Agents

LLM agents and the systems around them: harnesses and orchestration, tool use and MCP, multi-agent systems, agent memory and planning

600 articles · page 5 of 13

Agents

Multi-objective evolution yields stronger, more transferable agent harnesses

The authors introduce Harmony Harness, a framework that treats LLM agent harness evolution as a multi-objective optimization problem using decomposition-based evolutionary search. They also separate the harness into general-purpose mechanisms and environment-specific adaptations, composing them at inference time. On SecRepoBench and BaxBench, Harmony Harness consistently outperforms baselines, with gains exceeding 10%.

Agents

Co-Adaptive Routing and Retrieval Learns Query Rules Without Supervision

Qi et al. propose Co-Adaptive Rule Routing and Constraining Retrieval (CoRE), a framework that jointly learns which reasoning rule to apply to a query and how to retrieve evidence satisfying that rule, without explicit query-rule supervision. On STaRK-QA benchmarks across scientific discovery, biomedicine, and e-commerce, CoRE achieves the strongest average retrieval performance, and ablations confirm that co-adapting the router and retriever is essential.

Agents

Agent memory benefits more from event aggregation than retrieval alone

The authors diagnose two distinct bottlenecks in agent memory, access and aggregation, and show that retrieval alone fails when answers require combining many events. They propose Trajectory Memory for lossless, question-time evidence access and TallyMem for deterministic aggregation of event specifications. TallyMem improves exact match on held-out MINTEval repositories from 35.2% to 49.0% over Trajectory Memory.

Agents

Tool discovery paradigm crossover: Retrieve-then-Plan wins on simple queries, Plan-then-Retrieve on complex ones at scale

The authors present a head-to-head comparison of two paradigms for tool-augmented LLM agents: Retrieve-then-Plan (RtP), which embeds the user query and retrieves similar APIs, and Plan-then-Retrieve (PtR), which first generates abstract tool stubs (hypothesized function signatures) to use as retrieval queries. Testing on ToolBench with 46,984 APIs across single-tool, intra-category multi-tool, and cross-category multi-tool queries, they find a complexity-dependent crossover: RtP outperforms PtR on single-tool queries, while PtR outperforms RtP on multi-tool queries. This crossover is robust across retrieval budgets, registry sizes, and model families.

Agents

Video agents improve their own harnesses using unlabeled video

Xu and colleagues present a framework that turns unlabeled videos into supervision for improving video-understanding agents. A Questioner creates questions and reference answers from structured video memories, while a Solver uses execution feedback to revise its skills, tools, parameters and coded workflows. The authors report progressive accuracy gains across evolution rounds and transfer improvements across model backbones.

Agents

Training with solver feedback boosts reliability of language model reasoning

The authors introduce Sequent-Prover, a family of language model agents that reason by writing and executing SMT-LIB programs with the Z3 solver. Using a novel training recipe called verifier-conditioned trajectory distillation, they post-train models by sampling multi-turn trajectories, filtering for correctness and non-triviality, and applying supervised fine-tuning and preference optimization. The 4B model's per-attempt accuracy rises from 85.6% to 97.0%, and four-sample reliability from 64.1% to 91.7%, while reducing average SMT calls and tokens.

Agents

Self-evolving visual agents learn tool use from unlabeled images using margin reward.

The authors propose Marvo, a two-agent framework that learns tool invocation for visual agents from unlabeled images. A challenger generates question-answer pairs, and a solver learns to answer them via reinforcement learning. A margin reward, defined as the solver's accuracy difference with and without tools, steers the challenger toward questions where tool use materially helps. Across 11,691 questions on agentic, visual search, mathematical, and general reasoning benchmarks, Marvo raises Qwen2.5-VL-7B's average accuracy by up to 7.1%.

Agents

Task synthesis via artifact transitions yields long-horizon agentic trajectories.

The authors propose WorkBorne, a framework that models agentic tasks using artifact transitions—specifying who performs the work, which artifacts anchor the initial state and intended outcome, and the relation between them. They construct 1,061 artifact corpora and 7,882 executable task environments, producing 15,485 high-quality trajectories. Fine-tuning Qwen3.5-27B on these trajectories improves Terminal-Bench 2.1 accuracy from 43.1% to 61.8% and ClawEval General score from 75.0 to 80.2.

