Self-Correction Makes LLMs Better Parsers

Chen Gong; Yang Hou; Zhenghua Li; Ziyan Zhang

REVIEW 3 major objections 5 minor 2 cited by

Self-Correction Makes LLMs Better Parsers

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Rule-guided self-correction lets LLMs recover 10–20 F1 points in constituency parsing without retraining.

desk verdict Promising training-free self-correction for LLM parsing, but the paper's own rule statistics contradict each other by a factor of 5.6 and the main analysis needs reconciliation. read the letter →

arxiv 2504.14165 v1 pith:HO5EHTEA submitted 2025-04-19 cs.CL

Ziyan Zhang , Yang Hou , Chen Gong , Zhenghua Li This is my paper · ORCID

classification cs.CL

keywords constituencyparsinglargelanguagemodelsself-correctiontreebankgrammarrulesfew-shotlearningerroranalysiscross-domain

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that large language models' weak constituency parsing does not require retraining: most errors trace back to the models' failure to use grammar rules that already exist in treebanks, and a training-free self-correction loop that feeds those rules back as hints can fix them. The authors support this with an error analysis that splits LLM parse mistakes into span, label, flatness, and deepness errors, and with a correction pipeline that detects suspect subtrees by looking up their parent-child rules in the treebank. On five-shot parsing, the method lifts GPT-4's F1 on the Penn Treebank from 73.38 to 83.50 and on CTB5 from 44.69 to 64.99, with positive gains for every model and dataset tested. If right, the result means that a substantial part of LLM parsing weakness is recoverable knowledge rather than a hard architectural limit.

What carries the argument

The load-bearing mechanism is rule lookup against an existing treebank: every subtree is compressed to a rule, meaning the parent tag plus the sequence of its children's tags, and the treebank's train split is treated as the authority on which rules are valid. Absence from the treebank marks a subtree as needing correction; candidate rules are then aligned to the detected error type (label, flatness, deepness) and ranked by LCS of child-tag sequences, with frequency as tie-breaker, so that the LLM receives the most similar attested structures as in-context examples. The unmatch-correction step, which passes back targeted hints about added, omitted, split, or modified words, operates first to ensure the leaves of the predicted tree align with the original sentence.

What would settle it

Take the same two-pass correction pipeline but replace the rule-absence detector with an oracle that flags exactly the subtrees that disagree with gold parses; if the F1 gain is no larger than the paper's, then error detection via treebank lookup is not what drives the improvement. A cheaper check is to count how often the subtrees flagged by rule absence actually match the gold tree.

Watch

Extended reading notes

Core claim

The central claim is that LLMs underperform at constituency parsing largely because they do not systematically exploit the grammar rules (parent-child subtrees) of existing treebanks, and that supplying those rules through a self-correction mechanism, without any parameter updates, yields substantial and consistent gains. The method first repairs leaf mismatches between the predicted tree and the input sentence, then walks the predicted tree top-down; for each subtree whose rule is absent from the treebank, it treats the rule as likely erroneous, transforms candidate treebank rules to match three error types (label, flatness, deepness), ranks them by longest common subsequence of child label sequences, and prompts the LLM with the top five rule examples to regenerate that subtree. The paper reports F1 gains across GPT-4, GPT-3.5, LLaMA-3-8B, LLaMA-3-70B, Qwen-2.5-72B, and DeepSeek-V3 on PTB, CTB5, and MCTB, including cross-domain settings, and argues that correcting from the top down indirectly reduces span errors as well.

Load-bearing premise

The whole pipeline treats a grammar rule that is absent from the treebank as an error, even though the paper's own analysis shows many unseen rules are correct; if that equation is wrong often enough, the detector both flags valid subtrees and misses genuine errors.

