Pith. sign in

REVIEW 4 major objections 5 minor 4 cited by

A decoding-time method that keeps the format in a separate module improves LLM accuracy by 1–6% with guaranteed format compliance on math, judging, and extraction.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 11:36 UTC pith:P3BLKF4T

load-bearing objection A useful, honest extension of CtrlG-style control to template constraints, with solid engineering, but the HMM lookahead is uncalibrated and the gains are small — worth reviewing, not accepting as is. the 4 major comments →

arxiv 2510.03595 v2 pith:P3BLKF4T submitted 2025-10-04 cs.CL

Decoupling Task-Solving and Output Formatting in LLM Generation

Haikang Deng , Po-Nien Kung , Nanyun Peng This is my paper
classification cs.CL
keywords formatdeco-greasoningadherencecomplianceformattinggenerationintroduce
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Large language models often do worse when a prompt tells them both what to solve and exactly how to format the answer. This paper's method, DECO-G, splits those two jobs. The language model sees only the task; a separate 'Format Estimation Module' (FEM) holds the format rules. At every step, the FEM looks ahead—using a simplified model of the LLM's own writing habits—and boosts tokens that lead to a compliant output, while the main model keeps choosing tokens that make sense for solving the problem.

The simplified model is a hidden Markov model trained on one million responses the LLM gave to general instructions. The format rules are turned into a state machine with a trie-building trick, so templates like 'The final answer is ...' can be enforced cheaply. To keep it fast, the method keeps only the most likely hidden states when computing compliance, cutting the extra compute to a small fraction of the main model's cost.

On GSM8k math, SummEval LLM-as-judge, and ACE05 event extraction, DECO-G improves accuracy, F1, or correlation over prompting with format constraints and over hard structured generation, and hits 100% format compliance. The gains are 1–6%. The main caveats are that the steering strength had to be increased for one model family, and the code is not yet released.

Core claim

The central claim is that DECO-G, by separating format adherence from task solving, yields '1.0% to 6.0% relative gain over regular prompting practice with guaranteed format compliance' (Abstract) and 'achieves the best performance over all three models' on GSM8k (Section 4.1). If the paper is correct, a decoding-time module can enforce complex templates without degrading—indeed while improving—the LLM's underlying task performance, and does so at a small fraction of the main model's FLOPs.

Load-bearing premise

The load-bearing assumption is that an HMM distilled on one million generic instruction-response pairs from Natural-Instructions-v2 (Section 3.1, Appendix H) is a sufficiently accurate approximation of the LLM's token distribution on the specific target tasks, even though the HMM is run with an empty input and ignores the actual task prompt. If this approximation is miscalibrated, the FEM's lookahead—PFEM(α|x_t, x_<t)—will not reflect true future compliance likelihood and could steer generation badly, as the paper shows CtrlG doing in Appendix C. This assumption is structurally separate from the claim itself and is the most fragile part of the method.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes DECO-G, a decoding-time framework that removes format instructions from the LLM prompt and delegates format enforcement to a separately distilled HMM (the FEM). At each step, token probabilities are reweighted by PFEM(α|·)^γ, where the FEM computes the probability of eventual format compliance via a DFA-constrained HMM lookahead. Three efficiency contributions are claimed: instruction-aware HMM distillation, trie-based DFA construction, and top-k hidden-state pruning. Experiments on GSM8k, SummEval, and ACE05 across Llama-3.1-8B-Instruct and two Qwen models report 100% format compliance and 1.0–6.0% relative gains over prompting and structured-generation baselines.

