Reflexion: Language Agents with Verbal Reinforcement Learning
Open accessThe DOI matches its DataCite record; 41 of the 41 references it cites were found in a registry or an index; 7 of the 8 sources its claims cite were checked for support.
Abstract
Large language models (LLMs) have been increasingly used to interact with external environments (e.g., games, compilers, APIs) as goal-driven agents. However, it remains challenging for these language agents to quickly and efficiently learn from trial-and-error as traditional reinforcement learning methods require extensive training samples and expensive model fine-tuning. We propose Reflexion, a novel framework to reinforce language agents not by updating weights, but instead through linguistic feedback. Concretely, Reflexion agents verbally reflect on task feedback signals, then maintain their own reflective text in an episodic memory buffer to induce better decision-making in subsequent trials. Reflexion is flexible enough to incorporate various types (scalar values or free-form language) and sources (external or internally simulated) of feedback signals, and obtains significant improvements over a baseline agent across diverse tasks (sequential decision-making, coding, language reasoning). For example, Reflexion achieves a 91% pass@1 accuracy on the HumanEval coding benchmark, surpassing the previous state-of-the-art GPT-4 that achieves 80%. We also conduct ablation and analysis studies using different feedback signals, feedback incorporation methods, and agent types, and provide insights into how they affect performance.
Overview
CoChat Summary
TL;DR
Reflexion reinforces language agents through verbal self-reflection stored in episodic memory rather than weight updates, improving performance across decision-making, reasoning, and coding tasks. It reports gains including 91% pass@1 on HumanEval, a 22% absolute improvement on AlfWorld, and 20% on HotPotQA.
Key findings
- On HumanEval Python, Reflexion achieves 91.0 pass@1, surpassing the previous state-of-the-art GPT-4 at 80.1; on HumanEval Rust it reaches 68.0 vs. 60.0, MBPP Rust 75.4 vs. 70.9, and Leetcode Hard 15.0 vs. 7.5.
- In AlfWorld, ReAct + Reflexion completes 130 out of 134 tasks using a simple heuristic and improves over strong baselines by an absolute 22% over 12 iterative learning steps; ReAct-only converges at a 22% hallucination rate with no long-term recovery.
- In HotPotQA, Reflexion improves over baselines by 20%; self-reflection gives an 8% absolute boost over episodic memory alone, and although CoT (GT) fails on 39% of questions, Reflexion improves accuracy by 14% without access to ground-truth answers.
- Ablations on the 50 hardest HumanEval Rust problems show full Reflexion at 0.68 vs. baseline 0.60; omitting test generation drops to 0.52, and omitting self-reflection stays at 0.60.
Methods
- Reflexion uses a modular loop with an Actor that generates text/actions, an Evaluator that scores outputs, and a Self-Reflection model that converts feedback into verbal summaries stored in memory; trials repeat until the Evaluator passes or a max trial count is reached, with memory usually bounded to 1–3 experiences.
- Experiments cover AlfWorld (134 environments, six tasks, ReAct + GPT-3), HotPotQA (100 questions, CoT/ReAct, exact-match grading, memory size 3), and programming (HumanEval, MBPP, LeetcodeHardGym; Python/Rust via MultiPL-E; self-generated unit tests up to 6).
- Ablations and analyses vary feedback signals, feedback incorporation methods, and agent types; compromised approaches are tested on HumanEval Rust, and a WebShop experiment is run across 100 environments.
Limitations
- The paper states Reflexion may fall into non-optimal local minima, limits long-term memory to a sliding window with maximum capacity, and notes code-generation test-driven development is limited by non-deterministic, impure, hardware-dependent, or concurrent functions; it also relies on LLM self-evaluation/heuristics and lacks a formal success guarantee.
- The paper reports WebShop failure: after only four trials the agent shows no improvement and does not generate helpful self-reflections, concluding Reflexion cannot solve tasks requiring significant diversity and exploration.
