arxiv: 2605.05893 · v1 · submitted 2026-05-07 · 💻 cs.CL · cs.AI

Recognition: unknown

Logic-Regularized Verifier Elicits Reasoning from LLMs

Xinyu Wang , Changzhi Sun , Lian Cheng , Yuanbin Wu , Dell Zhang , Xiaoling Wang , Xuelong Li

Authors on Pith no claims yet

Pith reviewed 2026-05-08 10:51 UTC · model grok-4.3

classification 💻 cs.CL cs.AI

keywords LLM reasoningunsupervised verifierlogical consistencybinary latent variablereasoning pathsinternal activationsnegation consistency

0 comments

The pith

Logic rules on reasoning paths let LLMs build their own verifiers without labels.

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

LOVER learns a binary verifier directly from an LLM's internal activations by treating it as a latent variable and enforcing three logical constraints across multiple generated reasoning paths. The constraints require that negation of a path flips its correctness label, that paths sharing the same final answer receive the same label, and that paths with different answers receive opposite labels. These rules serve as priors that replace any need for human-annotated supervision, allowing the method to use only unlabeled examples and to work with any off-the-shelf model. Across ten reasoning datasets the resulting verifier beats other unsupervised selection methods and reaches 95 percent of the accuracy obtained by a fully supervised verifier.

Core claim

The paper shows that a binary verifier can be elicited by optimizing a latent variable to satisfy negation consistency, intra-group consistency, and inter-group consistency on reasoning paths grouped by their final answers, using only the model's internal activations and unlabeled data.

What carries the argument

The logic-regularized binary latent verifier that applies three consistency constraints to multiple reasoning paths grouped by final answer.

Load-bearing premise

Enforcing the three logical constraints on multiple reasoning paths is sufficient to elicit a reliable binary verifier from internal activations without labeled supervision.

What would settle it

If LOVER performs no better than simply selecting the highest-probability path on a new dataset where the logical constraints are frequently violated by correct answers, the claim that the constraints suffice would be falsified.

Figures

Figures reproduced from arXiv: 2605.05893 by Changzhi Sun, Dell Zhang, Lian Cheng, Xiaoling Wang, Xinyu Wang, Xuelong Li, Yuanbin Wu.

**Figure 1.** Figure 1: An illustration of our proposed LOVER. For any question q, we create xi by combining q with the i-th solution from N solutions. We form x + i and x − i by adding "This is a true/false answer." to xi , respectively. Choosing the correct solution involves determining which assertion, x + i or x − i , is correct. The hidden states of LLMs are used to represent x + i and x − i , which are then input into LOVER… view at source ↗

**Figure 2.** Figure 2: An accuracy comparison of LOVER and baselines across different numbers of solutions on GSM8K over mistral-7b. LOVER maintains high reasoning accuracy regardless of N. The reasoning accuracy of all methods increases as the number of solutions grows shown in view at source ↗

**Figure 3.** Figure 3: LOVER reliably improves reasoning performance across model scales (Qwen-2.5 family). hances reasoning accuracy across different model scales over the Qwen-2.5 family. LOVER enables a 7B-parameter LLM to achieve reasoning accuracy comparable to that of a 32B-parameter LLM. LOVER achieves an average accuracy gains of 1.7% across five models with varying parameter sizes and consistently outperforms the base… view at source ↗

read the original abstract

Verifiers are crucial components for enhancing modern LLMs' reasoning capability. Typicalverifiers require resource-intensive superviseddataset construction, which is costly and faceslimitations in data diversity. In this paper, wepropose LOVER, an unsupervised verifier regularized by logical rules. LOVER treats theverifier as a binary latent variable, utilizinginternal activations and enforcing three logical constraints on multiple reasoning paths:negation consistency, intra-group consistency,and inter-group consistency (grouped by thefinal answer). By incorporating logical rulesas priors, LOVER can leverage unlabeled examples and is directly compatible with any offthe-shelf LLMs. Experiments on 10 datasetsdemonstrate that LOVER significantly outperforms unsupervised baselines, achieving performance comparable to the supervised verifier(reaching its 95% level on average). The sourcecode is publicly available at https://github.com/wangxinyufighting/llm-lover.