Editorial extensions

If this is right

The improvement is training-free: any LLM that can follow few-shot prompts can adopt the method without gradient updates or task-specific fine-tuning.
Recall gains are consistently larger than precision gains, indicating the method pushes LLMs toward deeper, less flat trees and directly counters their known tendency to under-parse.
The method transfers to cross-domain parsing, improving performance on dialogue, forum, law, literature, and review domains even when the treebank comes from a different genre.
All four error types decrease after correction, and span errors—the most frequent category—are reduced indirectly through the top-down structural correction process.
Selecting examples by error-specific rule ranking beats both random example selection and POS-sequence-based selection, showing that the retrieval and ranking steps carry the benefit.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the paper's own numbers show that many rules absent from the treebank are actually correct, the detector is a noisy proxy; the gains may come less from accurate error identification than from giving the LLM a second chance with any valid treebank examples, a hypothesis that could be tested by comparing against a detector that flags subtrees using gold parses.
The indirect reduction of span errors suggests a broader design principle for structured prediction: correcting a full constituent structure from the top down can repair boundary errors without explicit span-level supervision, which might transfer to semantic parsing or named-entity recognition where treebank-style constraints exist.
The method's value should scale with how completely the treebank covers the test domain's rules; a natural testable extension is to measure gains on domains with deliberately low rule overlap, where the paper's logic predicts smaller improvements.
Since the paper reports that about 4% of originally correct rules are turned wrong during correction, running the self-correction loop for multiple iterations could accumulate such errors; a multi-round experiment would reveal whether gains saturate or reverse.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Read the paper. The core finding is credible: a training-free loop that detects likely wrong subtrees by checking against treebank grammar rules, retrieves LCS-ranked example rules, and asks the LLM to correct itself. On PTB, GPT-4 goes from 73.4 to 83.5 F1, and the gains are consistent across English/Chinese and open/closed models. The random-selection control (Table 4) is the right experiment and shows the retrieval component matters. That's a real, useful contribution for anyone building syntax-aware pipelines on top of LLMs.

The soft spots are addressable but real. No code or error bars; the hyperparameter sweep (minimum subtree height h) is run on the PTB test set itself, which inflates the headline number by an unknown amount. The contamination check covers only Llama-8B, not GPT-3.5/GPT-4, so the leak question for the main results remains open. And the load-bearing assumption that "rule not found in treebank => error" is known to be noisy from the paper's own numbers.

The bigger problem is internal consistency. Table 2 and Figure 5(a) report the same quantity for GPT-4 on PTB—the accuracy of known and unknown rules in the base output—and they disagree wildly. Table 2 gives unknown accuracy 37.14% (7,381 correct of 19,875). Figure 5(a) shows only 1,320 correct unknown rules (6.64%). Section 4.5 later quotes the pre-correction unknown accuracy as 35.3%. These cannot all be right. The known-rule numbers also differ (16,526 vs 15,897). The paper never defines "accuracy" precisely, and the discrepancy changes the interpretation of the method: if the true unknown accuracy is about 7%, the rule-absence detector is much more precise than the reader's analysis assumed, but the paper's story about LLMs lacking known rules needs rework. This is not a rounding error; it's a factor of 5.6.

Who is this for? Researchers working on LLM parsing or retrieval-augmented prompting for structured prediction. It deserves a serious referee, but it needs major revision before acceptance: reconcile the rule statistics, release code and variance numbers, and move the h selection off the test set. If they do that, this could be a solid paper.

Referee Report

3 major / 5 minor

Summary. The paper studies why LLMs underperform on constituency parsing and proposes a training-free self-correction method. It first analyzes parsing outputs by comparing grammar rules extracted from predicted trees against treebank rules and by classifying errors into span, label, flatness, and deepness categories. It then proposes a two-stage correction: unmatch correction to align leaf words, and structure correction that detects subtrees whose rules are absent from the treebank, retrieves similar rules via LCS ranking after error-type-specific processing, and prompts the LLM to rewrite the subtree using sampled examples. Experiments on PTB, CTB5, and MCTB with LLaMA-3, Qwen-2.5-72B, DeepSeek-V3, GPT-3.5, and GPT-4 report large positive gains over the few-shot baseline, e.g., GPT-4 PTB F1 73.38 to 83.50 and CTB5 44.69 to 64.99, with consistent improvements across all model/dataset combinations.

