Pith. sign in

REVIEW 3 major objections 4 minor 85 references

Reproducing and Stress-Testing Two Approaches to LLM Reasoning Reliability: Test-Time Probability Aggregation and Logic-Representation Editing

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read Independent rerun: aggregation reproduces, logic steering does not

desk verdict An honest RPC reproduction plus a plausible but unverified LCF negative result; the asymmetry is directional, not settled. read the letter →

arxiv 2608.08514 v1 pith:5ZW7Q2YL submitted 2026-08-09 cs.AI cs.CLcs.LG

classification cs.AIcs.CLcs.LG
keywords LLMreasoningreliabilityreproducibilitystudytest-timeaggregationself-consistencylogicrepresentationeditingprobingactivationsteeringcalibration
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 independently re-runs two published methods that claim to make large language model reasoning more reliable: a test-time aggregation rule that combines token probabilities with self-consistency votes (RPC), and a representation-editing pipeline that trains projectors to separate 'content' from 'logic' in hidden states and shifts the logic part toward valid conclusions (LCF). The aim is to test whether results reported by each method's own authors survive a faithful re-implementation and transfer across domains and models. It finds an asymmetry: RPC reproduces the original math grid exactly and never significantly hurts accuracy on new domains, but its edge over plain self-consistency never reaches significance; LCF's logic-validity direction is real but weak, is not controllable by the tested interventions, and significantly lowers the probability margin on two of four models. A reader choosing between the two mechanisms would trust the output-side aggregation more, at least at the 7-8B model sizes and small evaluation budgets tested.

What carries the argument

The argument is carried by two concrete mechanisms at opposite ends of the model. RPC is an aggregation rule over K sampled chain-of-thought paths: it weights each path by its mean token probability, combines the weighted votes with self-consistency, and prunes a low-probability component using a Weibull mixture. LCF trains projectors of shape $d \to 2048 \to 1024$ that split a hidden state into content and logic vectors, a cross-attention decoder that reconstructs the state, and an InfoNCE contrastive objective that pulls logic vectors of valid conclusions together; at inference the logic vector is shifted toward the 'valid' region with a scale $\eta$ and decoded back into the residual stream. The paper also uses a held-out linear probe to measure how strongly logic validity and a semantic control attribute are linearly encoded; the probe's 0.82-versus-0.95 contrast is the evidence that the logic signal is decodable but weak, and the additive-intervention sweep is the evidence that it is not controllable.

What would settle it

Run the original LCF code, or a version whose projector, contrastive objective, edited layers, and validity discriminator are supplied by the original authors, on Llama-2 and Mistral with the same 204-item fallacy task: if DeltaProb rises significantly on either model instead of falling, the paper's central negative claim is false. Likewise, a further independent dataset draw on BIRD at K=32, which has already turned a +2.5 lead into -0.25 when the sample grew from n=80 to n=200, would settle whether any aggregation edge exists.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that the two reliability mechanisms do not have the same status under independent testing. RPC's published pipeline, run on the original authors' reasoning paths, reproduces the reported grid nearly exactly (accuracy/ECE 26.15/12.32 versus the paper's 26.11/12.37, with the perplexity and self-consistency baselines also matching), and on four new domains plus three public benchmarks its difference from self-consistency is never significant (paired $p \ge 0.28$), with one K=32 lead of +2.5 accuracy points on BIRD reversing to -0.25 when the sample is enlarged to n=200. LCF's premise is partially confirmed: a held-out linear probe separates valid from invalid conclusions at 0.82 accuracy at the best sub-layer, versus 0.95 for a semantic control, showing a real but weak and localized logic-validity direction. But shifting along that direction—whether by the trained projector, a supervised best-layer direction, or a random same-norm direction—does not move behavior, the DeltaProb gain on Qwen3 is not significant ($p=0.56$), and the same recipe significantly reduces DeltaProb on Llama-2 and Mistral. The paper concludes that test-time aggregation is the safer intervention, while logic-representation editing is not a reliably reproducible, model-agnostic reliability fix at this scale.

Load-bearing premise

The paper's negative verdict on LCF assumes that its reimplementation matches the original unpublished recipe—the projector size, training objective, layer range, intervention strength, and validity judge—so if any of those details differ, the failures could be artifacts of the reimplementation rather than properties of the method.

