Pith. sign in

REVIEW 3 major objections 5 minor 18 references

Protoreasoning in Tiny Transformers

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that tiny transformers can learn a primitive form of chain-of-thought reasoning—deterministic pruning traces over Dyck bracket trees—and that this training signal substantially narrows their out-of-distribution…

desk verdict A cheap, well-designed reasoning testbed with a strong but overconfident headline; the Low pattern and missing paired statistics are the real soft spots. read the letter →

arxiv 2608.04980 v1 pith:LCAPWQQD submitted 2026-08-05 cs.CL cs.AIcs.LG

Eduardo Valle , Fergal Reid This is my paper
classification cs.CLcs.AIcs.LG
keywords protoreasoningchainofthoughttinytransformersDycklanguagesout-of-distributiongeneralizationformal-languagereasoningscratchpadtracespositional-encodingaugmentation
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

This paper tries to establish that step-by-step reasoning is available far below the scale where natural-language chain of thought is viable: a transformer with roughly one million parameters can profit from training on deterministic reasoning traces, which the authors call protoreasoning. On two tasks defined over Dyck-$k$ bracket languages (deepest path and maximum-order preleaf), adding such a trace to training samples substantially narrows the out-of-distribution generalization gap across six hold-out patterns, while in-distribution validity is already near perfect without it. A dot-filler ablation confirms that the gain comes from the trace's content rather than from the extra tokens it adds. The small scale matters because it makes dense, exactly verifiable experiments feasible—the study comprises over 1700 training runs—offering a cheap way to probe when models learn general algorithms instead of input-output heuristics.

What carries the argument

The load-bearing object is the protoreasoning trace: a deterministic scratchpad of pruning steps over the tree or forest encoded by a Dyck-$k$ sentence, where each step is the previous state after deleting nodes that cannot belong to the answer, so the trace ends exactly when only the answer remains. Step dropout randomly skips 15% of reasoning steps during training, and a positional-index augmentation called SkipAlign remaps position indices so that the longer trace sequences do not confound the comparison with vanilla training. The trace carries the argument because it converts each task from a one-shot input-output mapping into a sequence of intermediate states, and the dot-filler ablation isolates the content of those states as the source of the out-of-distribution gain.

What would settle it

Take any one of the six hold-out patterns and recompute the trace-versus-vanilla comparison with per-seed validity reported separately and with a fresh set of 128 validation prompts drawn for each seed; if the typical seed-level difference is near zero, or if the dot-filler replacement matches full-trace validity once token counts are equalized, the central claim fails.

Watch

Extended reading notes

Core claim

The central discovery is that tiny transformers trained on an input followed by a deterministic pruning trace generalize markedly better out of distribution than models trained on input-output pairs alone. In the deepest-path task the trace removes the shallowest remaining leaf at each step until only the path to the deepest leaf remains; in the maximum-order preleaf task it peels trees from the root and eliminates the smallest two-level sibling brood until one tree remains. Evaluated on held-out strata of depth or preleaf size, the trace format almost always beats the vanilla format across two tasks and six hold-out patterns, with the largest gains in interpolation; extrapolation remains hard for both formats. Replacing trace content with filler dots destroys the advantage, supporting the interpretation that trace content, not token count, drives the gain.

Load-bearing premise

The conclusion that trace content helps assumes that averaging output validity over 10 random seeds on fixed 128-prompt validation sets is a stable estimator, despite many seed-level standard deviations above 10-20 points in the reported tables; if the aggregates are dominated by a few lucky seeds or particular validation prompt draws, the trace advantage is not reliable.

Editorial extensions

If this is right

  • Step-by-step reasoning is not exclusive to frontier models: ~1M-parameter transformers already generalize better out of distribution when trained on deterministic reasoning traces than on input-output pairs alone.
  • The trace advantage appears across two tasks and six hold-out patterns and is largest for interpolation; extrapolation to unseen extremes remains hard for both trace and vanilla formats.
  • Extra tokens alone do not explain the gain: replacing trace content with filler dots dramatically deteriorates validity, and step dropout plus SkipAlign improve generalization by removing length and scaffold confounders.
  • Under mixed-task training, trace tolerates the harder setting far better than vanilla, and the absence of cross-task transfer indicates the generalization limit is about how each task's solution is wired rather than about input representability.
  • The tiny-model setup makes dense experimentation feasible, so reasoning can be studied with exact verification, controlled difficulty, and thousands of training runs on modest compute.