Significance. If the results are reliable, the paper makes a useful contribution: a training-free method for improving LLM constituency parsing that is simple, model-agnostic, and appears to generalize across English and Chinese. The large and consistent gains, especially the recall improvements, plus the random-selection control in Section 4.6, provide evidence that the retrieved rules and examples carry information beyond simply adding more examples. The analysis of rule statistics and error types is also a useful diagnostic. However, the reliability of the central empirical claims is currently weakened by an internal inconsistency in the rule statistics (Table 2 versus Figure 5), by apparent selection of hyperparameters on the PTB test set, and by a contamination check performed only on LLaMA-8B rather than on the closed models used for the headline results. These issues are fixable, so the paper is not beyond repair.

major comments (3)

[Table 2, Figure 5(a), Section 2.2] The paper reports irreconcilable numbers for the accuracy of unknown rules for GPT-4 on PTB. Table 2 reports 81.44% known-rule accuracy over 20,291 known rules and 37.14% unknown-rule accuracy over 19,875 unknown rules; under the natural span-level definition of rule correctness used in Figure 5(a), these rates imply 16,526 correct known rules and 7,381 correct unknown rules. Figure 5(a), with the same totals, shows 15,897 correct known rules and 1,320 correct unknown rules, corresponding to about 78.3% and 6.64% accuracy, respectively. The unknown-rule accuracy differs by a factor of about 5.6. This is not a minor rounding discrepancy: Section 2.2 uses the high unknown-rule accuracy to motivate the claim that LLMs lack known rules, and Section 3.2 step 1 relies on rule absence as an error signal, whose precision depends directly on this quantity. The authors must define the exact rule-accuracy computation, reconcile the two presentations, and ideally release the counting script or per-rule statistics.
[Appendix A.3, Section 2.1] The contamination check is performed only with LLaMA-8B, yet the main results include GPT-3.5 and GPT-4, and Table 1 uses LLaMA-70B, Qwen-72B, and DeepSeek-V3. The MKP analysis therefore supports a statement only about LLaMA-8B; the conclusion that "these datasets are unlikely to have been included in the pre-training data" overreaches when applied to the closed models. The authors should either qualify the claim to the open models actually tested or provide additional evidence for the closed models, and discuss how partial contamination would affect the interpretation of the baseline and corrected results.
[Figure 4, Appendix A.4, Section 4.2] Key hyperparameters appear to be selected on the PTB test set itself. Figure 4 reports the effect of the minimum correction height h on PTB, and Appendix A.4 compares the number of retrieved rules and examples, reporting that the setting "5 rules * 1 example" works better, again on PTB with GPT-4. No development split is described, and no error bars or significance tests are given for any of the main results in Table 3. This makes the reported improvements vulnerable to selection bias and leaves the reader unable to judge whether the gains are statistically distinguishable from the random-selection baseline. The authors should specify a dev/test protocol, report variance across runs or example sets, or otherwise justify that the chosen settings were not tuned on the evaluation set.

minor comments (5)

[Abstract and Section 2.2] The abstract contains an ungrammatical phrase, "LLMs may stem from limitations to fully leverage grammar rules," and the footnote in Section 2.2 misspells "involve" as "involve." These should be corrected.
[Tables 1 and 2] The model name "DeeSeekp-v3" is inconsistent with "DeepSeek-V3" used in the text and in Table 9; please unify the spelling.
[Figure 2] Figure 2 would be easier to read with an explicit legend or directly labeled bars, since the four error types are currently identified only through the caption and surrounding text.
[Section 4.5 and Appendix A.6] There are typos in "unknownn rules" and "we conducte a detailed analysis"; these should be fixed.
[Appendix A.3] The MKP interpretation would benefit from a brief explanation of why a higher MKP indicates a lower likelihood of pre-training inclusion, since the phrasing is otherwise unintuitive.

Circularity Check

0 steps flagged · score 0.0 of 10

No derivational circularity: main gains are measured on held-out test sets, and the treebank-rule feedback is a training-free method rather than an encoded fit to the test labels.

