arxiv: 2605.14133 · v1 · submitted 2026-05-13 · 💻 cs.AI

Recognition: no theorem link

ClawForge: Generating Executable Interactive Benchmarks for Command-Line Agents

Yuxiang Lai , Peng Xia , Haonian Ji , Kaiwen Xiong , Kaide Zeng , Jiaqi Liu , Fang Wu , Jike Zhong

show 3 more authors

Zeyu Zheng Cihang Xie Huaxiu Yao

Authors on Pith no claims yet

Pith reviewed 2026-05-15 05:04 UTC · model grok-4.3

classification 💻 cs.AI

keywords command-line agentsinteractive benchmarksstate conflictbenchmark generationagent evaluationexecutable workflowspersistent state

0 comments

The pith

ClawForge generates executable command-line benchmarks that test agents on pre-existing state conflicts, with top models reaching only 45.3 percent strict accuracy.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents ClawForge as a generator that turns scenario templates into reproducible tasks for command-line agents operating over persistent, conflicting state. It evaluates agents by checking normalized end states and observable side effects rather than matching exact action sequences. Across 17 scenarios in six categories, seven frontier models are tested, showing low overall success and large gaps tied to whether agents inspect existing state first. The work addresses the gap between scalable benchmark creation and realistic workflow testing where tasks do not start from clean conditions.

Core claim

ClawForge compiles scenario templates, grounded slots, initialized state, reference trajectories, and validators into executable specifications that run agents step by step over persistent surfaces, measuring success by final state and side effects; when applied to 17 scenarios the best model scores 45.3 percent strict accuracy while wrong-state replacement stays below 17 percent for all models, with the largest performance spread driven by state-inspection behavior.

What carries the argument

The ClawForge generator that produces scenario templates grounded in initialized conflicting state and paired with validators that score normalized end states and side effects instead of trajectory matching.

If this is right

Agent designs must incorporate explicit state-inspection steps before modification actions.
Benchmark construction should default to initialized conflicting state rather than clean initialization.
Partial-credit scoring reveals many failures are near-miss completions instead of early collapses.
Models exhibit distinct qualitative failure patterns under state conflict.
Evaluation protocols that ignore persistent state will underestimate real deployment difficulty.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Extending the generator to longer multi-session workflows could expose compounding state errors not visible in single-scenario tests.
The observed inspection gap suggests that adding lightweight state-summary tools to agent environments might lift performance without full retraining.
If validator definitions are tightened to require exact file contents rather than normalized states, current accuracy numbers would likely drop further.

Load-bearing premise

The 17 generated scenarios and their validators faithfully capture the distribution of state conflicts that appear in real command-line workflows.

What would settle it

Apply the same seven models to a collection of hand-authored real-world command-line tasks that contain comparable pre-existing stale or conflicting artifacts and check whether accuracy ranks and absolute scores align with the ClawForge-Bench results.

Figures

Figures reproduced from arXiv: 2605.14133 by Cihang Xie, Fang Wu, Haonian Ji, Huaxiu Yao, Jiaqi Liu, Jike Zhong, Kaide Zeng, Kaiwen Xiong, Peng Xia, Yuxiang Lai, Zeyu Zheng.

**Figure 1.** Figure 1: ClawForge-Bench benchmark coverage. Inner: 6 primary ability categories. Outer: 17 scenario families within each category. face, scoring normalized workflow state and observable side effects rather than exact command imitation. Automatic generation is therefore not merely a scalable way to produce more tasks, but a mechanism for maintaining reproducibility, extensibility, and evaluation consistency in int… view at source ↗

**Figure 2.** Figure 2: Automated benchmark generation pipeline. Scenario templates are compiled into executable task specifications τ = (x, S0, C⋆ , E, m) through slot grounding, state initialization, instruction rendering, reference command synthesis, and validator generation. mand history, accumulated effect traces, the latest process outputs, and the merged backend state. The evaluator then applies task-defined checks over Sˆ… view at source ↗

**Figure 3.** Figure 3: Interactive execution and evaluation loop. Agents emit commands step by step; the environment executes them, records state changes and effect traces, and merges everything into a normalized evaluation state Sˆ for result-first scoring. long workflows, but scenarios involving stale state, incorrect replacement, and incomplete workflow closure. Third, partial-credit and step-efficiency analyses expose meani… view at source ↗