Significance. The idea is valuable: if the FEM lookahead is a faithful estimate of the LLM's own compliance probability, Eq. (6) is a principled Bayes rearrangement that could reduce format-induced degradation while preserving exact format enforcement. The trie construction and top-k pruning analysis are concrete engineering contributions, and the appendix provides prompt templates and FLOPs estimates. However, the central empirical claim is currently under-supported: no calibration check ties the HMM's prompt-independent lookahead to task-conditioned LLM behavior, no error bars or significance tests accompany the small reported differences, and a key hyperparameter is set per model after observing the default's failure. These are fixable but load-bearing.

major comments (4)
  1. [§3.1, §3.3, Appendix H, Eq. (6)] The DECO-G posterior is a Bayes posterior only if PFEM(α|xt,x<t) approximates the target LLM's PLM(α|xt,x<t). Section 3.1 describes instruction-aware distillation, but Appendix H clarifies that the HMM is trained on LLM continuations only and is run with an empty input; conditioning the HMM on the prompt yields 'almost identical' results. That establishes prompt invariance of the HMM, not calibration to the task-conditioned LLM. No experiment compares FEM compliance estimates with empirical compliance rates of the LLM on the evaluated tasks. Appendix C shows precisely the failure mode of an unconditioned HMM: on GSM8k it forces the template early and drops accuracy to 60.6%. Without a calibration study, the gains in Tables 1–3 may be template-specific artifacts of the generic HMM rather than evidence for the decoupling mechanism.
  2. [Tables 1–3, 5–6] No error bars, significance tests, or repeated-seed variation are reported. Many headline differences are small: in Table 3, Llama DECO-G has average Spearman 0.418 vs. JSON-S 0.414, Qwen3 DECO-G has 0.507 vs. NL-S 0.499, and DECO-G is below JSON-S on Llama Consistency (0.439 vs. 0.447) and Fluency (0.331 vs. 0.334). These gaps are within typical evaluation noise for correlation metrics and F1. The abstract's claim of 'constant gains' over baselines is therefore not supported by the evidence as reported.
  3. [§5.2, Table 1, Appendix A] The steering strength γ is set to 2 for all Qwen models after observing that γ=1 is insufficient. No γ=1 results or sensitivity sweep are reported, and Appendix A acknowledges that finding the optimal γ may require empirical exploration. Since γ is a model-specific free parameter chosen after seeing test-set performance, the reported gains are conditional on a selection procedure not specified in the method. The paper should report results across a range of γ and state a principled selection rule (e.g., a small validation set).
  4. [§4.1, Appendix F Table 7, Figure 2] The evaluation may not actually compare decoupled vs. entangled prompts. The GSM8k 'task instruction' in Table 7 includes 'the final answer must be only a number', which is itself a format constraint, and the NL format instruction specifies 'The final answer is ...'. If DECO-G uses the same task instruction, the LLM is still told the output format, contrary to the paper's claim that it receives only task-specific information. If DECO-G uses a stripped prompt, that prompt is not shown. The exact DECO-G prompt for each task must be specified and shown to be free of format instructions; otherwise, the source of the gains is confounded with prompt differences.
minor comments (5)
  1. [§4.1, §5.2] Inconsistent notation: Section 4.1 says 'raise the control factor λ' while the method and Section 5.2 use γ. Also 'tamplate' should be 'template'.
  2. [§3.3, Eq. (6)] The conditional order is written inconsistently: Eq. (5) and the text use PFEM(α|xt,x<t), while Eq. (6) writes PFEM(α|x<t,xt). Align the notation.
  3. [Table 2] The Qwen3-8B JSON row reports '3.01' in the AI+ column; Table 5 shows the correct value is 30.1. This typo should be corrected.
  4. [§2.2, §5.1] Typos: 'langauge' in §2.2 and 'encuraging' in §5.1. The paper would also benefit from proofreading of hyphenation and spacing in the appendix.
  5. [Appendix H] The statement that empty-input HMM scores are reported 'in accordance with the distillation objective' is confusing given §3.1's 'instruction-aware' framing. Add a sentence explaining why the prompt is omitted at inference despite being used during distillation.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The method rests on an HMM proxy trained on generic instruction data and on the assumption that format constraints can be separated from task solving. The main free parameters are γ, h, k, and the distillation corpus size. No new physical or theoretical entities are introduced; the FEM is a software component, not an invented entity.