Editorial extensions

If this is right

  • If the asymmetry holds, users can apply RPC-style aggregation to 7-8B models without fear of significant accuracy loss, but should not expect a large gain over self-consistency on new domains.
  • LCF-style logic editing should not be treated as a model-agnostic reliability fix: at this scale it can significantly reduce the model's confidence in valid options, even where a decodable logic direction exists.
  • A reported gain that rests on an unreleased validity discriminator cannot be independently checked; any claim of near-perfect control needs the discriminator or an auditable substitute before it can be reproduced.
  • Separability of a concept direction does not imply controllability: a direction can be linearly decodable yet adding or subtracting it changes behavior no more than a random direction does.
  • The BIRD budget sweep shows that apparent gains at small sample sizes can vanish with more data, so aggregation advantages should be reported with power bounds.

Reading between the lines

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

  • A step the paper leaves implicit: the benefit of RPC-style aggregation appears only when path probabilities carry information about correctness, so one could measure the confidence-correctness correlation on a task before adopting aggregation.
  • Because the valid-conclusion labels in the LCF study were machine-generated and only 28% were accepted by an independent judge, the 0.82 separability may overestimate the logic signal; human-validated labels could yield a weaker or stronger estimate.
  • At 7-8B scale, model capability appears to dominate intervention effects: the only model helped was the one that already had the target competence, suggesting that logic-steering research should either move to stronger models or focus on unlocking existing capability.
  • If the original LCF recipe, including its discriminator, were released and restored controllability, the asymmetry would narrow; testing that is a direct way to decide whether the negative result is about LCF itself or about reproducibility.
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.

Referee Report

3 major / 4 minor

Summary. The paper reports an independent reproduction and stress-test of two LLM reliability methods: RPC, a test-time probability aggregation method, and LCF, a logic-representation editing method. RPC is reproduced using the authors' released code and reasoning paths, matching the published accuracy/ECE grid, and is then extended to four new task domains plus three public benchmarks; the edge over self-consistency is never significant, and a BIRD budget sweep shows a small, non-significant gain at K=32 that reverses to −0.25 at n=200. LCF, for which no public code exists, is re-implemented from scratch; the paper finds a weak probe-separable validity signal (0.82 at the best sub-layer), no controllability under additive or projective interventions, a non-significant positive DeltaProb effect on one model, and significant DeltaProb drops on two of four models. A training-free contrastive steering variant also fails except under a narrow matched-distribution condition. The conclusion is an asymmetry: RPC is safe and reproducible, whereas LCF's logic editing is a real but weak, locally encoded signal that is decodable but not controllable, and not model-agnostic. The limitations section candidly acknowledges small samples, model-generated labels with 28% judge agreement, and uncertainty about LCF reimplementation fidelity.

Significance. If upheld, the paper provides a valuable independent check on two influential methods: it shows that RPC's aggregate behavior transfers without measurable harm across domains, while LCF's headline effects do not reproduce as a model-agnostic intervention at 7–8B scale. The RPC side is a faithful reproduction using the authors' own code and data, with machine-checkable numbers that match the published grid; the released code, data, and experiment logs are a concrete strength. The LCF side is honest in its limitations, explicitly flagging the unverified reimplementation and the noisy validity labels. The main value is therefore the asymmetry claim: aggregation is the safer, more trustworthy intervention. However, the LCF-negative half of that asymmetry depends on an unverified from-scratch reimplementation and on labels of uncertain validity, which limits how strongly the conclusion can be drawn. The paper is most useful as a careful, reproducible stress-test of RPC and as a cautionary, clearly hedged negative result for LCF.