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

LOVER adds three logical consistency rules to an unsupervised latent verifier and reports gains on 10 datasets, but the rules may be satisfiable by answer-grouping heuristics that ignore activations.

read the letter

The paper's core move is to treat the verifier as a binary latent variable over LLM reasoning paths and regularize it with three logical constraints—negation consistency, intra-group consistency, and inter-group consistency grouped by final answer—using only unlabeled data. This lets it train without supervised labels while staying compatible with off-the-shelf models. The experiments claim it beats unsupervised baselines and reaches about 95% of a supervised verifier's performance across 10 datasets, with public code released. That combination of latent framing plus the specific named constraints is not in the cited prior verifier work, so the formulation itself is new. The public code and multi-dataset evaluation are also useful for anyone trying to reproduce or extend it. The main soft spot is whether the constraints actually elicit a verifier that reads internal activations to judge reasoning quality. The inter-group term groups paths by extracted answer and pushes consistency within those groups; a simple majority-vote or answer-frequency heuristic could satisfy it without ever consulting the activation features. Intra-group and negation terms look similarly open to constant or majority solutions. If the paper contains ablations that isolate the activation input or show that removing it collapses performance, that would address the concern directly. From the abstract and the stress-test framing, those controls are not described, so the reported gains could be driven more by the grouping logic than by learning a meaningful verifier. The work is aimed at researchers building reliable reasoning systems without large labeled sets. A reader interested in unsupervised verifier training or logical regularization would get value from the concrete constraints and the empirical numbers, even if they need to verify the mechanism themselves. It is coherent enough and grounded enough in a practical problem to deserve peer review, though any referee should press on the degeneracy question and ask for activation-specific ablations.

Referee Report

3 major / 2 minor

Summary. The paper proposes LOVER, an unsupervised verifier for LLM reasoning that treats the verifier as a binary latent variable over internal activations and regularizes it via three logical constraints (negation consistency, intra-group consistency, and inter-group consistency) applied to multiple reasoning paths grouped by final answer. Experiments on 10 datasets are reported to show that LOVER significantly outperforms unsupervised baselines while reaching approximately 95% of the performance of a supervised verifier on average; source code is released.

Significance. If the method genuinely elicits a non-degenerate verifier from activations using only unlabeled data and logical priors, the result would be significant for lowering the cost of supervision in LLM reasoning pipelines. Public code availability is a positive factor for reproducibility.

major comments (3)

[Method (loss definitions and optimization)] The inter-group consistency loss (defined by grouping paths according to extracted final answers) can be satisfied by a majority-vote solution that assigns the verifier output solely on the basis of group size and never consults the internal activation features. No ablation or diagnostic is presented to show that removing or randomizing the activation inputs causes a performance drop, leaving open the possibility that reported gains are driven by the grouping heuristic rather than the claimed elicitation mechanism.
[Experiments and results] The experimental claims rest on performance numbers across 10 datasets, yet the manuscript provides no information on the number of random seeds, statistical significance tests, variance across runs, or the precise functional form and weighting of the three consistency losses inside the training objective. Without these controls it is impossible to assess whether the reported 95% parity with the supervised verifier is robust.
[Method and analysis] The central assumption that the three consistency constraints are sufficient to force the latent verifier to extract a meaningful correctness signal from activations (rather than admitting constant or majority-based fixed points) is not tested. A simple sanity check—e.g., training with activations replaced by random noise or by a constant vector—would directly address whether the learned function depends on the claimed features.

minor comments (2)

[Abstract] The abstract contains a missing space ('Typicalverifiers').
[Method] Notation for the latent verifier variable and the exact form of each consistency term should be introduced with a single, consistent symbol table or equation block.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and constructive comments. We address each major point below with clarifications and commit to revisions that strengthen the empirical support for our claims.

read point-by-point responses

Referee: The inter-group consistency loss (defined by grouping paths according to extracted final answers) can be satisfied by a majority-vote solution that assigns the verifier output solely on the basis of group size and never consults the internal activation features. No ablation or diagnostic is presented to show that removing or randomizing the activation inputs causes a performance drop, leaving open the possibility that reported gains are driven by the grouping heuristic rather than the claimed elicitation mechanism.