full rationale

Section 2.2's rule statistics and Section 3.2's rule-absence error detector both use the training treebank to define 'known' rules, but the paper's headline improvements (Table 3) are F1 scores on held-out PTB, CTB5, and MCTB test parses, not on the treebank used to build the hints. A subtree being absent from the training treebank is a heuristic for 'error,' not a definition of the target; the gold trees come from a separate test split. The increase in known-rule counts after correction (Figure 5) is partly a definitional consequence of feeding treebank examples, but the central parsing-F1 claim does not reduce to that statistic: the model could copy examples and still fail on unseen test sentences, and the method improves on all datasets. The only self-citation (Zhou et al. 2023, which shares two authors with the present paper) appears as background in Related Work and is not load-bearing. Two non-circular concerns should be weighed separately: Figure 4 and Appendix A.4 select the minimum height and the number of retrieved rules using the PTB test set with GPT-4, so the 83.50 PTB number is partly a test-selected maximum; and Table 2's 37.14% unknown-rule accuracy for GPT-4/PTB conflicts with Figure 5(a)'s implied 6.64% for the same counts, an internal data-consistency problem. Neither concern makes the derivation equivalent to its inputs; the CTB5 and MCTB improvements stand as independent evidence.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a small set of hand-set hyperparameters (height threshold, number of retrieved examples) and four domain assumptions about treebank coverage, span-error correction, error taxonomy, and data contamination. The most fragile is the assumption that a missing rule is an error, which the paper's own Table 2 contradicts in part. No new entities are introduced.

free parameters (3)

Minimum correction height h = 3 (all subtrees of height >= 3)
Section 4.2 (Figure 4) sweeps h from 10 down to 3 on the PTB test set with GPT-4 and reports F1 for each; h=3 gives the best F1 and is used in all main experiments. This is a test-set-tuned hyperparameter.
Number of final rules and examples = 5 rules, 1 example each = 5 examples
Appendix A.4 compares 3 rules x 2 examples vs 5 rules x 1 example and adopts the latter based on PTB/GPT-4 results. Another test-set choice.
Base few-shot example count = 5
Section 2.1 samples five examples for few-shot parsing, following standard practice in prior work. It is a hand-set constant, not fitted in this paper, but it affects all baseline and corrected results.

assumptions (4)

domain assumption A subtree rule absent from the training treebank is likely incorrect and requires correction.
Quoted in Section 3.2 step 1; this is the trigger for all structure corrections. Table 2 shows unknown rules can be correct (e.g., GPT-4 PTB unknown-rule accuracy 37.14%), so the assumption is known to be imperfect.
domain assumption Top-down correction of subtrees at different heights will indirectly resolve span errors.
Section 3.2 step 4 says "we speculate that this process can also indirectly help resolve span errors"; Section 4.4 then attributes span-error reductions to this mechanism without a direct test.
domain assumption The four error types (span, label, flatness, deepness) cover all parsing mistakes.
Section 2.3 defines the taxonomy and states it "can encompass all the mistakes made by the parsers", while compound errors inside an already classified subtree are ignored. Error counts and the correction coverage argument rely on this.
domain assumption Test sets are not memorized by the evaluated LLMs, including GPT-4.
Appendix A.3 tests contamination only with Llama-8B via the MKP metric; the assumption is extended to GPT-3.5 and GPT-4 without a specific test.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Self-Correction Makes LLMs Better Parsers." pith.science (2026). https://pith.science/paper/HO5EHTEA

@misc{pith2026250414165,
  author       = {Pith},
  title        = {Pith review of: Self-Correction Makes LLMs Better Parsers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HO5EHTEA}},
  note         = {Machine review of arXiv:2504.14165}
}

read the original abstract

Large language models (LLMs) have achieved remarkable success across various natural language processing (NLP) tasks. However, recent studies suggest that they still face challenges in performing fundamental NLP tasks essential for deep language understanding, particularly syntactic parsing. In this paper, we conduct an in-depth analysis of LLM parsing capabilities, delving into the specific shortcomings of their parsing results. We find that LLMs may stem from limitations to fully leverage grammar rules in existing treebanks, which restricts their capability to generate valid syntactic structures. To help LLMs acquire knowledge without additional training, we propose a self-correction method that leverages grammar rules from existing treebanks to guide LLMs in correcting previous errors. Specifically, we automatically detect potential errors and dynamically search for relevant rules, offering hints and examples to guide LLMs in making corrections themselves. Experimental results on three datasets with various LLMs, demonstrate that our method significantly improves performance in both in-domain and cross-domain settings on the English and Chinese datasets.

Figures

Figures reproduced from arXiv: 2504.14165 by the authors.