major comments (3)
  1. [Section 5.2 and Limitations] The LCF negative results are load-bearing for the central asymmetry, but they rest entirely on a reimplementation whose fidelity is unverified. The manuscript infers projector dimensions (d→2048→1024), InfoNCE temperature (tau=0.1), the edited layer range, and the intervention scale eta from the paper text, and substitutes an unreleased validity discriminator with a GPT-4 judge and DistilBERT classifier. The Limitations concede that the implementation 'may differ in unstated details,' but the conclusion's asymmetry ('logic-representation editing rests on a real but weak, locally-encoded signal... decodable yet not controllable') is stated without this caveat at full strength. Every LCF number in Table 5, Figure 3, and Table 9 comes from this unverified pipeline. Because the RPC side of the comparison is an exact reproduction using the authors' released code, the two sides are not compared on equal footing. The paper should either add a hyperparameter sensitivity analysis around eta, layer range, and projector architecture, or clearly reframe the LCF conclusions as 'under our faithful re-implementation, not as a property of the original method.' As written, the asymmetry claim outruns the evidence.
  2. [Section 5.2, Figure 3, and Limitations] The probe claim that a 'logic-validity direction is real but weak' is confounded by the construction of the valid/invalid labels. Valid conclusions are generated by GPT-4o-mini while invalid ones come from LFUD's fallacious conclusions; an independent judge accepts only 28% of the generated 'valid' conclusions as strictly valid. The probe could therefore be separating surface style (e.g., formality, wording patterns) rather than logical validity. The Limitations say this 'weakens, not strengthens, our reading of the signal,' but that argument only works if the style confound is orthogonal to logic; nothing in the paper establishes that. A concrete control would be to probe on a style-matched invalid set (e.g., generated invalid conclusions) or to compare against a bag-of-words or embedding-distance baseline. Without such a control, the 0.82 separability does not establish that the signal is 'logic' at all, which is a central premise for the conclusion that the signal is 'real but weak.'
  3. [Section 6, Table 4, and Limitations] The conclusion that RPC 'never significantly degraded accuracy' and is therefore 'the safer intervention' overstates what the statistics support. The power analysis in the Limitations shows that at K=32 on BIRD (n=80) the study has 80% power only for gaps of at least 5.0 accuracy points, and the observed +2.5 gap reverses to −0.25 at n=200. A non-significant difference at these sample sizes does not establish safety; it only bounds the effect size. The phrase 'never significantly degraded' is technically true but the prose in the Conclusion ('aggregation is the safer intervention') reads as a stronger claim. The asymmetry should be restated as 'no significant evidence of harm at these evaluation sizes, and the only observed positive gain disappears with more data,' which is still a meaningful result but less categorical than the current wording.
minor comments (4)
  1. [Abstract and Section 6] The abstract reports the BIRD enlargement reversal as '−0.25' in text and elsewhere as '−0.25' but the full text, including the Conclusion, uses '−0.25' consistently; however, in Table 4 the caption refers to a '+2.5 lead' and the Limitations say '−0.25', which is fine, but the abstract's phrase 'its edge over self-consistency is never significant' is contradicted in the same sentence by the BIRD K=32 p=0.16; clarify that the edge is non-significant rather than absent.
  2. [Table 4] The K=32 row is a single deterministic run with no seed variance, whereas K=8 and K=16 report 95% CIs over 20 seeds; the caption does explain this, but it would be clearer to mark the K=32 row explicitly as 'no CI' in the table itself to avoid reader confusion.
  3. [Section 5.2 and Limitations] The phrase 'we find no evidence of fabrication; we find that the headline leans on an unauditable component' is a useful clarification, but it appears only in the results section; the same nuance would be worth one sentence in the Limitations where the discriminator substitution is discussed.
  4. [Section 3 and Table 1] The LFUD split is described as '45:5:17' for conclusion generation and '540/204' for RPC MCQ; the relationship between these two splits is not explained in the main text. Please clarify whether the 204-item LCF evaluation is a subset of the 540-item MCQ or a separate split.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the RPC reproduction is explicit code reuse, and the LCF negative results rest on conceded reimplementation and label-noise limitations, not on a conclusion smuggled into the inputs.

full rationale

The paper is a reproducibility study, and its central asymmetry is an empirical comparison rather than a derivation. RPC's exact grid match comes from reusing the authors' released aggregation and evaluation code over their published reasoning paths (Section 4: 'For RPC we reuse the authors' released aggregation and evaluation code over their published reasoning paths'), so the match is a code-reproducibility check, not a first-principles prediction; the authors disclose this reuse explicitly. The new-domain RPC results are anchored to public benchmarks (BIRD, GSM8K, FOLIO, LogiQA) and to self-consistency baselines with paired significance tests, so no fitted parameter is renamed as a prediction. LCF is re-implemented from scratch because no code was released, and the paper repeatedly concedes the consequences: the validity labels are GPT-4o-mini-generated ('valid conclusions are model-generated (GPT-4o-mini) while invalid ones are the dataset's, so the 0.82 separability may partly reflect surface style rather than logic'), an independent judge accepts only 28% of generated valid conclusions, and 'our conclusions about LCF are negative results under a faithful re-implementation, not proof that no implementation can reproduce the paper.' These concessions work against the paper's own LCF-negative reading rather than bootstrapping it. The only self-citation of note is the KCC dataset (Cho et al., 2026), a released SIGIR-published corpus used as one of several evaluation domains; removing KCC would not change the RPC/LCF asymmetry, and KCC is not invoked to exclude alternatives or to justify a uniqueness claim. I found no step in which a predicted quantity reduces by construction to a fitted input, no load-bearing self-citation chain, no imported uniqueness theorem, and no ansatz smuggled in via the authors' own prior work.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claims rest on reusing RPC authors' artifacts, on a from-scratch LCF reimplementation without released code, on generated validity labels, and on small evaluation sets. None of these are independently machine-checked. The paper's own Limitations section concedes label noise, power limits, and unknown reimplementation fidelity. Free parameters are inherited hyperparameters and post hoc layer selection rather than newly fitted physical constants. No invented entities are introduced.