Authors: We agree that the inter-group consistency term by itself admits a majority-vote fixed point. However, the full objective also enforces negation consistency and intra-group consistency, which together penalize solutions that ignore activation features. To directly test dependence on activations, we will add an ablation in the revised manuscript that replaces activation inputs with random noise or constant vectors while keeping all other components fixed, and report the resulting performance drop. This diagnostic will confirm that the learned verifier relies on the internal features rather than the grouping heuristic alone. revision: yes
Referee: The experimental claims rest on performance numbers across 10 datasets, yet the manuscript provides no information on the number of random seeds, statistical significance tests, variance across runs, or the precise functional form and weighting of the three consistency losses inside the training objective. Without these controls it is impossible to assess whether the reported 95% parity with the supervised verifier is robust.

Authors: We acknowledge the need for greater experimental transparency. In the revision we will specify the number of random seeds (we used 5), report mean and standard deviation across runs for all main results, include paired statistical significance tests against baselines, and explicitly state the functional form of the total loss together with the exact weighting coefficients applied to the three consistency terms. These additions will allow readers to evaluate robustness directly. revision: yes
Referee: The central assumption that the three consistency constraints are sufficient to force the latent verifier to extract a meaningful correctness signal from activations (rather than admitting constant or majority-based fixed points) is not tested. A simple sanity check—e.g., training with activations replaced by random noise or by a constant vector—would directly address whether the learned function depends on the claimed features.

Authors: This concern is well-founded and overlaps with the first comment. We will therefore include the requested sanity check—training with randomized or constant activation vectors—in the revised experiments section. The expected outcome is a substantial performance degradation relative to the original setting, which would support that the logical constraints elicit a non-degenerate correctness signal from the activations rather than permitting trivial fixed points. revision: yes

Circularity Check

0 steps flagged

No significant circularity; logical priors and activation-based optimization remain independent

full rationale

The paper defines LOVER by treating the verifier as a binary latent variable optimized over internal activations subject to three externally stated logical constraints (negation consistency, intra-group consistency, inter-group consistency) applied to unlabeled reasoning paths grouped by final answer. These constraints function as priors in the loss rather than quantities defined in terms of the verifier outputs or fitted parameters. Reported results compare LOVER against unsupervised baselines and a supervised verifier (reaching 95% of supervised performance on average across 10 datasets) without evidence that the metrics are constructed from the same parameters used to define the method or via self-citation chains. The approach is presented as directly compatible with off-the-shelf LLMs and includes public code, confirming the derivation chain is self-contained against external benchmarks rather than reducing to its inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that the three logical constraints function as effective priors for the latent verifier; no free parameters or invented entities are mentioned in the abstract.

axioms (1)

domain assumption Negation consistency, intra-group consistency, and inter-group consistency hold across multiple reasoning paths for the same question.
These three rules are introduced as logical priors that regularize the verifier without labeled data.

pith-pipeline@v0.9.0 · 5464 in / 1207 out tokens · 23518 ms · 2026-05-08T10:51:01.638128+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

94 extracted references · 30 canonical work pages · 13 internal anchors

[1]

The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

Chain-of-Thought Reasoning Without Prompting , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=
[2]

The Eleventh International Conference on Learning Representations , year=

Discovering Latent Knowledge in Language Models Without Supervision , author=. The Eleventh International Conference on Learning Representations , year=
[3]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Loren: Logic-regularized reasoning for interpretable fact verification , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=
[4]

A Logic-Driven Framework for Consistency of Neural Models , author=. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , pages=

2019
[5]

Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

Scaling llm test-time compute optimally can be more effective than scaling model parameters , author=. arXiv preprint arXiv:2408.03314 , year=

work page internal anchor Pith review arXiv
[6]

Advances in neural information processing systems , volume=

Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=
[7]

Advances in Neural Information Processing Systems , volume=

Tree of thoughts: Deliberate problem solving with large language models , author=. Advances in Neural Information Processing Systems , volume=
[8]

The Twelfth International Conference on Learning Representations , year=

Let's verify step by step , author=. The Twelfth International Conference on Learning Representations , year=
[9]

Solving math word problems with process- and outcome-based feedback

Solving math word problems with process-and outcome-based feedback , author=. arXiv preprint arXiv:2211.14275 , year=

work page internal anchor Pith review arXiv
[10]

ToolChain*: Efficient Action Space Navigation in Large Language Models with A* Search , author=
[11]