Reading between the lines

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

  • Beyond the paper: if protoreasoning traces hold up, they offer a cheap testbed for mechanistic interpretability of reasoning, since the trace provides exact intermediate states that internal representations can be probed against.
  • Beyond the paper: the mixed-task result suggests compositionality will not emerge from trace training alone; a natural next test is whether traces that share sub-steps across tasks produce modular transfer.
  • Beyond the paper: the dot-filler ablation implies that 'thinking tokens' without task-relevant content should not be expected to improve out-of-distribution generalization in small models, a prediction that could be tested in larger models by perturbing chain-of-thought content.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces "protoreasoning traces" for tiny (~1M-parameter) transformers: deterministic pruning scratchpads that solve two tasks defined over Dyck languages (deepest path and maximum preleaf) by repeatedly removing tree nodes that cannot be part of the answer. Across six hold-out patterns and a mixed-task setting, the paper compares vanilla input-output training with trace training, reporting that traces "almost always" improve out-of-distribution output validity (Tables 2-3). A dot-erasure ablation (Table 4) is used to argue that the improvement comes from trace content rather than extra tokens. The authors also examine the roles of SkipAlign positional augmentation and step dropout, and report that extrapolation remains difficult for both formats.

Significance. If the reported effects are robust, the paper provides a valuable minimal testbed for studying step-by-step reasoning at model scales far below natural-language competence. Strengths of the design include the use of two exactly verifiable tasks with controllable difficulty axes, ten seeds, disjoint train/validation/test strata, deterministic trace construction independent of the results, and a dense experimental campaign (over 1800 final training runs). The finding that trace content, not just token count, drives generalization would be an interesting contribution to the literature on scratchpads and Chain-of-Thought. However, the statistical support for the headline claims is currently incomplete, and the key content ablation is missing in the main experimental setting, so the significance of the claims is not yet fully established.

major comments (3)
  1. [Tables 2-3; Sec. 2.3, A.4] The headline claim that traces "almost always" beat vanilla is not supported by any inference that accounts for seed variance. For example, in Table 2 the DP Low row shows trace at 21.6±7.4 versus vanilla at 39.6±2.0, and PL Low shows 41.0±5.4 versus 43.9±3.0; in Table 3, DP High Mixed shows trace at 43.0±4.9 versus vanilla at 46.6±5.2. Several of the positive comparisons, such as DP Even (98.8±0.9 vs 98.5±2.5), are smaller than the seed standard deviations. The paper reports no paired seed-level differences, no confidence intervals for the differences, and no multiple-comparison-aware testing across the 24 task×pattern comparisons. Since the abstract and conclusion generalize from these aggregates, the "almost always" claim needs paired effect-size reporting (e.g., mean paired difference with confidence interval over seeds, or per-seed scatter plots) and ideally variance estimates over prompt draws.
  2. [Sec. 3.4, Table 4] The dot-erasure ablation, which is the sole evidence for the claim that trace content rather than extra tokens drives the gain, is reported only under RoPE and not under SkipAlign, the positional augmentation used in all main comparisons. This matters because SkipAlign has large effects on trace validity; for instance, Table 4 shows DP Random at 59.8±24.4 under RoPE versus 85.2±8.5 under SkipAlign for the trace condition. Without the erased condition under SkipAlign, the content-not-tokens conclusion is not deconfounded from the positional-encoding regime. Please add the SkipAlign erased condition or explicitly justify why it is unnecessary.
  3. [Sec. A.4, Sec. 2.3] The fixed 128-prompt validation set is used for both hyperparameter selection and best-checkpoint selection, but the paper does not state how many test prompts per stratum are used for the reported OOD numbers or whether test sentences are freshly drawn per seed. If the test evaluation also uses small fixed prompt sets, the reported means and standard deviations may reflect prompt-specific draws rather than model behavior. Please report the test sampling procedure and, if the test set is fixed and small, re-evaluate with fresh prompts or supply bootstrap intervals over prompt draws.
minor comments (5)
  1. [Title/Author block and Abstract] Several words are run together or split in the author block and abstract (e.g., "Eduardo V alle", "F ergal Reid", "wecallprotoreasoning"); these should be cleaned up in the final version.
  2. [Sec. A.4] The contribution statement says "over 1700 training runs" while Appendix A.4 reports 1860 final runs plus 1488 hyperparameter-search runs (3348 total); please clarify which number is meant.
  3. [Table 4] The table header formatting is confusing: the grouping of columns under "Trace", "Erased", and "Vanilla" is not clear from the text. Please reformat so each condition and augmentation is unambiguous.
  4. [Sec. 2.2] Please specify whether the dot-by-dot replacement preserves the exact token count and positional indices of the trace, including delimiters and newlines, since the "extra tokens" confounder is central to the ablation.
  5. [Fig. 7] The top x-axis in Fig. 7 is labeled "number of unique bracket nestings (log)" without specifying the base; the text uses log10, so the figure should match.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's central claims are empirical comparisons over held-out strata with externally generated pruning traces, not derivations from fitted inputs.