free parameters (5)
  • LCF intervention scale eta = not stated in text; adopted from original paper
    Shift magnitude for the logic-vector edit; if mis-set relative to the original, LCF ineffectiveness in this reproduction could be an artifact.
  • LCF edited layer range = not stated in text; single sub-layer or layer range inferred from original
    The probe signal is 0.82 at the best sub-layer but chance (0.52) pooled over edited layers; the intervention outcome depends on which layers are edited.
  • LCF projector dimensions and InfoNCE temperature = d -> 2048 -> 1024, tau = 0.1
    Reimplementation hyperparameters taken from Wu et al. 2025; central LCF results depend on them and they are not independently justified here.
  • Probe layer selection = single best sub-layer (held-out probe 0.82)
    The 0.82 separability is the maximum over layers; selecting the best layer after probing can inflate the reported estimate.
  • RPC Weibull mixture parameters = from authors' released code
    RPC's pruning component is part of the released aggregation; the exact reproduction claim is conditional on these parameters and was not independently re-estimated.
assumptions (6)
  • domain assumption RPC authors' released reasoning paths and aggregation code faithfully implement the published RPC method.
    The exact-reproduction claim in Section 5.1 rests on reusing these artifacts; no independent reimplementation of RPC is attempted.
  • domain assumption GPT-4o-mini generated valid conclusions and the GPT-4 and Claude validity judges approximate ground-truth logical validity.
    LCF training labels and the 0.82 probe rely on this; the paper reports only 28% judge agreement on generated valid conclusions in the Limitations section.
  • domain assumption Linear probe accuracy on hidden states is a valid measure of how strongly an attribute is encoded.
    Representation-level claims in Figure 3 assume linear separability reflects usable encoding, a common but nontrivial assumption.
  • domain assumption Self-consistency majority vote is the appropriate baseline for RPC.
    RPC's edge is measured against SC, the original method's own baseline choice; if another baseline were used, the relative claim could differ.
  • domain assumption MoodRisk suicide-risk labels and representations serve as a valid semantic-attribute control.
    The 0.95 control separability comes from a separate corpus with different labels and pooling; its comparability to logic validity is assumed.
  • standard math Paired Wilcoxon tests on these small, non-random samples are valid for the reported significance statements.
    Ties and small n make Wilcoxon conservative, but the tests do not correct for multiple comparisons and power is low, as the paper's own power analysis concedes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reproducing and Stress-Testing Two Approaches to LLM Reasoning Reliability: Test-Time Probability Aggregation and Logic-Representation Editing." pith.science (2026). https://pith.science/paper/5ZW7Q2YL

@misc{pith2026260808514,
  author       = {Pith},
  title        = {Pith review of: Reproducing and Stress-Testing Two Approaches to LLM Reasoning Reliability: Test-Time Probability Aggregation and Logic-Representation Editing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5ZW7Q2YL}},
  note         = {Machine review of arXiv:2608.08514}
}
abstract

We independently reproduce two recent methods for making large language model (LLM) reasoning more reliable, and stress-test them across domains and models (RPC across four new task domains with Qwen3-8B, LCF across four 7-8B models). The first, RPC, aggregates token probabilities and self-consistency at inference; the second, LCF, trains projectors that split hidden states into "content" and "logic" and edits the logic part toward a valid region. Validating such reliability claims matters because the original evaluations are run by each method's own authors and were never independently reproduced or stress-tested across models and domains, and LCF shipped no public code. We re-run RPC's published-path aggregation and re-implement LCF's projector, contrastive, and intervention pipeline, then extend both to text-to-SQL, legal extraction, fallacy identification, and precedent grading, and probe LCF's representation directly. RPC reproduces the original grid exactly on the authors' released reasoning paths; on four new domains its edge over self-consistency is never significant (ties or small mixed differences, paired p >= 0.28), and on BIRD, the one domain where we vary the budget, the edge grows with K as predicted but its largest gap (+2.5 accuracy at K=32, p=0.16) reverses to -0.25 when we enlarge the sample to n=200. LCF's logic-validity direction is real but weak (0.82 separability at the single best sub-layer versus 0.95 for a semantic-attribute control); its one positive effect (Qwen3 $\Delta$Prob) is not significant (p=0.56), while it significantly reduces $\Delta$Prob on two of the other three models.

Figures

Figures reproduced from arXiv: 2608.08514 by the authors.

