference without additional training, evaluated on eight benchmark tasks against seven baselines.
该论文研究了参考答案（gold reference）的选择对多语言基准评测分数的影响，发现交换参考标注可使F1分数移动数个点，并揭示了一个未记录的标注聚合默认设置导致参考答案与被评输出部分重合的问题，提出应将参考敏感性区间与分数一同报告。
The paper shows that keyword-matching benchmarks can credit small language models with tool use they never perform, documenting a false positive in matched-architecture Spanish security LMs and proposing a cheap diagnostic ladder plus a targeted SFT recipe that repairs the failing 1B model.
该论文提出 RASO（检索增强技能优化）框架，通过 Cross-Harness Adaptation 从外部技能语料库中检索先验知识来初始化和迭代优化 agent 技能，在四个 agent 基准和两个模型上显著优于无检索增强的基线方法。
The paper presents TW3Cast, a time-series forecasting system that uses a frozen router of lightly fine-tuned foundation models to achieve top performance on the GIFT-Eval benchmark without agents or language models.
This paper audits a multilingual affective generation benchmark and reveals that system rankings are driven by measurement artifacts rather than genuine performance differences, with annotator variability playing a key role.
Vals AI evaluated Grok 4.7, finding it ranks #24 on the Vals Index with a score of 54.2%, down from Grok 4.6, but shows improvements in legal and medical domains.
This paper presents RoboDawn, a method to transfer Vision-Language Model intelligence to robotic control, achieving state-of-the-art results on benchmarks with zero-shot and one-shot learning and successful real-world applications.
A Yale-led paper finds that frontier AI models' poor performance on physics benchmarks is largely due to benchmark flaws rather than model limitations, suggesting benchmark quality is a critical bottleneck for accurate evaluation.
The paper proposes CLEAR, an agentic framework for cross-source evidence adjudication to improve large language models in medicine by handling conflicts from multiple knowledge sources. It demonstrates competitive performance across benchmarks, with significant gains in settings where direct inference or retrieval is weak.