- Evident from the text: self-correction is an emergent quality of stronger, larger models—starchat-beta shows no improvement (baseline 0.26 vs. Reflexion 0.26)—and flaky or incorrect test suites can cause false positives/negatives, with false negatives preferred.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
4 Valid · 2 Not valid · 0 Not checked · 59 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Therefore, in the case in which the model generates a flaky test suite, it is possible that all tests pass on an incorrect solution and lead to a false positive label on a code completion [11].”
§ Reflexion: Language Agents with Verbal Reinforcement Learning > Analysis
What the sources say
- Partly supportedcites [11] 11. Lam, W., Winter, S., Wei, A., Xie, T., Marinov, D., and Bell, J. (2020). A large-scale longitudinal study of flaky tests. Proc. ACM Program. Lang., 4(OOPSLA). sourceExperimental — accuracy not yet measured
The Introduction (page not provided) says a flaky test can pass and fail on the same code and that changes may be merged when all tests pass. It does not address model-generated test suites, incorrect code completions, or false-positive labels, so those parts of the claim are not established by this paper. (LLM, conf=0.82)
hedgedstates a possible cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“3 shows an inferior 52% vs 60% (baseline) accuracy, which suggests that the agent is unable to determine if the current implementation is correct without unit tests.”
§ Reflexion: Language Agents with Verbal Reinforcement Learning > Ablation study
hedgedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“ReAct + Reflexion significantly outperforms ReAct by completing 130 out of 134 tasks using the simple heuristic to detect hallucinations and inefficient planning.”
§ Reflexion: Language Agents with Verbal Reinforcement Learning > Results
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“4 shows that self-reflection improves learning by an 8% absolute boost over the episodic memory learning advantage.”
§ Reflexion: Language Agents with Verbal Reinforcement Learning > Analysis
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“on reasoning questions in HotPotQA [28] by 20%”
- Also stated in: § 4 Experiments · plain
“Most notably, Reflexion improves performance over strong baselines by 22% in AlfWorld, 20% in HotPotQA, and 11% on HumanEval.”
§ Reflexion: Language Agents with Verbal Reinforcement Learning > 1 Introduction
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § 4 Experiments · plain
30 more claims
Create a free account to see all 65 claims and every source verdict.
- It may state two separate claims as one. 18
- Probably not a substantial claim. 11
- The quote only partly states this claim. 3
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The cited source doesn’t support the claim as written.
In the citing paper
“Memory and reflection systems such as MemAgent (Yu et al., 2025a), Reflexion (Shinn et al., 2023), and Generative Agents (Park et al., 2023) improve retrieval and episodic memory, but do not perform context shaping, state restructuring, or nextinstruction refinement.”
From PAACE: A Plan-Aware Automated Agent Context Engineering Framework (2025) · cites this paper as Shinn et al., 2023
Checked by a reasoning model — may be wrong.
Experimental — accuracy not yet measuredThe cited source doesn’t support the claim as written.
In the citing paper
“However, Reflexion computes verbal reward over a full trajectory, yielding a single end-of-episode signal that conflates credit across reasoning steps: a failure due to a single intermediate error receives the same critique pressure as one due to a flawed strategy.”
From Process Supervision via Verbal Critique Improves Reasoning in Large Language Models (2026) · cites this paper as Shinn et al. [2023]
Checked by a reasoning model — may be wrong.
Experimental — accuracy not yet measuredThe cited source doesn’t support the claim as written.
In the citing paper
“Unlike Reflexion, which only reflects on unsuccessful trajectories, our approach separately reflects on both successful and unsuccessful ones.”
From An Efficient Context-Dependent Memory Framework for LLM-Centric Agents (2025) · cites this paper as Shinn et al., 2023
Checked by a reasoning model — may be wrong.
Experimental — accuracy not yet measuredThe cited source doesn’t support the claim as written.
In the citing paper
“The dominant approach is to curate memory at write time.”
From Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents (2026) · cites this paper as Shinn et al., 2023
Checked by a reasoning model — may be wrong.
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Interaction and code generation are increasingly used together, with code as the modality of choice for actions [48, 59], tool construction [13, 51, 69], and reasoning [39, 66, 67].”
From SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering (2024) · cites this paper as [39]
Experimental — accuracy not yet measured
107 more claims about this paper
Create a free account to see every claim other papers make about this one.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Reflexion: Language Agents with Verbal Reinforcement Learning > 3 Reflexion: reinforcement via verbal reflection
“while M e not pass or t < max trials do”
§ Reflexion: Language Agents with Verbal Reinforcement Learning > Self-reflection
“The Actor, Evaluator, and Self-Reflection models work together through trials in a loop until the Evaluator deems τ t to be correct.”
Citations
Cites
The publisher hasn’t deposited this paper’s reference list.
Cited by
Survey of Hallucination in Natural Language Generation
Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023) · ACM Computing Surveys · doi:10.1145/3571730 · Cited by 5,025
A Survey of Large Language Models
Zhao, W. X., Zhou, K., Li, J., Tang, T., Dong, Z., Hou, Y., Zhang, B., Min, Y., Zhang, J., Liu, P., et al. (2026) · Frontiers of Computer Science · doi:10.1007/s11704-026-60308-3 · Cited by 2,093
A survey on large language model based autonomous agents
Wang, L., Ma, C., Feng, X., Zhang, Z., Yang, H., Zhang, J., Chen, Z., Tang, J., Chen, X., Lin, Y., et al. (2024) · Frontiers of Computer Science · doi:10.1007/s11704-024-40231-1 · Cited by 1,941
“Common metrics include success rate [156, 151, 125, 90], reward/score [156, 151, 99], coverage [161], and accuracy [113, 1, 61].”
InfluentialA Comprehensive Overview of Large Language Models
Naveed, H., Khan, A. U., Qiu, S., Saqib, M., Anwar, S., Usman, M., Akhtar, N., Barnes, N., & Mian, A. (2025) · ACM Transactions on Intelligent Systems and Technology · doi:10.1145/3744746 · Cited by 1,257
“…pre-trained language tasks OPT-IML • Task size sampling to create a batch with most of the task examples is important for better performance • Only example proportional sampling is not enough, training datasets/benchmarks should also be proportional for better generalization/performance • Fully…”
InfluentialImproving Factuality and Reasoning in Language Models through Multiagent Debate
Du, Y., Li, S., Torralba, A., Tenenbaum, J. B., & Mordatch, I. (2023) · arXiv · doi:10.48550/arxiv.2305.14325 · Cited by 1,202
Metadata
- Authors
- Shinn, N., Cassano, F., Berman, E., Gopinath, A., Narasimhan, K., & Yao, S.
- Venue
- arXiv
- Publisher
- arXiv
- Published
- 2023
- Pages
- 8634-8652
- ISSN
- 1049-5258 (ISSN-L)
- arXiv
- 2303.11366
- OpenAlex
- W7133222210
- Cited by
- 1,087
- References
- 0
- Citations by year
- 2026: 760 · 2025: 167 · 2024: 105 · 2023: 55
- FWCI
- 192.15
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- Locations
- neural information processing systems (publishedVersion)
- arXiv (Cornell University) (submittedVersion, cc-by)
- arXiv (Cornell University) (cc-by)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Reinforcement Learning in Robotics
- Physical Sciences › Computer Science › Artificial Intelligence › Speech and dialogue systems
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Keywords
- language agents, verbal behavior, language acquisition
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Reflexion: Language Agents with Verbal Reinforcement Learning. https://app.cochat.ai/papers/10.48550/arxiv.2303.11366
- The paper
- Shinn, N., Cassano, F., Berman, E., Gopinath, A., Narasimhan, K., & Yao, S. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. arXiv. https://doi.org/10.48550/arxiv.2303.11366
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.