full rationale

This paper makes no formal derivation; its central claims are empirical comparisons of output validity across formats, tasks, hold-out patterns, and ablations. The protoreasoning traces are generated by an explicit, deterministic pruning procedure (Section 2.2) defined independently of model performance: for DP the algorithm repeatedly eliminates the shallowest leaf until the tree degenerates into a path, and for PL it peels trees by the root and eliminates the smallest two-level tree, with ties broken leftmost. None of the trace-construction constants, the step-dropout probability, the SkipAlign augmentation, or the hold-out splits is fitted to the reported test validity numbers. Hyperparameters are selected using validation strata that are explicitly disjoint from both training and test strata (Section A.4), and the test results are reported on unseen test strata; this is a selection procedure, not circular reasoning. The abstract's claim that the trace's content rather than its extra tokens drives the gain is supported by the dot-erasure ablation in Table 4, which is an empirical control and not a re-labelled fit. The weaknesses flagged in the reader's take, such as large seed standard deviations, fixed 128-prompt validation sets, and the absence of paired significance tests, are legitimate statistical-evidence concerns about robustness, but they do not make any reported quantity equal to its input by construction. There are no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation. The paper is therefore self-contained as an empirical study, and no circular step can be exhibited.

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

No new physical entities are postulated. The paper introduces the term protoreasoning for deterministic pruning traces, but this is an experimental protocol, not an entity with independent empirical handles. The central claims rest on the assumptions above, mainly the transferability of trace training to no-trace evaluation and the stability of noisy seed-averaged metrics.

free parameters (6)
  • learning rate and weight decay = selected per treatment by 8-trial Sobol search (LR 1e-4 to 1e-2, WD 1e-7 to 1e-1)
    Chosen on OOD validation strata, not test; affects every reported validity number.
  • step dropout probability = 0.15
    Hand-picked; applied independently to each trace step, and ablations show it strongly affects DP results.
  • SkipAlign Dirichlet alpha = 0.4
    Hand-picked; controls position-gap budget splitting in positional augmentation.
  • model size = width 128, depth 4, 16 heads (~1M params)
    Chosen from the model/language scaling experiment (Fig. 9) as the smallest setting with near-perfect expressivity.
  • training schedule = batch 256, 2K pretrain + 4K finetune steps, 600 warmup/decay
    Hand-picked fixed schedule; the central comparisons hold it constant across formats.
  • tie-breaking rules = rightmost for answers, leftmost for pruning
    Hand-chosen to make tasks and traces deterministic; affects task difficulty and trace length.
assumptions (6)
  • standard math Dyck sentences correspond bijectively to forests, with inner brackets as descendants
    Invoked in Section 2.1 to define DP and PL tasks.
  • domain assumption Every input has exactly one correct output and one deterministic gold trace
    Section 2.2 constructs traces by pruning; if traces were not deterministic, the training signal would be ambiguous.
  • domain assumption Training with traces transfers to no-trace prompts at test time
    Section 2.2 removes traces from evaluation; this is the key transfer assumption behind the reported OOD gains.
  • domain assumption Hyperparameters selected on OOD validation strata generalize to disjoint OOD test strata
    Appendix A.4 selects LR/WD and best checkpoint on validation; if validation and test strata are not comparable, test numbers are optimistic.
  • domain assumption SkipAlign position-index remapping preserves task semantics
    Section A.3 inserts position gaps only at semantic boundaries; the model must treat remapped positions as equivalent to natural positions.
  • ad hoc to paper Step dropout yields an order-preserving subsequence from which the model can recover the answer
    Section A.3 defines trace validity as subsequence matching; this scoring rule is introduced specifically for this paper's trace format.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Protoreasoning in Tiny Transformers." pith.science (2026). https://pith.science/paper/LCAPWQQD

@misc{pith2026260804980,
  author       = {Pith},
  title        = {Pith review of: Protoreasoning in Tiny Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LCAPWQQD}},
  note         = {Machine review of arXiv:2608.04980}
}
read the original abstract

We show that tiny transformers can profitably employ a simple form of Chain of Thought, which we call protoreasoning, allowing us to study step-by-step reasoning on ~1M-parameter models and opening up opportunities for much more detailed experimentation and analysis than is feasible for larger models. Current Large Language Models exhibit impressive step-by-step reasoning, but we have yet to understand its generality, i.e., when and how LLMs learn genuinely general algorithms rather than "bags of heuristics." Such questions are hard to settle on compute-intensive frontier models trained on opaque data. To work at model scales far below the threshold for natural-language competence, we define reasoning-friendly tasks on Dyck languages (sentences of correctly nested brackets). We find that protoreasoning traces substantially close the out-of-distribution generalization gap, and ablations confirm that the trace's content, not merely its extra tokens, drives the gain.