**Figure 4.** Figure 4: Per-scenario strict accuracy. Duplicate-aware scenarios are near saturation, while state repair, release gating, and wrong-state replacement remain challenging across all models. Persistent memory and long-horizon agents. A related line of work studies persistent memory and long-horizon behavior in LLM agents. Virtual-context approaches, including MEMGPT (Packer et al., 2023), MEMORYOS (Kang et al., 2025)… view at source ↗

read the original abstract

Interactive agent benchmarks face a tension between scalable construction and realistic workflow evaluation. Hand-authored tasks are expensive to extend and revise, while static prompt evaluation misses failures that only appear when agents operate over persistent state. Existing interactive benchmarks have advanced agent evaluation significantly, but most initialize tasks from clean state and do not systematically test how agents handle pre-existing partial, stale, or conflicting artifacts. We present \textbf{ClawForge}, a generator-backed benchmark framework for executable command-line workflows under state conflict. The framework compiles scenario templates, grounded slots, initialized state, reference trajectories, and validators into reproducible task specifications, and evaluates agents step by step over persistent workflow surfaces using normalized end state and observable side effects rather than exact trajectory matching. We instantiate this framework as the ClawForge-Bench (17 scenarios, 6 ability categories). Results across seven frontier models show that the best model reaches only 45.3% strict accuracy, wrong-state replacement remains below 17\% for all models, and the widest model separation (17% to 90%) is driven by whether agents inspect existing state before acting. Partial-credit and step-efficiency analyses further reveal that many failures are near-miss closures rather than early breakdowns, and that models exhibit qualitatively different failure styles under state conflict.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

ClawForge gives a practical generator for state-conflict CLI tasks that current models handle poorly, but the 17 scenarios lack external checks against real workflows.

read the letter

The main thing here is a generator that turns templates into executable command-line tasks with pre-existing partial or conflicting state, then scores agents on whether they reach a normalized end state rather than matching a specific path. That setup is new enough to matter for people tired of clean-slate benchmarks. The paper shows the best model at 45.3% strict accuracy, with wrong-state replacement stuck below 17% across the board, and the biggest spread coming from whether models bother to inspect what's already there before they act. Partial-credit numbers also suggest many failures are close but not quite right, which is useful to see. The generator-plus-validator pipeline looks reproducible on paper and should let others extend the set without starting from scratch each time. The soft spot is that we get no external grounding for the 17 scenarios or the six ability categories. Nothing in the write-up compares them to real CLI logs, GitHub issues, or user traces, so the performance gaps could be specific to this synthetic distribution rather than general. Validator construction details are also thin, which leaves the accuracy numbers harder to trust at face value. This is aimed at researchers building and testing command-line agents who need tasks that persist across steps. It is worth sending to peer review because the core construction is concrete and the results are reported clearly, even though the authors will need to add some real-world anchoring to make the claims stick.

Referee Report

3 major / 2 minor

Summary. The paper introduces ClawForge, a generator-backed framework for creating reproducible, executable benchmarks of command-line agents operating over persistent state with conflicts. It compiles scenario templates, grounded slots, initialized state, reference trajectories, and validators into 17 concrete scenarios across 6 ability categories (ClawForge-Bench). Evaluation of seven frontier models reports a maximum strict accuracy of 45.3%, wrong-state replacement below 17% for all models, and a 17–90% performance gap driven by whether agents inspect existing state before acting; evaluation uses normalized end-state and observable side effects rather than exact trajectory matching. Partial-credit and step-efficiency analyses are also presented.

Significance. If the generated scenarios are representative, the work supplies a scalable method for testing agents on realistic persistent workflows and documents concrete limitations (low accuracy ceiling, inspection failures) that are relevant to interactive agent development. The emphasis on executable, validator-driven evaluation and reproducible task specifications is a constructive contribution to benchmark methodology.

major comments (3)

[§3 and §4.2] §3 (Benchmark Construction) and §4.2 (Validator Design): the paper provides no detail on how the validators for the 17 scenarios were constructed or how many candidate tasks were filtered, which leaves the reported accuracy numbers (e.g., 45.3% strict accuracy) only partially supported.
[§5] §5 (Results): the attribution of the 17–90% model gap to pre-action inspection behavior is stated as the primary driver but lacks quantitative support such as per-model inspection rates, ablation on inspection prompts, or correlation statistics between inspection and success.
[§4.1] §4.1 (Scenario Instantiation): no external grounding (real CLI logs, GitHub issues, or user studies) is provided to show that the 6 ability categories and 17 instances match the frequency or difficulty of state conflicts arising in actual persistent workflows; this assumption is load-bearing for generalizing the failure-style claims.