Figure 1
Figure 1. The two reliability methods diverge under faithful re-implementation. (a) On BIRD, RPC’s accuracy moves above SC (majority-vote baseline) only at K=32; error bars are 95% CIs over 20 seeds for K=8, 16 (K=32 is a single deterministic run), and the +2.5 gap is not significant (p=0.16). (b) ∆Prob (the probability margin the model puts on the valid option) under LCF: it rises on Qwen3 (gap +1.7, n.s.) but significantly … view at source ↗
Figure 2
Figure 2. Where the two methods act. Both consume the same sampled chain-of-thought but inter￾vene at opposite ends of the model. LCF (green) works inside the network: at a single layer ℓ it splits the hidden state into a content and a logic vector, shifts the logic vector toward the valid region, and decodes it back into the residual stream. RPC (blue) leaves the weights and activations untouched and instead aggregates the o… view at source ↗
Figure 3
Figure 3. The logic-validity direction is real but [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

85 extracted references · 44 canonical work pages

  1. [1]

    A theoretical study on bridging internal probability and self-consistency for

    Zhou, Zhi and Tan, Yuhao and Li, Zenan and Yao, Yuan and Guo, Lan-Zhe and Li, Yu-Feng and Ma, Xiaoxing , booktitle=. A theoretical study on bridging internal probability and self-consistency for

  2. [2]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Content-free logical modification of large language model by disentangling and modifying logic representation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  3. [3]

    International Conference on Learning Representations , year=

    Self-consistency improves chain of thought reasoning in language models , author=. International Conference on Learning Representations , year=

  4. [4]

    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=

  5. [5]

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

    Steering llama 2 via contrastive activation addition , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  6. [7]

    Advances in Neural Information Processing Systems , volume=

    Can llm already serve as a database interface? a big bench for large-scale database grounded text-to-sqls , author=. Advances in Neural Information Processing Systems , volume=

  7. [8]

    International conference on learning representations , volume=

    Programming refusal with conditional activation steering , author=. International conference on learning representations , volume=

  8. [9]

    Advances in neural information processing systems , volume=

    Inference-time intervention: Eliciting truthful answers from a language model , author=. Advances in neural information processing systems , volume=

Show all 85 references
  1. [10]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Mitigating content effects on reasoning in language models through fine-grained activation steering , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  2. [11]

    Findings of the Association for Computational Linguistics: NAACL 2024 , pages=

    Reason from fallacy: Enhancing large language models’ logical reasoning through logical fallacy understanding , author=. Findings of the Association for Computational Linguistics: NAACL 2024 , pages=

  3. [12]

    Proceedings of the 2024 conference of the north American chapter of the association for computational linguistics: human language technologies (volume 1: long papers) , pages=

    A survey of confidence estimation and calibration in large language models , author=. Proceedings of the 2024 conference of the north American chapter of the association for computational linguistics: human language technologies (volume 1: long papers) , pages=

  4. [14]

    Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

    KCC: Korean Civil Case Dataset for Legal Information Retrieval , author=. Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

  5. [15]

    International Conference on Learning Representations , year=

    Scaling llm test-time compute optimally can be more effective than scaling model parameters , author=. International Conference on Learning Representations , year=

  6. [16]

    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=

  7. [20]

    International Conference on Machine Learning , pages=

    Position: Trustllm: Trustworthiness in large language models , author=. International Conference on Machine Learning , pages=. 2024 , organization=

  8. [21]

    ACM transactions on information systems , volume=

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions , author=. ACM transactions on information systems , volume=. 2025 , publisher=

  9. [22]

    Editing models with task arithmetic , booktitle =

    Gabriel Ilharco and Marco T. Editing models with task arithmetic , booktitle =. 2023 , url =

  10. [24]

    Albert Q. Jiang and Alexandre Sablayrolles and Arthur Mensch and Chris Bamford and Devendra Singh Chaplot and Diego de Las Casas and Florian Bressand and Gianna Lengyel and Guillaume Lample and Lucile Saulnier and L. Mistral 7B , journal =. 2023 , eprint =

  11. [25]

    International conference on machine learning , pages=

    On calibration of modern neural networks , author=. International conference on machine learning , pages=. 2017 , organization=

  12. [26]

    Advances in neural information processing systems , volume=

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

  13. [27]

    International Conference on Learning Representations , volume=

    Let's verify step by step , author=. International Conference on Learning Representations , volume=

  14. [28]

    See https://vicuna

    Vicuna: An open-source chatbot impressing gpt-4 with 90\ author=. See https://vicuna. lmsys. org (accessed 14 April 2023) , volume=

  15. [29]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Folio: Natural language reasoning with first-order logic , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

  16. [30]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

  17. [31]