**Figure 1.** Four types of errors. be involved when identifying label errors. Therefore, span error represents that the boundary of the parent node does not align with the gold. For instance, the span of the subtree “a young man” in [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 3.** The process of structure correction. For clarify, we provide a typical example of rule for each processing [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: The effect of unmatch correction and structure [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 6.** Figure 6: The prompt for few-shot learning. Type Error LLM Output Hint Length Unmatch extra (S (NP (DT That)) (VP (VBZ 's) (ADJP (RB really) (RB so) (JJ cool)))) Do not additionally add any words, especially “really” missing (S (NP (DT That)) (VP (VBZ 's) (ADJP (JJ cool)))) Do n…

**Figure 7.** Figure 7: The process of Unmatch Correction. A Appendix A.1 Prompt for Constituency Parsing Previous research (Bai et al., 2023; Tian et al., 2024) has demonstrated that linearizing constituency trees into sequences is effective when using LLMs for tree parsing. Therefore, foll…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

SAC-Opt: Semantic Anchors for Iterative Correction in Optimization Modeling
cs.AI 2025-09 conditional novelty 6.0 of 10

A semantic-anchor alignment loop that reconstructs and verifies constraints from generated solver code lifts average modeling accuracy by 7.8 percentage points over seven datasets.
Lost in Speech: Benchmarking, Evaluation, and Parsing of Spoken Bilingual Conversational Language Beyond Standard UD Assumptions
cs.CL 2026-02 reject novelty 5.0 of 10

A 126-sentence spoken code-switching benchmark, a flexible scoring metric, and an agentic LLM parser are introduced, but the paper's tables contradict its claimed accuracy gains.

Reference graph

Works this paper leans on

14 extracted references · 6 canonical work pages · cited by 2 Pith papers

[5]

arXiv preprint arXiv:2310.01798

Large language models cannot self-correct reasoning yet. arXiv preprint arXiv:2310.01798. Nikita Kitaev and Dan Klein

arXiv
[6]

In Proceed- ings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 8174–8185, Singapore

LLM-enhanced self-training for cross-domain constituency parsing. In Proceed- ings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 8174–8185, Singapore. Association for Computational Linguis- tics. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumo...

work page 2023
[7]

In Find- ings of the Association for Computational Linguis- tics: EMNLP 2023, pages 12009–12024, Singapore

Leveraging GPT- 4 for automatic translation post-editing. In Find- ings of the Association for Computational Linguis- tics: EMNLP 2023, pages 12009–12024, Singapore. Association for Computational Linguistics. Weijia Shi, Anirudh Ajith, Mengzhou Xia, Yangsibo Huang, Daogao Liu, Terra Blevins, Danqi Chen, and Luke Zettlemoyer

work page 2023
[8]

Preprint, arXiv:2310.16789

Detecting pretrain- ing data from large language models. Preprint, arXiv:2310.16789. Yuanhe Tian, Fei Xia, and Yan Song

arXiv
[9]

Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou

Llms cannot find rea- soning errors, but can correct them! arXiv preprint arXiv:2311.08516. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou

arXiv
[10]

Preprint, arXiv:2203.11171

Self-consistency improves chain of thought reasoning in language models. Preprint, arXiv:2203.11171. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou

arXiv
[11]

arXiv preprint arXiv:2407.10671

Qwen2 technical report. arXiv preprint arXiv:2407.10671. Kevin Yang, Yuandong Tian, Nanyun Peng, and Dan Klein. 2022a. Re3: Generating longer stories with recursive reprompting and revision. In Proceedings of the 2022 Conference on Empirical Methods in Nat- ural Language Processing, pages 4393–4479, Abu Dhabi, United Arab Emirates. Association for Com- pu...

arXiv 2022
[12]

really” missing (S (NP (DT That)) (VP (VBZ 's) (ADJP (JJ cool)))) Do not omit any words, especially “so

How well do large language models understand syn- tax? an evaluation by asking natural language ques- tions. Preprint, arXiv:2311.08287. LLM Prompt Y ou will be given one sentence for constituency parsing. Every word that is separated by a space should be considered an independent word and have its own constituency label. Please parse the sentence with gi...

arXiv 2023

Show all 14 references

[13]

(S (NP (PRP He)) (VP (MD could) (RB not) (VP (VB speak))))

has demonstrated that linearizing con- stituency trees into sequences is effective when using LLMs for tree parsing. Therefore, follow- ing Bai et al. (2023), we represent constituency tree structures using bracket notation, such as “(S (NP (PRP He)) (VP (MD could) (RB not) (V...