free parameters (4)
  • Steering strength γ = γ=1 for Llama, γ=2 for Qwen
    Chosen per model family; the paper justifies via an entropy hypothesis but reports no held-out validation. The same γ is applied across tasks for Qwen, and γ=2 was set after observing γ=1 underperformed on GSM8k.
  • HMM hidden state size h = 4096
    Fixed by design; no sensitivity analysis is presented. Larger h raises the cost of the state-transition computation.
  • Pruning top-k = k=200 (top 5% of hidden states)
    Chosen to retain ≈98% probability mass; only one threshold is tested, and the effect on downstream accuracy is reported only for GSM8k.
  • HMM distillation corpus size = 1M completions from 1,000 Natural-Instructions-v2 prompts
    An arbitrary large scale; no study of how control quality varies with distillation data size.
axioms (4)
  • domain assumption An HMM distilled on generic instruction-following responses, run with empty input, is a sufficient proxy for the target-task-conditioned LLM distribution for estimating format compliance.
    Invoked in Section 3.1 and Appendix H; the HMM ignores the actual task prompt, so the lookahead is not conditioned on the specific problem.
  • domain assumption Task instructions and format constraints can be cleanly separated without loss.
    Central design premise; contradicted by the GSM8k task prompts in Table 7, which include 'the final answer must be only a number' inside the task instruction.
  • standard math The trie-based DFA construction correctly encodes all valid format-compliant sequences.
    Section 3.2; needed for the 100% format-compliance guarantee. No formal correctness proof is given.
  • standard math The backward recurrence for marginalizing the HMM over the DFA is correct.
    Taken from Zhang et al. (2024) and reproduced in Appendix B, Eq. (9); the paper does not prove it independently.

pith-pipeline@v1.3.0-alltime-deepseek · 18782 in / 12589 out tokens · 101669 ms · 2026-08-04T11:36:43.308826+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Decoupling Task-Solving and Output Formatting in LLM Generation." pith.science (2026). https://pith.science/paper/P3BLKF4T

@misc{pith2026251003595,
  author       = {Pith},
  title        = {Pith review of: Decoupling Task-Solving and Output Formatting in LLM Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P3BLKF4T}},
  note         = {Machine review of arXiv:2510.03595}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large language models (LLMs) are increasingly adept at solving complex problems, such as mathematical reasoning and automatic evaluation. However, performance often degrades when prompts intertwine task instructions with rigid formatting requirements. This entanglement creates competing goals for the model, hindering its reasoning capabilities. To address this, we introduce Deco-G, a decoding framework that explicitly decouples format adherence from problem solving. Deco-G delegates format adherence to a separate Format Estimation Module (FEM), which performs probabilistic lookahead to estimate future format compliance rate and reweighs token probabilities, allowing the LLM to focus solely on task resolution. To make this approach both practical and efficient, we introduce three key innovations: instruction-aware distillation, a flexible trie-building algorithm, and HMM state pruning. Experiments across mathematical reasoning, event argument extraction, and LLM-as-a-judge demonstrate that Deco-G constantly gains over prompting or structured generation baselines, with guaranteed format compliance. We release our code at https://github.com/haikangdeng/deco-g.

Figures

Figures reproduced from arXiv: 2510.03595 by Haikang Deng, Nanyun Peng, Po-Nien Kung.

Figure 1
Figure 1. Figure 1: Example of GSM8k responses. LLM prompted without format constraints gets the correct [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: DECO-G decouples task and format—prompting LLM with task-only information and sending format constraints to FEM. DECO-G decodes from the posterior constructed by multiplying LLM token probabilities with FEM estimated satisfaction rate. Probabilistic reasoning over logical constraints. The core idea of these prior frameworks is to use the TPM to perform a probabilistic lookahead—that is, to efficiently comp… view at source ↗
Figure 3
Figure 3. Figure 3: DECO-G steers Llama to generate predefined template “The final answer is ...” by boosting probabilities of template tokens. Llama-3.1-8B-Instruct Qwen2.5-7B-Instruct Qwen3-8B 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Entropy (bits) NL NL-S JSON JSON-S Deco-G NL NL-S JSON JSON-S Deco-G NL NL-S JSON JSON-S Deco-G Comparison of Model Entropy [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: LLM’s token-level entropy for different models and methods. [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Average retention rate (of total mass) over [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 4 Pith papers

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

  1. The Extrapolation Cliff in On-Policy Distillation of Near-Deterministic Structured Outputs

    cs.LG 2026-05 unverdicted novelty 7.0

    On-policy distillation has an extrapolation cliff at closed-form lambda*(p,b,c) set by teacher modal probability, warm-start mass, and clip strength, past which training shifts from format-preserving to format-collapsing.

