Researchers introduced JOINGR, a join-aware table retrieval method for Text-to-SQL, treating the database join graph as the retrieval space. It selects semantically similar anchor tables, traverses join edges, and aggregates scores. JOINGR improves recall over baselines on the BEAVER benchmark and transfers across benchmarks.

The authors propose the MCRI Framework, a four-dimensional framework for analyzing and evaluating agent skills, and operationalize it as MCRI-Eval, a large language model-based evaluation method. They evaluate MCRI-Eval on 63,812 public skills from the OpenClaw skill Hub and achieve promising results, including improved skill selection and ranking agreement.

This paper introduces Revision-Aware Independent Agent Graphs (RIAG), a policy that enables dynamic task routing and manages document versions for AI agents. The authors evaluate RIAG on six benchmarks, achieving higher accuracy and lower call counts compared to existing methods.

MemFit is a long-term memory system for conversational agents that reduces memory construction time and cost by storing each turn verbatim in an append-only store, using an LLM-free insertion and indexing strategy.

A study on the effectiveness of using off-the-shelf small language models (SLMs) in agent harnesses, finding an eligibility gap for microtasks involving large language models. The study proposes a benchmark and analysis framework to assess SLMs and suggests using a baseline that meets a context-informed (CI-backed) threshold.

CompoWorld is a compositional environment scaling approach for training general agents. It composes a library of reusable services, enabling agents to connect information and actions across multiple services. This approach improves performance on benchmarks, surpassing frontier models.

This paper proposes PACE, a model-native VLM policy that learns to optimize long-horizon reasoning in AI agents by dynamically determining the execution depth.

Cascadia is a control-plane-free system for serving large language models on commodity hardware, using a libp2p QUIC mesh for peer-to-peer communication and a certificate authority for admission and fleet management.

Researchers introduce E2E-SWE, a benchmark for evaluating the ability of LLM-powered coding agents to build complete, functional software repositories from scratch. The benchmark contains 186 tasks across 11 programming languages, with varying levels of success across 13 frontier models.

SkillFM generates textual skills for LLM agents via latent flow matching, improving skill synthesis and reducing reliance on manual curation or indirect reinforcement learning. It achieves state-of-the-art performance in embodied tasks and question answering.

Researchers introduced a new model-free controller called Decode-Latency Feedback Prefill (DLFP) to reduce interference in concurrent autoregressive inference. The controller adjusts prefilled chunks based on observed latency and achieves a 27.7% reduction in P99 inter-token latency on a 0.6B Qwen3 model. However, the mechanism does not generalize to larger models or multi-GPU configurations.

This paper proposes a framework for self-evolving harnesses, where a language-model agent improves its own code organization and execution control. The framework uses a recursive self-improvement process, where the frozen model solves tasks and then edits its own harness based on run records. The results show improved performance on in-distribution and out-of-distribution tasks, surpassing or matching Codex on some benchmarks.

HARISSA is a method for local language model deployment that makes decisions on whether to spend more computation on a query or deliver a potentially incorrect answer. It uses the model's own hidden states to make these decisions, improving efficiency and safety.

The paper proposes KV-Kaizen, a method for learning context-adaptive cache compression choices for Large Language Models (LLMs). This approach can reduce memory usage without compromising accuracy, enabling the use of larger models and improving inference performance.

Researchers introduced BudgetPM, a resource-allocation formulation for external observations required by stored intentions under a shared episode budget. BudgetPM offers two policy variants, BudgetPM-Static and BudgetPM-Sequential, which outperform adapted memory-agent systems and hand-designed monitoring rules in two benchmarks.

AnyAct proposes a universal action layer for self-evolving agents to tackle challenges in large-scale, dynamic tool ecosystems, including the scale dilemma, non-stationarity, and heterogeneity of feedback formats.

This paper proposes CADOC, an online algorithm for compressible context management in long-horizon agents. It schedules replacements of structured objects with compact Cards, preserving exact on-demand retrieval of original contents, and achieves a 40% reduction in input cost on average while maintaining task performance.

The paper proposes Last-Chance Policy Identification (LCPI) and Risk-Budgeted Compatibility Planning (RBCP) for agents to handle irreversible resource depletion. LCPI formalizes a principle to distinguish or homogenize fault models through agent actions, while RBCP searches a compatibility-aware frontier under a hard failure constraint.

This paper presents a causal mechanistic audit of a self-discovered reinforcement learning (RL) rule, analyzing its internal update machinery and learning history. The authors test whether learning history acts as an asset or a burden, finding that it actively expands usable reward scales and can be a burden due to perpetual clamping. This work establishes a foundational audit standard for next-generation, self-evolving RL algorithms.

A new framework, Think Short, Defer Smart (TSDS), is proposed for edge LLM agents. TSDS integrates a lightweight convergence probe with a perplexity-based deferral rule to manage reasoning budget and defer to a cloud-side model when uncertainty is high. This results in significant reductions in per-episode thinking compute and maintained guarantees on reward and cloud-call rate.