Forty-first International Conference on Machine Learning , year=

Alphazero-like tree-search can guide large language model decoding and training , author=. Forty-first International Conference on Machine Learning , year=
[12]

Training Verifiers to Solve Math Word Problems

Training verifiers to solve math word problems , author=. arXiv preprint arXiv:2110.14168 , year=

work page internal anchor Pith review arXiv
[13]

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021 , pages=

Probabilistic Graph Reasoning for Natural Proof Generation , author=. Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021 , pages=

2021
[14]

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

Learning Logic Rules for Document-Level Relation Extraction , author=. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

2021
[15]

Advances in neural information processing systems , volume=

Probabilistic logic neural networks for reasoning , author=. Advances in neural information processing systems , volume=
[16]

Advances in Neural Information Processing Systems , volume=

Towards interpretable natural language understanding with explanations as latent variables , author=. Advances in Neural Information Processing Systems , volume=
[17]

Computational Linguistics , volume=

Probing classifiers: Promises, shortcomings, and advances , author=. Computational Linguistics , volume=. 2022 , publisher=

2022
[18]

Aho and Jeffrey D

Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

1972
[19]

Publications Manual , year = "1983", publisher =

1983
[20]

Chandra and Dexter C

Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

work page doi:10.1145/322234.322243 1981
[21]

Scalable training of

Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of
[22]

Advances in neural information processing systems , volume=

Training language models to follow instructions with human feedback , author=. Advances in neural information processing systems , volume=
[23]

Limo: Less is more for reasoning

LIMO: Less is More for Reasoning , author=. arXiv preprint arXiv:2502.03387 , year=

work page arXiv
[24]

Logiqa: A challenge dataset for machine reading comprehension with logical reasoning

Logiqa: A challenge dataset for machine reading comprehension with logical reasoning , author=. arXiv preprint arXiv:2007.08124 , year=

work page arXiv 2007
[25]

Dan Gusfield , title =. 1997

1997
[26]

Tetreault , title =

Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

2015
[27]

A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =
[28]

Measuring Mathematical Problem Solving With the MATH Dataset , author=
[29]

HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering

HotpotQA: A dataset for diverse, explainable multi-hop question answering , author=. arXiv preprint arXiv:1809.09600 , year=

work page internal anchor Pith review arXiv
[30]

Chi and Quoc V

Jason Wei and Xuezhi Wang and Dale Schuurmans and Maarten Bosma and Brian Ichter and Fei Xia and Ed H. Chi and Quoc V. Le and Denny Zhou , editor =. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , booktitle =. 2022 , url =

2022
[31]

The Llama 3 Herd of Models

The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

work page internal anchor Pith review arXiv
[32]

2023 , eprint=

Mistral 7B , author=. 2023 , eprint=

2023
[33]

The 4th Workshop on Mathematical Reasoning and AI at NeurIPS'24 , year=

Scaling Inference Computation: Compute-Optimal Inference for Problem-Solving with Language Models , author=. The 4th Workshop on Mathematical Reasoning and AI at NeurIPS'24 , year=
[34]

2023 , eprint=

Large Language Models are Zero-Shot Reasoners , author=. 2023 , eprint=

2023
[35]

2023 , eprint=

Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models , author=. 2023 , eprint=

2023
[36]

The Twelfth International Conference on Learning Representations,

Chao Chen and Kai Liu and Ze Chen and Yi Gu and Yue Wu and Mingyuan Tao and Zhihang Fu and Jieping Ye , title =. The Twelfth International Conference on Learning Representations,. 2024 , url =

2024
[37]

The Eleventh International Conference on Learning Representations,

Collin Burns and Haotian Ye and Dan Klein and Jacob Steinhardt , title =. The Eleventh International Conference on Learning Representations,. 2023 , url =

2023
[38]

The internal state of an LLM knows when it ' s lying

Amos Azaria and Tom M. Mitchell , editor =. The Internal State of an. Findings of the Association for Computational Linguistics:. 2023 , url =. doi:10.18653/V1/2023.FINDINGS-EMNLP.68 , timestamp =

work page doi:10.18653/v1/2023.findings-emnlp.68 2023
[39]

Kenneth Li and Oam Patel and Fernanda B. Vi. Inference-Time Intervention: Eliciting Truthful Answers from a Language Model , booktitle =. 2023 , url =