  2. Constraint Tax in Open-Weight LLMs: An Empirical Study of Tool Calling Suppression Under Structured Output Constraints

    cs.CL 2026-06 conditional novelty 6.0

    Open-weight LLMs exhibit tool suppression under joint tool-calling and JSON-schema constraints due to grammar token masking; a two-pass inference method restores tool use.

  3. Decoupled Smart Contract Audits: Lightweight LLM Framework via Distillation and Aggregation

    cs.CR 2026-06 unverdicted novelty 5.0

    A decoupled four-stage LLM pipeline with rsLoRA, distillation, and CoVe aggregation outperforms larger models on smart contract vulnerability detection and explanation using only 0.6B-4B parameter models.

  4. It's Not the Capability: Harness Sensitivity Is Non-Monotone Across LLM Agent Tiers

    cs.AI 2026-05 unverdicted novelty 5.0

    A 432-run experiment across capability tiers refutes the assumption of a monotone inverse relationship between LLM capability and optimal harness complexity, showing model-type-specific patterns instead.

Reference graph

Works this paper leans on

47 extracted references · 18 linked inside Pith · cited by 4 Pith papers

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Director: Generator-classifiers for supervised language modeling

    Kushal Arora, Kurt Shuster, Sainbayar Sukhbaatar, and Jason Weston. Director: Generator-classifiers for supervised language modeling. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pp.\ 512-...

    2022

  3. [3]

    An inequality and associated maximization technique in statistical estimation for probabilistic functions of markov processes

    Leonard E Baum et al. An inequality and associated maximization technique in statistical estimation for probabilistic functions of markov processes. Inequalities, 3 0 (1): 0 1--8, 1972

    1972

  4. [4]

    Guiding LLM s the right way: Fast, non-invasive constrained generation

    Luca Beurer-Kellner, Marc Fischer, and Martin Vechev. Guiding LLM s the right way: Fast, non-invasive constrained generation. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp (eds.), Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of...

    2024

  5. [5]

    Scaling instruction-finetuned language models

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. Journal of Machine Learning Research, 25 0 (70): 0 1--53, 2024

    2024

  6. [6]

    Training verifiers to solve math word problems

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021

    Pith/arXiv arXiv 2021

  7. [7]

    Safe rlhf: Safe reinforcement learning from human feedback

    Josef Dai, Xuehai Pan, Ruiyang Sun, Jiaming Ji, Xinbo Xu, Mickel Liu, Yizhou Wang, and Yaodong Yang. Safe rlhf: Safe reinforcement learning from human feedback. In The Twelfth International Conference on Learning Representations, 2024

    2024

  8. [8]

    Plug and play language models: A simple approach to controlled text generation

    Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu. Plug and play language models: A simple approach to controlled text generation. arXiv preprint arXiv:1912.02164, 2019

    Pith/arXiv arXiv 1912

  9. [9]

    Reward-augmented decoding: Efficient controlled text generation with a unidirectional reward model

    Haikang Deng and Colin Raffel. Reward-augmented decoding: Efficient controlled text generation with a unidirectional reward model. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp.\ 11781--11791, 2023

    2023

  10. [10]

    The automatic content extraction (ace) program-tasks, data, and evaluation

    George R Doddington, Alexis Mitchell, Mark A Przybocki, Lance A Ramshaw, Stephanie M Strassel, and Ralph M Weischedel. The automatic content extraction (ace) program-tasks, data, and evaluation. In Lrec, volume 2, pp.\ 837--840. Lisbon, 2004

    2004

  11. [11]

    Summeval: Re-evaluating summarization evaluation

    Alexander R Fabbri, Wojciech Kry \'s ci \'n ski, Bryan McCann, Caiming Xiong, Richard Socher, and Dragomir Radev. Summeval: Re-evaluating summarization evaluation. Transactions of the Association for Computational Linguistics, 9: 0 391--409, 2021