2023
[14]

Length unmatch means the number of words in the predicted tree differ from the original sentence, while word unmatch in- dicates a word in the sentence has been altered

A.2 Hint for Unmatch Correction Following the evaluation script of EV ALB, we cat- egorize unmatch errors into two types: length un- match and word unmatch. Length unmatch means the number of words in the predicted tree differ from the original sentence, while word unmatch in-...

2023
[2020]

In Ad- vances in Neural Information Processing Systems , volume 33, pages 1877–1901

Language models are few-shot learners. In Ad- vances in Neural Information Processing Systems , volume 33, pages 1877–1901. Curran Associates, Inc. Cheng-Han Chiang and Hung-yi Lee

1901
[2021]

In Proceedings of the Society for Computation in Linguistics 2021, pages 323–333, Online

What’s in a span? evaluating the creativity of a span-based neural constituency parser. In Proceedings of the Society for Computation in Linguistics 2021, pages 323–333, Online. Association for Computational Linguistics. DeepSeek-AI, Aixin Liu, Bei Feng, Bing Xue, Bingx- uan W...

2021
[2023]

Preprint, arXiv:2310.19462

Constituency parsing using llms. Preprint, arXiv:2310.19462. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom H...
[2024]

Preprint, arXiv:2412.19437

Deepseek-v3 technical report. Preprint, arXiv:2412.19437. Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Chaganty, Yicheng Fan, Vincent Zhao, Ni Lao, Hongrae Lee, Da-Cheng Juan, and Kelvin Guu. 2023a. RARR: Researching and revis- ing what language models sa...

arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [5]

arXiv preprint arXiv:2310.01798

Large language models cannot self-correct reasoning yet. arXiv preprint arXiv:2310.01798. Nikita Kitaev and Dan Klein

arXiv

[2] [6]

In Proceed- ings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 8174–8185, Singapore

LLM-enhanced self-training for cross-domain constituency parsing. In Proceed- ings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 8174–8185, Singapore. Association for Computational Linguis- tics. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumo...

work page 2023

[3] [7]

In Find- ings of the Association for Computational Linguis- tics: EMNLP 2023, pages 12009–12024, Singapore

Leveraging GPT- 4 for automatic translation post-editing. In Find- ings of the Association for Computational Linguis- tics: EMNLP 2023, pages 12009–12024, Singapore. Association for Computational Linguistics. Weijia Shi, Anirudh Ajith, Mengzhou Xia, Yangsibo Huang, Daogao Liu, Terra Blevins, Danqi Chen, and Luke Zettlemoyer

work page 2023

[4] [8]

Preprint, arXiv:2310.16789

Detecting pretrain- ing data from large language models. Preprint, arXiv:2310.16789. Yuanhe Tian, Fei Xia, and Yan Song

arXiv

[5] [9]

Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou

Llms cannot find rea- soning errors, but can correct them! arXiv preprint arXiv:2311.08516. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou

arXiv

[6] [10]

Preprint, arXiv:2203.11171

Self-consistency improves chain of thought reasoning in language models. Preprint, arXiv:2203.11171. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou

arXiv

[7] [11]

arXiv preprint arXiv:2407.10671

Qwen2 technical report. arXiv preprint arXiv:2407.10671. Kevin Yang, Yuandong Tian, Nanyun Peng, and Dan Klein. 2022a. Re3: Generating longer stories with recursive reprompting and revision. In Proceedings of the 2022 Conference on Empirical Methods in Nat- ural Language Processing, pages 4393–4479, Abu Dhabi, United Arab Emirates. Association for Com- pu...

arXiv 2022

[8] [12]

really” missing (S (NP (DT That)) (VP (VBZ 's) (ADJP (JJ cool)))) Do not omit any words, especially “so

How well do large language models understand syn- tax? an evaluation by asking natural language ques- tions. Preprint, arXiv:2311.08287. LLM Prompt Y ou will be given one sentence for constituency parsing. Every word that is separated by a space should be considered an independent word and have its own constituency label. Please parse the sentence with gi...

arXiv 2023