2023
[40]

Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

Charlie Snell and Jaehoon Lee and Kelvin Xu and Aviral Kumar , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2408.03314 , eprinttype =. 2408.03314 , timestamp =

work page Pith review doi:10.48550/arxiv.2408.03314 2024
[41]

Solving math word problems with process- and outcome-based feedback

Jonathan Uesato and Nate Kushman and Ramana Kumar and H. Francis Song and Noah Y. Siegel and Lisa Wang and Antonia Creswell and Geoffrey Irving and Irina Higgins , title =. CoRR , volume =. 2022 , url =. doi:10.48550/ARXIV.2211.14275 , eprinttype =. 2211.14275 , timestamp =

work page Pith review doi:10.48550/arxiv.2211.14275 2022
[42]

URLhttps://www.nature.com/articles/ s41586-023-06924-6

Bernardino Romera. Mathematical discoveries from program search with large language models , journal =. 2024 , url =. doi:10.1038/S41586-023-06924-6 , timestamp =

work page doi:10.1038/s41586-023-06924-6 2024
[43]

Nature , volume=

Large language models encode clinical knowledge , author=. Nature , volume=. 2023 , publisher=

2023
[44]

Costello and Gordon Pennycook and David G

Thomas H. Costello and Gordon Pennycook and David G. Rand , title =. Science , volume =. 2024 , doi =. https://www.science.org/doi/pdf/10.1126/science.adq1814 , abstract =

work page doi:10.1126/science.adq1814 2024
[46]

Coarse-to-Fine n-Best Parsing and MaxEnt Discriminative Reranking , booktitle =

Eugene Charniak and Mark Johnson , editor =. Coarse-to-Fine n-Best Parsing and MaxEnt Discriminative Reranking , booktitle =. 2005 , url =. doi:10.3115/1219840.1219862 , timestamp =

work page doi:10.3115/1219840.1219862 2005
[47]

GPT-4 Technical Report

OpenAI , title =. CoRR , volume =. 2023 , url =. doi:10.48550/ARXIV.2303.08774 , eprinttype =. 2303.08774 , timestamp =

work page Pith review doi:10.48550/arxiv.2303.08774 2023
[48]

Qwen2 Technical Report

An Yang and Baosong Yang and Binyuan Hui and Bo Zheng and Bowen Yu and Chang Zhou and Chengpeng Li and Chengyuan Li and Dayiheng Liu and Fei Huang and Guanting Dong and Haoran Wei and Huan Lin and Jialong Tang and Jialin Wang and Jian Yang and Jianhong Tu and Jianwei Zhang and Jianxin Ma and Jianxin Yang and Jin Xu and Jingren Zhou and Jinze Bai and Jinzh...

work page internal anchor Pith review doi:10.48550/arxiv.2407.10671 2024
[49]

PaLM 2 Technical Report

Rohan Anil and Andrew M. Dai and Orhan Firat and Melvin Johnson and Dmitry Lepikhin and Alexandre Passos and Siamak Shakeri and Emanuel Taropa and Paige Bailey and Zhifeng Chen and Eric Chu and Jonathan H. Clark and Laurent El Shafey and Yanping Huang and Kathy Meier. PaLM 2 Technical Report , journal =. 2023 , url =. doi:10.48550/ARXIV.2305.10403 , eprin...

work page internal anchor Pith review doi:10.48550/arxiv.2305.10403 2023
[50]

DeepSeek-V3 Technical Report

Deepseek-v3 technical report , author=. arXiv preprint arXiv:2412.19437 , year=

work page internal anchor Pith review arXiv
[51]

2025 , eprint=

Qwen2.5 Technical Report , author=. 2025 , eprint=

2025
[52]

Self-Consistency Improves Chain of Thought Reasoning in Language Models , author=
[53]

Advances in Neural Information Processing Systems , volume=

Solving quantitative reasoning problems with language models , author=. Advances in Neural Information Processing Systems , volume=
[54]

Physics of language models: Part 3.2, knowledge manipula- tion.arXiv:2309.14402, 2023