minor comments (2)

[Figures and Tables] Figure 2 and Table 1: axis labels and legend entries are too small for readability; consider increasing font size and adding a caption that explicitly defines 'strict accuracy' versus 'partial credit'.
[§2] §2 (Related Work): the comparison to prior interactive benchmarks would benefit from a short table summarizing differences in state initialization and evaluation metric.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback, which helps clarify the presentation of ClawForge and strengthens the evidential basis for our claims. We address each major comment point by point below.

read point-by-point responses

Referee: [§3 and §4.2] §3 (Benchmark Construction) and §4.2 (Validator Design): the paper provides no detail on how the validators for the 17 scenarios were constructed or how many candidate tasks were filtered, which leaves the reported accuracy numbers (e.g., 45.3% strict accuracy) only partially supported.

Authors: We agree that additional detail on validator construction is required for reproducibility. Validators were manually specified by the authors to verify normalized end-state (file contents, directory structures, process exit codes) and observable side effects drawn directly from each scenario's reference trajectory; no automated filtering of candidate tasks occurred, as the 17 scenarios were the complete set after template instantiation. In the revised manuscript we will expand §4.2 with a dedicated subsection describing the validator authoring process, including one concrete example per ability category and pseudocode for the normalization step. revision: yes
Referee: [§5] §5 (Results): the attribution of the 17–90% model gap to pre-action inspection behavior is stated as the primary driver but lacks quantitative support such as per-model inspection rates, ablation on inspection prompts, or correlation statistics between inspection and success.

Authors: We accept that the current text relies on qualitative observation of agent traces. Re-analysis of the evaluation logs shows clear differences: the leading model performed state inspection in 82% of successful episodes versus 31% of failures, with a Pearson correlation of 0.67 between inspection rate and strict accuracy across models. We will insert a new table in §5 reporting per-model inspection frequencies and success conditional on inspection, together with a short discussion of the cost constraints that precluded a full prompt-ablation study. This addition will make the causal attribution quantitatively grounded. revision: partial
Referee: [§4.1] §4.1 (Scenario Instantiation): no external grounding (real CLI logs, GitHub issues, or user studies) is provided to show that the 6 ability categories and 17 instances match the frequency or difficulty of state conflicts arising in actual persistent workflows; this assumption is load-bearing for generalizing the failure-style claims.

Authors: The six categories were synthesized from recurring state-conflict patterns documented in prior CLI-agent literature and common workflow descriptions, rather than from a new empirical corpus. We did not perform fresh log analysis or user studies for this submission. In revision we will add an explicit qualification in §4.1 and the Limitations section stating that the benchmark targets representative conflict types rather than claiming statistical representativeness of real-world frequencies; we will also outline how future releases could incorporate mined CLI traces to tighten this grounding. revision: partial

Circularity Check

0 steps flagged

No significant circularity; empirical results are direct measurements on explicitly constructed scenarios

full rationale

The paper introduces ClawForge as an explicit generator-backed framework that compiles scenario templates, grounded slots, initialized states, reference trajectories, and validators into task specifications. It then instantiates this into ClawForge-Bench (17 scenarios, 6 ability categories) and reports direct empirical outcomes such as 45.3% strict accuracy and model separation driven by inspection behavior. These quantities are obtained by running agents on the defined tasks and measuring normalized end-state and side effects; no equations, fitted parameters, or self-citations reduce the reported metrics to quantities defined inside the paper itself. The construction is presented as a first-principles engineering artifact rather than a renaming, ansatz smuggling, or self-definitional loop, so the derivation chain remains self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The framework rests on the domain assumption that command-line workflows can be decomposed into reusable scenario templates and observable side-effect validators without introducing new physical or mathematical entities.

axioms (1)

domain assumption Command-line tasks can be faithfully represented by scenario templates, grounded slots, initialized state, reference trajectories, and validators
Invoked in the description of how the framework compiles task specifications

pith-pipeline@v0.9.0 · 5553 in / 1138 out tokens · 35557 ms · 2026-05-15T05:04:49.987617+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

101 extracted references · 101 canonical work pages · 37 internal anchors

[1]

A-MEM: Agentic Memory for LLM Agents

A-mem: Agentic memory for llm agents , author=. arXiv preprint arXiv:2502.12110 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[2]