Agents

RL framework mitigates reasoning-tool trade-off in MCP agents

The authors introduce MCPDisCo, a reinforcement learning framework for agents operating under the Model Context Protocol (MCP). By explicitly modeling reasoning alongside routing, execution, and termination in a Thought-Markov Decision Process, and by mitigating the reasoning–tool seesaw through token-masked reward attribution and bidirectional gradient surgery, MCPDisCo achieves higher task success with fewer chain-of-thought tokens and generalizes to unseen servers and broader agentic tasks.

Agents

Progress-aware training helps small language model agents escape repetitive loops

Cao et al. examine progress stagnation, a failure mode in which distilled small language model agents repeatedly pursue unsuccessful strategies during long interactions. They propose Progress-Aware Distillation, or ProD, which compares student and teacher trajectories and applies a stronger training penalty when the student shows more stagnation. Tests across eight models and both in-domain and out-of-domain tasks indicate that this approach reduces looping while improving success and efficiency.

Agents

Open training stack scales mobile agents across apps and tools

Yan and colleagues built an open training stack for agents that operate smartphones, including a controllable environment spanning more than 110 apps, nearly 12,000 demonstration trajectories and over 2,000 reinforcement-learning tasks. Models trained on the stack achieved competitive benchmark results and transferred to real-device evaluation, while also supporting tasks that mix screen interaction with structured app functions.

Agents

Coevolving specialized models and LLM agents for autonomous self-improvement

Ye et al. propose Heterogeneous Agentic Coevolution (HACE), a framework that coevolves a specialized foundation model with a heterogeneous LLM agent via reciprocal learning. The specialist improves through supervised learning while the agent undergoes on-policy reinforced self-training guided by verified gains on the specialist; an evolutionary search optimizes their interaction harness. Across three domains, HACE surpasses existing self-improving methods, demonstrating that specialists can continue to improve beyond supervised convergence.

Agents

Structural baselines mask real progress in agent failure detection

Naagarajan and Streif demonstrate that reported performance on detecting and attributing silent failures in LLM agents is largely driven by structural regularities—failing runs are longer and the decisive step is usually the last. A rule that always predicts the last step is correct 67–96% of the time on the omission-dominant splits. When performance is reported as gain over this baseline, no method tested improves on give-up failures, and only internal activations from the agent's model hold up against text-based baselines for detection.

Agents

Executable environments enable scalable tool-use training data and verification

The authors introduce Davinci-Tool, a 30-billion-parameter model for reliable multi-step tool use, and Davinci-ToolData, an execution-grounded dataset. By preserving restorable states, action traces, environment outcomes, and executable verifiers as a unified supervision substrate, they scale both environment coverage and outcome verification together. Supervised fine-tuning followed by outcome-grounded Direct Preference Optimization raises performance on three held-out benchmarks from 52.4 to 88.2, 30.4 to 62.8, and 16.9 to 56.2, respectively, surpassing a 480B-parameter model.

Agents

Learning the search harness improves frozen models’ scientific problem solving

Ma et al. train a language model to improve the computational harness around a frozen problem-solving model, rather than updating the solver itself. Each proposed harness governs a budgeted search, and the resulting solution quality supplies reinforcement-learning feedback. Across 14 mathematical, computational and scientific tasks, the authors report that this approach consistently outperforms a fixed-harness baseline and can transfer useful harnesses to larger executors.

Agents

Evolving environment beliefs helps LLM agents adapt without retraining

The authors present EvoBelief, a self-evolving world model that represents an LLM agent's environment knowledge as falsifiable hypotheses. During deployment, it proposes, tests, and validates beliefs about environment dynamics, then uses only validated beliefs to guide predictions and plans. On three benchmarks, this approach significantly outperformed baselines, achieving on ALFWorld success rate improvements of up to 31.3% and completing scientific discovery tasks that none of the compared methods could finish.

Agents

Alternating AI agents across sessions beats single agents on machine learning tasks

The authors identify two failure modes in long-horizon machine learning agents: long-horizon deceleration, where reasoning intensity drops as a session stretches, and harness staleness, where custom tooling helps less as base models improve. Their workflow, Humanize MLE Agents (HMA), alternates two agents on a shared workspace with context resets. On 75 Kaggle tasks in MLE-bench it reached a 78.2% any-medal rate, 8.0 points above the strongest open-source baseline.