Zeyuan Allen. Physics of Language Models: Part 3.2, Knowledge Manipulation , journal =. 2023 , url =. doi:10.48550/ARXIV.2309.14402 , eprinttype =. 2309.14402 , timestamp =

work page doi:10.48550/arxiv.2309.14402 2023
[55]

Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them

Challenging big-bench tasks and whether chain-of-thought can solve them , author=. arXiv preprint arXiv:2210.09261 , year=

work page internal anchor Pith review arXiv
[56]

The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track , year=

Mmlu-pro: A more robust and challenging multi-task language understanding benchmark , author=. The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track , year=
[57]

2022 , eprint=

Solving math word problems with process- and outcome-based feedback , author=. 2022 , eprint=

2022
[58]

2024 , eprint=

Self-Rewarding Language Models , author=. 2024 , eprint=

2024
[59]

2024 , eprint=

Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations , author=. 2024 , eprint=

2024
[60]

2024 , eprint=

V-STaR: Training Verifiers for Self-Taught Reasoners , author=. 2024 , eprint=

2024
[61]

2023 , eprint=

Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena , author=. 2023 , eprint=

2023
[62]

2024 , eprint=

ReST-MCTS*: LLM Self-Training via Process Reward Guided Tree Search , author=. 2024 , eprint=

2024
[63]

2022 , url =

Paul Christiano and Ajeya Cotra and Mark Xu , title =. 2022 , url =

2022
[64]

2024 , url =

OpenAI , title =. 2024 , url =

2024
[65]

2024 , eprint=

Eliciting Latent Knowledge from Quirky Language Models , author=. 2024 , eprint=

2024
[66]

Understanding intermediate layers using linear classifier probes

Understanding intermediate layers using linear classifier probes , author=. arXiv preprint arXiv:1610.01644 , year=

work page internal anchor Pith review arXiv
[67]

Qwen2.5 Technical Report

Qwen2. 5 Technical Report , author=. arXiv preprint arXiv:2412.15115 , year=

work page internal anchor Pith review arXiv
[68]

2025 , eprint=

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning , author=. 2025 , eprint=

2025
[69]

Yoshitaka Inoue, Tianci Song, and Tianfan Fu

O1 Replication Journey--Part 2: Surpassing O1-preview through Simple Distillation, Big Progress or Bitter Lesson? , author=. arXiv preprint arXiv:2411.16489 , year=

work page arXiv
[70]

The Thirteenth International Conference on Learning Representations , year=

Physics of language models: Part 2.1, grade-school math and the hidden reasoning process , author=. The Thirteenth International Conference on Learning Representations , year=
[71]

International Conference on Learning Representations , year=

Decoupled Weight Decay Regularization , author=. International Conference on Learning Representations , year=
[72]

InFindings of the Association for Computational Linguistics: ACL 2025, pages 18974–18988, Vienna, Austria

Teaching large language models to reason with reinforcement learning , author=. arXiv preprint arXiv:2403.04642 , year=

work page arXiv
[73]

Improving large language model fine-tuning for solving math problems,

Improving large language model fine-tuning for solving math problems , author=. arXiv preprint arXiv:2310.10047 , year=

work page arXiv
[74]

Large language models are reasoning teachers,

Large language models are reasoning teachers , author=. arXiv preprint arXiv:2212.10071 , year=

work page arXiv
[75]

Scaling Laws for Neural Language Models

Scaling laws for neural language models , author=. arXiv preprint arXiv:2001.08361 , year=

work page internal anchor Pith review arXiv 2001
[76]

Advances in neural information processing systems , volume=

Language models are few-shot learners , author=. Advances in neural information processing systems , volume=
[77]

Advances in neural information processing systems , volume=

Large language models are zero-shot reasoners , author=. Advances in neural information processing systems , volume=
[78]

doi: 10.48550/arXiv.2310.01714

Large language models as analogical reasoners , author=. arXiv preprint arXiv:2310.01714 , year=

work page arXiv
[79]

2024 , eprint=

Top- n : Not All Logits Are You Need , author=. 2024 , eprint=

2024
[80]

Proceedings of the 29th ACM International Conference on Information & Knowledge Management , pages=

Neural logic reasoning , author=. Proceedings of the 29th ACM International Conference on Information & Knowledge Management , pages=
[81]

Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining , pages=

Graph collaborative reasoning , author=. Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining , pages=

Showing first 80 references.