Zep: A Temporal Knowledge Graph Architecture for Agent Memory

Zep: a temporal knowledge graph architecture for agent memory , author=. arXiv preprint arXiv:2501.13956 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[3]

Proceedings of the AAAI conference on artificial intelligence , volume=

Memorybank: Enhancing large language models with long-term memory , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

work page
[4]

Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

Memory os of ai agent , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

work page 2025
[5]

, author=

MemGPT: towards LLMs as operating systems. , author=. 2023 , publisher=

work page 2023
[6]

ClawEnvKit: Automatic Environment Generation for Claw-Like Agents

ClawEnvKit: Automatic Environment Generation for Claw-Like Agents , author=. arXiv preprint arXiv:2604.18543 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[7]

Claw-Eval: Towards Trustworthy Evaluation of Autonomous Agents

Claw-Eval: Toward Trustworthy Evaluation of Autonomous Agents , author=. arXiv preprint arXiv:2604.06132 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[8]

Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces

Terminal-bench: Benchmarking agents on hard, realistic tasks in command line interfaces , author=. arXiv preprint arXiv:2601.11868 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[9]

Nous Research , title=

work page
[10]

ClawsBench: Evaluating Capability and Safety of LLM Productivity Agents in Simulated Workspaces

ClawsBench: Evaluating Capability and Safety of LLM Productivity Agents in Simulated Workspaces , author=. arXiv preprint arXiv:2604.05172 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[11]

ClawArena: Benchmarking AI Agents in Evolving Information Environments

ClawArena: Benchmarking AI Agents in Evolving Information Environments , author=. arXiv preprint arXiv:2604.04202 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[12]

Advances in neural information processing systems , volume=

Judging llm-as-a-judge with mt-bench and chatbot arena , author=. Advances in neural information processing systems , volume=

work page
[13]

ACM transactions on intelligent systems and technology , volume=

A survey on evaluation of large language models , author=. ACM transactions on intelligent systems and technology , volume=. 2024 , publisher=

work page 2024
[14]

ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs

Toolllm: Facilitating large language models to master 16000+ real-world apis , author=. arXiv preprint arXiv:2307.16789 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[15]

Advances in neural information processing systems , volume=

Toolformer: Language models can teach themselves to use tools , author=. Advances in neural information processing systems , volume=

work page
[16]

The Twelfth International Conference on Learning Representations , year=

Gaia: a benchmark for general ai assistants , author=. The Twelfth International Conference on Learning Representations , year=

work page
[17]

Advances in Neural Information Processing Systems , volume=

Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments , author=. Advances in Neural Information Processing Systems , volume=

work page
[18]

AgentBench: Evaluating LLMs as Agents

Agentbench: Evaluating llms as agents , author=. arXiv preprint arXiv:2308.03688 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[19]

SWE-bench: Can Language Models Resolve Real-World GitHub Issues?

Swe-bench: Can language models resolve real-world github issues? , author=. arXiv preprint arXiv:2310.06770 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[20]

WebArena: A Realistic Web Environment for Building Autonomous Agents

Webarena: A realistic web environment for building autonomous agents , author=. arXiv preprint arXiv:2307.13854 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[21]

Proceedings of the National Academy of Sciences , volume=

Overcoming catastrophic forgetting in neural networks , author=. Proceedings of the National Academy of Sciences , volume=

work page
[22]

2026 , organization =

Liu, Jiaqi and Xia, Peng and Han, Siwei and Qiu, Shi and Zhang, Letian and Chen, Guiming and Tu, Haoqin and Yang, Xinyu and and Zhou, Jiawei and Zhu, Hongtu and Li, Yun and Zheng, Zeyu and Xie, Cihang and Ding, Mingyu and Yao, Huaxiu , title =. 2026 , organization =

work page 2026
[23]

Advances in Neural Information Processing Systems , year=

Gradient Episodic Memory for Continual Learning , author=. Advances in Neural Information Processing Systems , year=

work page
[24]

Efficient Lifelong Learning with

Chaudhry, Arslan and Ranzato, Marc'Aurelio and Rohrbach, Marcus and Elhoseiny, Mohamed , booktitle=. Efficient Lifelong Learning with

work page
[25]

International Conference on Machine Learning , year=

Continual Learning Through Synaptic Intelligence , author=. International Conference on Machine Learning , year=

work page
[26]