    2021

  12. [12]

    The llama 3 herd of models

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Guidance: A guidance language for controlling large language models

    guidance-ai . Guidance: A guidance language for controlling large language models. https://github.com/guidance-ai/guidance, 2024

    2024

  14. [14]

    Does prompt formatting have any impact on llm performance?, 2024

    Jia He, Mukund Rungta, David Koleczek, Arshdeep Sekhon, Franklin X Wang, and Sadid Hasan. Does prompt formatting have any impact on llm performance?, 2024. URL https://arxiv.org/abs/2411.10541

    Pith/arXiv arXiv 2024

  15. [15]

    Measuring massive multitask language understanding

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020

    Pith/arXiv arXiv 2009

  16. [16]

    Scaling laws for neural language models

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020

    Pith/arXiv arXiv 2001

  17. [17]

    Ctrl: A conditional transformer language model for controllable generation

    Nitish Shirish Keskar, Bryan McCann, Lav R Varshney, Caiming Xiong, and Richard Socher. Ctrl: A conditional transformer language model for controllable generation. arXiv preprint arXiv:1909.05858, 2019

    Pith/arXiv arXiv 1909

  18. [18]

    Generalization through memorization: Nearest neighbor language models

    Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generalization through memorization: Nearest neighbor language models. In International Conference on Learning Representations, 2021

    2021

  19. [19]

    Gedi: Generative discriminator guided sequence generation

    Ben Krause, Akhilesh Deepak Gotmare, Bryan McCann, Nitish Shirish Keskar, Shafiq Joty, Richard Socher, and Nazneen Fatema Rajani. Gedi: Generative discriminator guided sequence generation. In Findings of the Association for Computational Linguistics: EMNLP 2021, pp.\ 4929--4952, 2021

    2021

  20. [20]

    The power of scale for parameter-efficient prompt tuning

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pp.\ 3045--3059, 2021

    2021

  21. [21]

    Prefix-tuning: Optimizing continuous prompts for generation

    Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pp.\ 4582--4597, 2021

    2021

  22. [22]

    Dexperts: Decoding-time controlled text generation with experts and anti-experts

    Alisa Liu, Maarten Sap, Ximing Lu, Swabha Swayamdipta, Chandra Bhagavatula, Noah A Smith, and Yejin Choi. Dexperts: Decoding-time controlled text generation with experts and anti-experts. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Vo...

    2021

  23. [23]

    G-eval: Nlg evaluation using gpt-4 with better human alignment

    Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. G-eval: Nlg evaluation using gpt-4 with better human alignment. arXiv preprint arXiv:2303.16634, 2023

    Pith/arXiv arXiv 2023

  24. [24]

    Chen, and Min-Yen Kan

    Do Xuan Long, Ngoc-Hai Nguyen, Tiviatis Sim, Hieu Dao, Shafiq Joty, Kenji Kawaguchi, Nancy F. Chen, and Min-Yen Kan. LLM s are biased towards output formats! systematically evaluating and mitigating output format bias of LLM s. In Luis Chiruzzo, Alan Ritter, and Lu Wang (eds.), Proceedings of the 2025 Conference of the Nations of the Americas Chapter of t...

    2025

  25. [25]

    Cross-task generalization via natural language crowdsourcing instructions

    Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. Cross-task generalization via natural language crowdsourcing instructions. In ACL, 2022

    2022

  26. [26]

    ChatGPT , 2025

    OpenAI . ChatGPT , 2025. URL https://chat.openai.com. Large language model. Accessed May 19, 2025

    2025

  27. [27]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 0 27730--27744, 2022

    2022

  28. [28]

    Sutherland

    Yi Ren and Danica J. Sutherland. Learning dynamics of llm finetuning, 2025. URL https://arxiv.org/abs/2407.10490

    Pith/arXiv arXiv 2025

  29. [29]

    Self-diagnosis and self-debiasing: A proposal for reducing corpus-based bias in nlp