    International Conference on Learning Representations , volume=

    Sparse autoencoders find highly interpretable features in language models , author=. International Conference on Learning Representations , volume=

  18. [32]

    Advances in Neural Information Processing Systems , volume=

    Self-evaluation guided beam search for reasoning , author=. Advances in Neural Information Processing Systems , volume=

  19. [33]

    Decodingtrust: A comprehensive assessment of trustworthiness in

    Wang, Boxin and Chen, Weixin and Pei, Hengzhi and Xie, Chulin and Kang, Mintong and Zhang, Chenhui and Xu, Chejian and Xiong, Zidi and Dutta, Ritik and Schaeffer, Rylan and others , booktitle=. Decodingtrust: A comprehensive assessment of trustworthiness in

  20. [34]

    Measuring mathematical problem solving with the

    Hendrycks, Dan and Burns, Collin and Kadavath, Saurav and Arora, Akul and Basart, Steven and Tang, Eric and Song, Dawn and Steinhardt, Jacob , booktitle=. Measuring mathematical problem solving with the

  21. [35]

    Scientific data , volume=

    Mathodyssey: Benchmarking mathematical problem-solving skills in large language models using odyssey math data , author=. Scientific data , volume=. 2025 , publisher=

  22. [36]

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

    He, Chaoqun and Luo, Renjie and Bai, Yuzhuo and Hu, Shengding and Thai, Zhen and Shen, Junhao and Hu, Jinyi and Han, Xu and Huang, Yujie and Zhang, Yuxiang and Liu, Jie and Qi, Lei and Liu, Zhiyuan and Sun, Maosong , title =. Proceedings of the 62nd Annual Meeting of the Assoc...

  23. [39]

    Bowman and Kyunghyun Cho , title =

    Angelica Chen and Jason Phang and Alicia Parrish and Vishakh Padmakumar and Chen Zhao and Samuel R. Bowman and Kyunghyun Cho , title =. Transactions on Machine Learning Research , year =

  24. [40]

    Nature , volume =

    Farquhar, Sebastian and Kossen, Jannik and Kuhn, Lorenz and Gal, Yarin , title =. Nature , volume =. 2024 , doi =

  25. [41]

    and Dasgupta, Ishita and Chan, Stephanie C

    Lampinen, Andrew K. and Dasgupta, Ishita and Chan, Stephanie C. Y. and Sheahan, Hannah R. and Creswell, Antonia and Kumaran, Dharshan and McClelland, James L. and Hill, Felix , title =. PNAS Nexus , volume =. 2024 , doi =

  26. [42]

    A Looming Replication Crisis in Evaluating Behavior in Language Models? Evidence and Solutions , journal =

    Laur. A Looming Replication Crisis in Evaluating Behavior in Language Models? Evidence and Solutions , journal =. 2024 , eprint =

  27. [43]

    Reasoning Circuits in Language Models: A Mechanistic Interpretation of Syllogistic Inference , booktitle =

    Geonhee Kim and Marco Valentino and Andr. Reasoning Circuits in Language Models: A Mechanistic Interpretation of Syllogistic Inference , booktitle =. 2025 , eprint =

  28. [44]

    CoRR , volume =

    Hojin Kim and Jaehyung Kim , title =. CoRR , volume =. 2026 , eprint =

  29. [45]

    CoRR , volume =

    Ming Liu , title =. CoRR , volume =. 2026 , eprint =

  30. [47]

    2025 , eprint =

    Qwen3 Technical Report , journal =. 2025 , eprint =

  31. [48]

    Leonardo Bertolazzi, Sandro Pezzelle, and Raffaella Bernardi. 2026. https://arxiv.org/abs/2510.06700 How language models conflate logical validity with plausibility: A representational analysis of content effects . In Findings of the Association for Computational Linguistics: ACL 2026

  32. [49]

    Bowman, and Kyunghyun Cho

    Angelica Chen, Jason Phang, Alicia Parrish, Vishakh Padmakumar, Chen Zhao, Samuel R. Bowman, and Kyunghyun Cho. 2024. https://openreview.net/forum?id=5nBqY1y96B Two failures of self-consistency in the multi-step reasoning of LLMs . Transactions on Machine Learning Research

  33. [50]

    Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, and 1 others. 2023. Vicuna: An open-source chatbot impressing gpt-4 with 90\ quality. See https://vicuna. lmsys. org (accessed 14 April ...

  34. [51]

    Minhan Cho, Soyoung Park, S Shyam Sundar, Daejin Choi, and Jinyoung Han. 2026. Kcc: Korean civil case dataset for legal information retrieval. In Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 3117--3123

  35. [52]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. https://arxiv.org/abs/2110.14168 Training verifiers to solve math word problems . ...