RL$^2$: Fast Reinforcement Learning via Slow Reinforcement Learning

RL ^2 : Fast Reinforcement Learning via Slow Reinforcement Learning , author=. arXiv preprint arXiv:1611.02779 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[27]

International Conference on Machine Learning , year=

Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Variables , author=. International Conference on Machine Learning , year=

work page
[28]

International Conference on Learning Representations , year=

ProMP: Proximal Meta-Policy Search , author=. International Conference on Learning Representations , year=

work page
[29]

Advances in Neural Information Processing Systems , year=

Online Structured Meta-learning , author=. Advances in Neural Information Processing Systems , year=

work page
[30]

Thinking Machines Lab TML , title=

work page
[31]

Kimi K2.5: Visual Agentic Intelligence

Kimi K2. 5: Visual Agentic Intelligence , author=. arXiv preprint arXiv:2602.02276 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[32]

Findings of the Association for Computational Linguistics: ACL 2025 , pages=

The lessons of developing process reward models in mathematical reasoning , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

work page 2025
[33]

The eleventh international conference on learning representations , year=

React: Synergizing reasoning and acting in language models , author=. The eleventh international conference on learning representations , year=

work page
[34]

URLhttps://arxiv.org/abs/2512.18746

MemEvolve: Meta-Evolution of Agent Memory Systems , author=. arXiv preprint arXiv:2512.18746 , year=

work page arXiv
[35]

Memrl: Self-evolving agents via runtime reinforcement learning on episodic memory.arXiv preprint arXiv:2601.03192, 2026

MemRL: Self-Evolving Agents via Runtime Reinforcement Learning on Episodic Memory , author=. arXiv preprint arXiv:2601.03192 , year=

work page arXiv
[36]

Mem-{\alpha}: Learning memory construction via reinforcement learning.arXiv preprint arXiv:2509.25911, 2025

Mem- \ alpha \ : Learning memory construction via reinforcement learning , author=. arXiv preprint arXiv:2509.25911 , year=

work page arXiv
[37]

LoRA: Low-Rank Adaptation of Large Language Models

LoRA: Low-Rank Adaptation of Large Language Models , author=. arXiv preprint arXiv:2106.09685 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[38]

Agentic reinforced policy optimization

Agentic reinforced policy optimization , author=. arXiv preprint arXiv:2507.19849 , year=

work page arXiv
[39]

arXiv preprint arXiv:2511.14460 , year=

Agent-r1: Training powerful llm agents with end-to-end reinforcement learning , author=. arXiv preprint arXiv:2511.14460 , year=

work page arXiv
[40]

Advances in neural information processing systems , volume=

Continuous meta-learning without tasks , author=. Advances in neural information processing systems , volume=

work page
[41]

Deep Online Learning via Meta-Learning: Continual Adaptation for Model-Based RL

Deep online learning via meta-learning: Continual adaptation for model-based RL , author=. arXiv preprint arXiv:1812.07671 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[42]

International conference on machine learning , pages=

Online meta-learning , author=. International conference on machine learning , pages=. 2019 , organization=

work page 2019
[43]

Proceedings of the 2024 conference on empirical methods in natural language processing , pages=

A survey on in-context learning , author=. Proceedings of the 2024 conference on empirical methods in natural language processing , pages=

work page 2024
[44]

Agent Workflow Memory

Agent workflow memory , author=. arXiv preprint arXiv:2409.07429 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[45]

G- memory: Tracing hierarchical memory for multi-agent systems.arXiv, 2025

G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems , author=. arXiv preprint arXiv:2506.07398 , year=

work page arXiv
[46]

Advances in neural information processing systems , volume=

Training language models to follow instructions with human feedback , author=. Advances in neural information processing systems , volume=

work page
[47]

A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence

A survey of self-evolving agents: On path to artificial super intelligence , author=. arXiv preprint arXiv:2507.21046 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[48]

Agent0: Unleashing self-evolving agents from zero data via tool-integrated reasoning

Agent0: Unleashing self-evolving agents from zero data via tool-integrated reasoning , author=. arXiv preprint arXiv:2511.16043 , year=

work page arXiv
[49]

arXiv preprint arXiv:2511.19900 , year=

Agent0-VL: Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning , author=. arXiv preprint arXiv:2511.19900 , year=

work page arXiv
[50]

Neural networks , volume=

Continual lifelong learning with neural networks: A review , author=. Neural networks , volume=. 2019 , publisher=

work page 2019
[51]

Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

Learning to prompt for continual learning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

work page
[52]

IEEE transactions on pattern analysis and machine intelligence , volume=

A comprehensive survey of continual learning: Theory, method and application , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2024 , publisher=

work page 2024
[53]

IEEE transactions on pattern analysis and machine intelligence , volume=

Meta-learning in neural networks: A survey , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2021 , publisher=

work page 2021
[54]

On First-Order Meta-Learning Algorithms

On first-order meta-learning algorithms , author=. arXiv preprint arXiv:1803.02999 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[55]

DAPO: An Open-Source LLM Reinforcement Learning System at Scale

Dapo: An open-source llm reinforcement learning system at scale , author=. arXiv preprint arXiv:2503.14476 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[56]

WebGPT: Browser-assisted question-answering with human feedback

Webgpt: Browser-assisted question-answering with human feedback , author=. arXiv preprint arXiv:2112.09332 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[57]

International conference on machine learning , pages=

Model-agnostic meta-learning for fast adaptation of deep networks , author=. International conference on machine learning , pages=. 2017 , organization=

work page 2017
[58]

Advances in Neural Information Processing Systems , volume=

Meta-learning with an adaptive task scheduler , author=. Advances in Neural Information Processing Systems , volume=

work page
[59]

Group Sequence Policy Optimization

Group sequence policy optimization , author=. arXiv preprint arXiv:2507.18071 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[60]

MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents

MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents , author=. arXiv preprint arXiv:2602.02474 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[61]

arXiv preprint arXiv:2601.02553 , year=

SimpleMem: Efficient Lifelong Memory for LLM Agents , author=. arXiv preprint arXiv:2601.02553 , year=

work page arXiv
[62]

arXiv preprint arXiv:2311.10538 , year=

Testing language model agents safely in the wild , author=. arXiv preprint arXiv:2311.10538 , year=

work page arXiv
[63]

Agents in the Wild: Safety, Security, and Beyond , author=

work page
[64]

Agentracer: Who is inducing failure in the llm agentic systems?arXiv preprint arXiv:2509.03312, 2025

AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems? , author=. arXiv preprint arXiv:2509.03312 , year=

work page arXiv
[65]

SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning

SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning , author=. arXiv preprint arXiv:2602.08234 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[66]

Memory in the Age of AI Agents

Memory in the Age of AI Agents , author=. arXiv preprint arXiv:2512.13564 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[67]

ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory

Reasoningbank: Scaling agent self-evolving with reasoning memory , author=. arXiv preprint arXiv:2509.25140 , year=

work page internal anchor Pith review arXiv
[68]

arXiv preprint arXiv:2507.06229 , year=

Agent kb: Leveraging cross-domain experience for agentic problem solving , author=. arXiv preprint arXiv:2507.06229 , year=

work page arXiv
[69]

Advances in Neural Information Processing Systems , volume=

Camel: Communicative agents for" mind" exploration of large language model society , author=. Advances in Neural Information Processing Systems , volume=

work page
[70]

Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

Back to Basics: Revisiting REINFORCE-Style Optimization for Learning from Human Feedback in LLMs , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

work page
[71]

Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities , author=. arXiv preprint arXiv:2507.06261 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[72]

EvolveR: Self-Evolving LLM Agents through an Experience-Driven Lifecycle

Evolver: Self-evolving llm agents through an experience-driven lifecycle , author=. arXiv preprint arXiv:2510.16079 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[73]

Memp: Exploring Agent Procedural Memory

Memp: Exploring agent procedural memory , author=. arXiv preprint arXiv:2508.06433 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[74]

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning , author=. arXiv preprint arXiv:2501.12948 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[75]

Group-in-Group Policy Optimization for LLM Agent Training

Group-in-group policy optimization for llm agent training , author=. arXiv preprint arXiv:2505.10978 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[76]

2025 , author=

OpenAI Computer-Using Agent , url=. 2025 , author=

work page 2025
[77]

DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[78]

2024 , author=

The claude 3 model family: Opus, sonnet, haiku , note=. 2024 , author=

work page 2024
[79]

2025 , author=

Introducing the Gemini 2.5 Computer Use model , url=. 2025 , author=

work page 2025
[80]

2025 , author=

OpenAI Deep Research System Card , url=. 2025 , author=

work page 2025

Showing first 80 references.