    Timo Schick, Sahana Udupa, and Hinrich Sch \"u tze. Self-diagnosis and self-debiasing: A proposal for reducing corpus-based bias in nlp. Transactions of the Association for Computational Linguistics, 9: 0 1408--1424, 2021

    2021

  30. [30]

    Classifiers are better experts for controllable text generation

    Askhat Sitdikov, Nikita Balagansky, Daniil Gavrilov, and Alexander Markov. Classifiers are better experts for controllable text generation. arXiv preprint arXiv:2205.07276, 2022

    Pith/arXiv arXiv 2022

  31. [31]

    Learning to summarize with human feedback

    Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. Advances in Neural Information Processing Systems, 33: 0 3008--3021, 2020

    2020

  32. [32]

    Let me speak freely? a study on the impact of format restrictions on performance of large language models, 2024

    Zhi Rui Tam, Cheng-Kuang Wu, Yi-Lin Tsai, Chieh-Yen Lin, Hung yi Lee, and Yun-Nung Chen. Let me speak freely? a study on the impact of format restrictions on performance of large language models, 2024. URL https://arxiv.org/abs/2408.02442

    Pith/arXiv arXiv 2024

  33. [33]

    Finetuned language models are zero-shot learners

    Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021

    Pith/arXiv arXiv 2021

  34. [34]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 0 24824--24837, 2022

    2022

  35. [35]

    Grace: gradient-guided controllable retrieval for augmenting attribute-based text generation

    Zhihua Wen, Zhiliang Tian, Zhen Huang, Yuxin Yang, Zexin Jian, Changjian Wang, and Dongsheng Li. Grace: gradient-guided controllable retrieval for augmenting attribute-based text generation. In Findings of the Association for Computational Linguistics: ACL 2023, pp.\ 8377--8398, 2023

    2023

  36. [36]

    Efficient guided generation for llms

    Brandon T Willard and R \'e mi Louf. Efficient guided generation for llms. arXiv preprint arXiv:2307.09702, 2023

    Pith/arXiv arXiv 2023

  37. [37]

    Qwen3 technical report, 2025 a

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, ...

    Pith/arXiv arXiv 2025

  38. [38]

    Fudge: Controlled text generation with future discriminators

    Kevin Yang and Dan Klein. Fudge: Controlled text generation with future discriminators. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pp.\ 3511--3535, 2021

    2021

  39. [39]

    Qwen2.5 technical report, 2025 b

    Qwen: An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Ti...

    Pith/arXiv arXiv 2025

  40. [40]

    Tree of thoughts: Deliberate problem solving with large language models

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36: 0 11809--11822, 2023

    2023

  41. [41]

    Technical report: Auxiliary tuning and its application to conditional text generation

    Yoel Zeldes, Dan Padnos, Or Sharir, and Barak Peleg. Technical report: Auxiliary tuning and its application to conditional text generation. arXiv preprint arXiv:2006.16823, 2020

    Pith/arXiv arXiv 2006

  42. [42]

    Token-level direct preference optimization

    Yongcheng Zeng, Guoqing Liu, Weiyu Ma, Ning Yang, Haifeng Zhang, and Jun Wang. Token-level direct preference optimization. In Proceedings of the 41st International Conference on Machine Learning, pp.\ 58348--58365, 2024

    2024

  43. [43]

    Tractable control for autoregressive language generation, 2023

    Honghua Zhang, Meihua Dang, Nanyun Peng, and Guy Van den Broeck. Tractable control for autoregressive language generation, 2023. URL https://arxiv.org/abs/2304.07438

    Pith/arXiv arXiv 2023

  44. [44]

    Adaptable logical control for large language models, 2024

    Honghua Zhang, Po-Nien Kung, Masahiro Yoshida, Guy Van den Broeck, and Nanyun Peng. Adaptable logical control for large language models, 2024. URL https://arxiv.org/abs/2406.13892

    Pith/arXiv arXiv 2024

  45. [45]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

  46. [46]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

  47. [47]

    However, as prompts grow more complex, models often struggle to adhere to all instructions

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

    arXiv 2025