The paper introduces RegretBench, a benchmark of 205 tasks that capture moments where users regret not backing up state, and evaluates six advanced coding agents on their backup awareness. While agents complete 86.1% of tasks, they only back up before regret moments 15.2% of the time, revealing a gap between task success and state preservation. To address this, the authors propose BackUpper, a plug‑in that automatically preserves affected state before tool calls, achieving 92.2% similarity to the original state on RegretBench at a cost of $0.014 per task and aligning with user preferences with an 84.5% F1 score.
arXiv:2610.27954v1 Announce Type: new Abstract: Recent advances in large (vision) language models have enabled computer-use agents that automate digital workflows via fine-grained keyboard and mouse...
Long-horizon coding requires agents to maintain an evolving view of the project they are changing: what problem is currently being solved, what has already been learned or modified, which other artifa...
arXiv:2608. 12851v1 Announce Type: new Abstract: Self-improving LLM agents convert successful trajectories into persistent cross-task state.
The paper investigates how large language model (LLM) agents can improve performance by storing and reusing experience in external memory. It identifies a problem called error reproduction, where mistakes in stored abstractions propagate through cycles of abstraction, execution, and re-abstraction, leading to persistent errors. To mitigate this, the authors propose Anti-Persistence, a memory design that retains raw trajectories and outcomes but reconstructs task-specific guidance on demand, discarding it after use. Experiments across multiple benchmarks show that Anti-Persistence yields higher held‑out success rates and lower LLM‑call costs compared to other memory designs, supporting the idea of separating experience accumulation from abstraction persistence.
Agent-Controlled Forgetting for Tool-Using Agents: Reversible Context Curation in Practice explores a method where an agent selectively replaces detailed tool results with concise notes while archiving the originals for recovery. Using a Python harness, the approach reduces prompt tokens from 912,492 to 231,951 and cuts cumulative input tokens by 50%, yielding significant API cost savings. However, the method increases request count, slightly slows execution, and its effectiveness varies across different workloads.
The paper investigates how deleting portions of a software engineering agent’s past trajectory affects performance and cost. Across 9,400 runs with two state‑of‑the‑art agents, it finds that large deletions (even up to 100% of reasoning text) can maintain overall effectiveness if applied with a delay, but immediate deletion harms performance. The study also shows that cost savings depend on the pricing model: a flat model yields savings in 26 of 42 cases, while a caching‑aware model actually increases costs because it breaks a low‑rate cached prefix.
The paper introduces Self‑Direction, a method that lets agents learn from mistakes by reflecting on failed trajectories, distilling lessons into lightweight low‑rank adapters, and storing them for future use. These adapters add only about 0.02% of the base model’s parameters, keeping the original weights frozen and avoiding costly fine‑tuning. Experiments on Qwen‑based agents across research, tool‑use, and terminal tasks show 8.9–17.0 percentage‑point gains in success rates, reduced repeated errors, and transfer of lessons to new tasks.
Self-evolving agents can continually improve their behavior, while tools define the executable action space through which they interact with the environment. However, exposing the full tool library to...
The paper introduces the evidence-action gap, where coding agents persist in tasks even when completion is impossible, despite having sufficient evidence that outside help is needed. To quantify this, the authors present ImpasseBench, a benchmark that injects realistic blockers into repair tasks and marks the step at which evidence becomes sufficient. Across eight models and 19,371 trajectories, they find that many agents still exhaust their budget or make unsafe requests after the evidence threshold, but an escalation tool and external monitor can significantly reduce these failures.
HarnessMM (Harness Minus Minus) is a method that improves harness token efficiency by using guarded decision reuse and recursive self‑improvement (RSI). It refines harness configurations from training feedback, reuses resolved decisions to replace later reasoning callbacks, and stores actions in Executable Decision Graphs. Experiments show significant token savings while maintaining or improving success rates on benchmarks such as HumanEval and DeepSeek Harness.
The paper introduces an online skill‑evolution framework that transforms interaction traces and evaluator feedback into a persistent, versioned library of reusable procedures for computer‑use agents. By executing each iteration against a frozen library snapshot, the system updates skills without altering the underlying model parameters. Experiments across four OSWorld domains show that the evolving library consistently outperforms an empty‑library baseline, with gains ranging from 5.7 to 18.6 percentage points, while also revealing domain‑specific temporal stability and challenges in skill retrieval and revision.
The paper introduces Traverse, a benchmark of 2,518 agent trajectories and 6,967 annotated mistakes across software engineering, computer use, and science tasks, revealing that failures often go unrecovered and can cause irreversible harm before a run is deemed successful. It shows that human judges struggle to detect the first mistake in most runs, while a 4‑billion‑parameter verifier called Scout can locate failures more effectively and improve task success when used to select among candidate runs. The study demonstrates that making failure detection inexpensive and reliable can enable long‑horizon agents to learn from their own mistakes and increase trustworthiness in autonomous AI.