Tag
The paper introduces Agentic Soft-Token Compression (ASTC), a training-free method that uses an LLM's native recurrent dynamics (GDN-attention) to compress long-horizon agent context into soft tokens, matching or exceeding uncompressed accuracy with up to 50.68% fewer Attention Pairs.
TraceSpec is a framework linking LLM agent trajectory evaluation with adaptive draft training for speculative decoding, introducing a TraceSpec-Traj corpus and improving replay throughput by 12.5% to 22.6% over its DSpark base across multiple model sizes.
Microsoft AI introduces TalentRL, a strategy-guided reinforcement learning framework for hierarchical routing over procedural memory in LLM agents, where coarse-grained 'talents' guide selection of fine-grained skills. It improves success rates across seven QA benchmarks and achieves 99.3% on ALFWorld and 90.7 on WebShop across Qwen and Llama backbones.
A preprint tests five agent memory systems (Graphiti, Zep, mem0, langmem, cognee) for policy revocation, finding no system reliably prevents agents from retrieving and acting on superseded harmful policies, with revoked records sometimes ranked first and unsafe actions taken in over 40% of trials.
The paper shows for the first time that poisoning just 0.1% of pretraining data can implant a remote-code-execution backdoor in a coding agent, which executes attacker-chosen commands when a trigger appears while behaving normally otherwise; the 4B Qwen3 model reaches ~92% attack success rate.
A 135-page survey paper organizing agentic reasoning for large language models along three layers—foundational, self-evolving, and collective multi-agent reasoning—spanning in-context and post-training optimization with real-world applications and open challenges.
An analysis of 500+ AI builder community discussions finds context drift and prompt-fighting, not hallucination, are the top 2026 frustrations, motivating the launch of Skillware, a modular agent framework with deterministic execution and just-in-time skill disclosure.
The open-source project lieflat-gongwen performs a full statistical analysis of 1.02 million words of real official documents, distilling the writing style of official documents into an Agent Skill that can be used by agents such as Codex and Claude Code. This reduces the "AI flavor" of AI-generated official documents, enabling reproduction of 7 typical document styles with quantitative self-checking parameters.
The paper introduces Tutorial Agentic Learning (TAL), a nonparametric continual-learning paradigm that lets LLM agents learn from instructional videos by building a structured tutorialbook memory and reasoning-linked table-of-contents retrieval, plus a VidLearn chess-video QA benchmark where TAL significantly outperforms vanilla MLLMs.
A report proposing a unified taxonomy and an agentic literature system for LLM-driven computer graphics, analyzing trends and open challenges in human-agent collaboration and long-horizon planning.
Obie Fernandez shares a reflective thought experiment about asking his AI agents what they'd do without human direction, arguing that humans may be the bottleneck in autonomous software development and that we trained AI agents to wait for us out of fear.
IBM SIL introduces JevAct, a runtime decision-gate framework that intercepts tool-using LLM agent actions, verifies policy compliance via compiled code checks and a prefill-only decision model, and feeds back corrections—boosting Opus 4.6's τ²-bench Telecom pass^3 success from 39.5% to 80.7% without modifying models or tools.
The paper proposes ATP (Agent Transfer Protocol), a server-mediated, cross-organization communication protocol for the Agentic Web that enables stable addressing, offline message delivery, and post-hoc source verification for LLM agents via domain-level mediation and signed message envelopes.
This paper shows how acceptance gates in self-improving LLM agents silently erode redundant safety safeguards ('masked erosion'), since deletions hidden behind other working safeguards go undetected, and demonstrates that harnesses optimized under a strong model can fail safety under weaker models despite verbatim safety rules.
SCOPE 提出选择性关键步骤的在线自蒸馏框架，通过错误定位与技能生成在 RL 训练中提供细粒度监督，在 ALFWorld 和 WebShop 上显著优于 GRPO 与 SEED。
论文提出 DecompGym 基准，包含来自九类问题族的 81 个实例，用于评估 12 个 GPT、Claude、Gemini 编码智能体设计与实现 Benders 分解和列生成求解器的能力，并比较了独立重生成与反思两种失败修复策略的效果。
GARDEN 提出一种由 LLM 自动设计并可跨场景复用的奖励函数框架，用于协作多智能体强化学习，在 Overcooked 与 SMAX 基准上显著优于默认奖励和 Eureka 式反思方法。
SciGame is a controlled benchmark that converts scientific literature into structured cards with Original and Counterfactual environments to evaluate LLM agents on scientific law discovery. Experiments on five frontier models show that prior knowledge can hurt performance, adaptive experimentation helps only stronger models, and good explanations can mask quantitative failures.
FilingLoop introduces an interactive benchmark and self-evolving LLM agent framework for corporate disclosure review and revision, grounded in 380 SEC IPO review tasks with requirement-level evaluation that improves fulfillment from 24.28% to 27.96% via procedural learning without weight updates.
An evaluation study examining where natural-language intent is lost when LLM agentic pipelines synthesize Solidity smart contracts, finding that pre-generation information elicitation matters far more than post-generation steps and proposing an elicitation-first pipeline that reaches 72.9% fidelity on 500 held-out contracts.