  36. [53]

    Ishita Dasgupta, Andrew K Lampinen, Stephanie CY Chan, Hannah R Sheahan, Antonia Creswell, Dharshan Kumaran, James L McClelland, and Felix Hill. 2022. Language models show human-like content effects on reasoning tasks. arXiv preprint arXiv:2207.07051

  37. [54]

    Meng Fang, Xiangpeng Wan, Fei Lu, Fei Xing, and Kai Zou. 2025. Mathodyssey: Benchmarking mathematical problem-solving skills in large language models using odyssey math data. Scientific data, 12(1):1392

  38. [55]

    Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. 2024. https://doi.org/10.1038/s41586-024-07421-0 Detecting hallucinations in large language models using semantic entropy . Nature, 630:625--630

  39. [56]

    Jiahui Geng, Fengyu Cai, Yuxia Wang, Heinz Koeppl, Preslav Nakov, and Iryna Gurevych. 2024. A survey of confidence estimation and calibration in large language models. In Proceedings of the 2024 conference of the north American chapter of the association for computational ling...

  40. [57]

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In International conference on machine learning, pages 1321--1330. PMLR

  41. [58]

    Simeng Han, Hailey Schoelkopf, Yilun Zhao, Zhenting Qi, Martin Riddell, Wenfei Zhou, James Coady, David Peng, Yujie Qiao, Luke Benson, and 1 others. 2024. Folio: Natural language reasoning with first-order logic. In Proceedings of the 2024 Conference on Empirical Methods in Na...

  42. [59]

    Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Thai, Junhao Shen, Jinyi Hu, Xu Han, Yujie Huang, Yuxiang Zhang, Jie Liu, Lei Qi, Zhiyuan Liu, and Maosong Sun. 2024. https://doi.org/10.18653/v1/2024.acl-long.211 Olympiadbench: A challenging benchmark for promoting agi w...

  43. [60]

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the MATH dataset. In Advances in Neural Information Processing Systems Datasets and Benchmarks Track

  44. [61]

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and 1 others. 2025. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM transactions ...

  45. [62]

    Yue Huang, Lichao Sun, Haoran Wang, Siyuan Wu, Qihui Zhang, Yuan Li, Chujie Gao, Yixin Huang, Wenhan Lyu, Yixuan Zhang, and 1 others. 2024. Position: Trustllm: Trustworthiness in large language models. In International Conference on Machine Learning, pages 20166--20270. PMLR

  46. [63]

    Robert Huben, Hoagy Cunningham, Logan Smith, Aidan Ewart, and Lee Sharkey. 2024. Sparse autoencoders find highly interpretable features in language models. In International Conference on Learning Representations, volume 2024, pages 7827--7845

  47. [64]

    Gabriel Ilharco, Marco T \' u lio Ribeiro, Mitchell Wortsman, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. 2023. https://openreview.net/forum?id=6t0Kwf8-jrj Editing models with task arithmetic . In The Eleventh International Conference on Learning Representations, ICL...

  48. [65]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, L \' e lio Renard Lavaud, Marie - Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, ...

  49. [66]

    Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, and 1 others. 2022. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221

  50. [67]

    Geonhee Kim, Marco Valentino, and Andr \' e Freitas. 2025. https://arxiv.org/abs/2408.08590 Reasoning circuits in language models: A mechanistic interpretation of syllogistic inference . In Findings of the Association for Computational Linguistics: ACL 2025

  51. [68]

    Hojin Kim and Jaehyung Kim. 2026. https://doi.org/10.48550/ARXIV.2601.13735 Reasoning or fluency? dissecting probabilistic confidence in best-of-n selection . CoRR, abs/2601.13735

  52. [69]

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199--22213

  53. [70]

    Lampinen, Ishita Dasgupta, Stephanie C

    Andrew K. Lampinen, Ishita Dasgupta, Stephanie C. Y. Chan, Hannah R. Sheahan, Antonia Creswell, Dharshan Kumaran, James L. McClelland, and Felix Hill. 2024. https://doi.org/10.1093/pnasnexus/pgae233 Language models, like humans, show content effects on reasoning tasks . PNAS N...

  54. [71]

    Bruce W Lee, Inkit Padhi, Karthikeyan Natesan Ramamurthy, Erik Miehling, Pierre Dognin, Manish Nagireddy, and Amit Dhurandhar. 2025. Programming refusal with conditional activation steering. In International conference on learning representations, volume 2025, pages 90960--90985

  55. [72]

    Jinyang Li, Binyuan Hui, Ge Qu, Jiaxi Yang, Binhua Li, Bowen Li, Bailin Wang, Bowen Qin, Ruiying Geng, Nan Huo, and 1 others. 2023 a . Can llm already serve as a database interface? a big bench for large-scale database grounded text-to-sqls. Advances in Neural Information Proc...

  56. [73]

    Kenneth Li, Oam Patel, Fernanda Vi \'e gas, Hanspeter Pfister, and Martin Wattenberg. 2023 b . Inference-time intervention: Eliciting truthful answers from a language model. Advances in neural information processing systems, 36:41451--41530

  57. [74]

    Yanda Li, Dixuan Wang, Jiaqing Liang, Guochao Jiang, Qianyu He, Yanghua Xiao, and Deqing Yang. 2024. Reason from fallacy: Enhancing large language models’ logical reasoning through logical fallacy understanding. In Findings of the Association for Computational Linguistics: NAA...

  58. [75]

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2024. Let's verify step by step. In International Conference on Learning Representations, volume 2024, pages 39578--39601

  59. [76]

    Jian Liu, Leyang Cui, Hanmeng Liu, Dandan Huang, Yile Wang, and Yue Zhang. 2020. https://doi.org/10.24963/IJCAI.2020/501 Logiqa: A challenge dataset for machine reading comprehension with logical reasoning . In Proceedings of the Twenty-Ninth International Joint Conference on ...

  60. [77]

    Ming Liu. 2026. https://doi.org/10.48550/ARXIV.2605.05715 Decodable but not corrected by fixed residual-stream linear steering: Evidence from medical LLM failure regimes . CoRR, abs/2605.05715

  61. [78]

    Qwen Team . 2025. https://doi.org/10.48550/ARXIV.2505.09388 Qwen3 technical report . CoRR, abs/2505.09388

  62. [79]

    Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. 2024. Steering llama 2 via contrastive activation addition. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15504--15522

  63. [80]

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2025. Scaling llm test-time compute optimally can be more effective than scaling model parameters. In International Conference on Learning Representations

  64. [81]

    Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher D Manning. 2023. Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback. In Proceed...

  65. [82]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, and 1 others. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

  66. [83]

    Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. 2023. Steering language models with activation engineering. arXiv preprint arXiv:2308.10248

  67. [84]

    Marco Valentino, Geonhee Kim, Dhairya Dalal, Zhixue Zhao, and Andr \'e Freitas. 2026. Mitigating content effects on reasoning in language models through fine-grained activation steering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 33314--33322

  68. [85]

    Laur \` e ne Vaugrante, Mathias Niepert, and Thilo Hagendorff. 2024. https://doi.org/10.48550/ARXIV.2409.20303 A looming replication crisis in evaluating behavior in language models? evidence and solutions . CoRR, abs/2409.20303

  69. [86]

    Boxin Wang, Weixin Chen, Hengzhi Pei, Chulin Xie, Mintong Kang, Chenhui Zhang, Chejian Xu, Zidi Xiong, Ritik Dutta, Rylan Schaeffer, and 1 others. 2023 a . Decodingtrust: A comprehensive assessment of trustworthiness in GPT models. In Advances in Neural Information Processing ...

  70. [87]

    Cunxiang Wang, Xiaoze Liu, Yuanhao Yue, Xiangru Tang, Tianhang Zhang, Cheng Jiayang, Yunzhi Yao, Wenyang Gao, Xuming Hu, Zehan Qi, and 1 others. 2023 b . Survey on factuality in large language models: Knowledge, retrieval and domain-specificity. arXiv preprint arXiv:2310.07521

  71. [88]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023 c . Self-consistency improves chain of thought reasoning in language models. In International Conference on Learning Representations

  72. [89]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824--24837

  73. [90]

    Xin Wu, Yuqi Bu, Yifei Chen, and Yi Cai. 2025. Content-free logical modification of large language model by disentangling and modifying logic representation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 25507--25515

  74. [91]

    Yuxi Xie, Kenji Kawaguchi, Yiran Zhao, James Xu Zhao, Min-Yen Kan, Junxian He, and Michael Xie. 2023. Self-evaluation guided beam search for reasoning. Advances in Neural Information Processing Systems, 36:41618--41650

  75. [92]

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

  76. [93]

    Zhi Zhou, Yuhao Tan, Zenan Li, Yuan Yao, Lan-Zhe Guo, Yu-Feng Li, and Xiaoxing Ma. 2025. A theoretical study on bridging internal probability and self-consistency for LLM reasoning. In Advances in Neural Information Processing Systems, volume 38, pages 87380--87413

  77. [94]

    Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, and 1 others. 2023. Representation engineering: A top-down approach to ai transparency. arXiv preprint arXiv:2310.01405

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.