Figures

Figures reproduced from arXiv: 2608.04980 by the authors.

Figure 1
Figure 1. A Dyck sentence is formed by well-balanced brack￾ets and corresponds to a for￾est of trees, with inner brack￾ets descending from outer ones. Dyck languages are a central ob￾ject in theoretical linguistics for their ability to represent arbi￾trary depths of recursion. LLMs have demonstrated remarkable step-by-step reasoning capabilities for tasks ranging from mathematical word problems to coding, spurring intense re￾… view at source ↗
Figure 2
Figure 2. Tasks. Deepest path (DP, answer in blue): path to the deepest leaf. Maximum-order preleaf (PL, an￾swer in orange): largest all-leaf sibling brood. Ties are broken rightmost and the answer has only the opening brackets. In￾put Dyck sentence at the figure’s bottom. Dyck-k languages are the sets of all sentences with k types of brackets correctly nested and balanced. For example, ([ ]) is a Dyck-k sentence for k ≥ 2, b… view at source ↗
Figure 3
Figure 3. Training samples for input and tasks of Fig. 2. A token separates input and expected output: [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Six experiments showcasing the gap between expressivity and generalization for the tasks proposed. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Same results as in Tab. 2, shown per stratum. Averaged output validities (lines) [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Same results as in Tab. 3, shown per stratum. Averaged output validities (lines) [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: shows precisely how we split each task’s strata into training, validation, and test sets for single-task training (Sec. 3.2). The top x-axis shows the log10 count of unique bracket nestings in each stratum. We select exactly 24 strata for each task, covering the region…
Figure 8
Figure 8. Figure 8: The three hold-out patterns (Even-, Middle-, and High-Mixed) for training by mixing both tasks. Each square is a joint stratum on both criteria (depth and maximum preleaf size). For a given pattern (columns), each row shows how the examples of a task (rows) are sampled…
Figure 9
Figure 9. Figure 9: In-domain expressivity (top row) and out-of-domain generalization for models trained and tested [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 8 canonical work pages

  1. [5]

    The illusion of the illusion of thinking: A comment on Shojaee et al

    Alex Lawsen. The illusion of the illusion of thinking: A comment on Shojaee et al. (2025). 2506.09250, 6

    arXiv 2025

  2. [9]

    An expressivity analysis of hierarchical modelling in deep transformers via bounded-depth grammars

    URLhttps://openreview.net/forum?id=9XFSbDPmdW. VinothNandakumar, QiangQu, PramodThebe, SakshiKhachariyapatel, andTongliangLiu. Anexpressivity analysis of hierarchical modelling in deep transformers via bounded-depth grammars. arXiv:2606.17522,

    work page Pith review arXiv

  3. [10]

    Jacob Pfau, William Merrill, and Samuel R

    URLhttps://proceedings.iclr.cc/paper_files/paper/2025/hash/ 8c5f30296296d2ae402ebbd09aaa9c12-Abstract-Conference.html. Jacob Pfau, William Merrill, and Samuel R. Bowman. Let’s think dot by dot: Hidden computation in transformer language models. InProceedings of the 2024 Conference on Language Modeling,

    work page 2025

  4. [11]

    URLhttps://dl.acm.org/doi/10.1145/3774896

    doi: 10.1145/3774896. URLhttps://dl.acm.org/doi/10.1145/3774896. Rylan Schaeffer, Brando Miranda, and Sanmi Koyejo. Are emergent abilities of large lan- guage models a mirage? InAdvances in Neural Information Processing Systems, volume 36, pp. 55565–55581,

    doi:10.1145/3774896

  5. [12]

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom

    URLhttps://proceedings.neurips.cc/paper_files/paper/2023/hash/ adc98a266f45005c403b8311ca7e8bd7-Abstract-Conference.html. Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. In Advances in Neural Information Proces...

    work page 2023

  6. [13]

    Parshin Shojaee, Iman Mirzadeh, Keivan Alizadeh, Maxwell Horton, Samy Bengio, and Mehrdad Farajtabar

    URL https://openreview.net/forum?id=sPIXjSkDFG. Parshin Shojaee, Iman Mirzadeh, Keivan Alizadeh, Maxwell Horton, Samy Bengio, and Mehrdad Farajtabar. The illusion of thinking: Understanding the strengths and limitations of reasoning models via the lens of problem complexity. arXiv:2506.06941, 6

    arXiv

  7. [14]

    doi: 10.1162/tacl_a_00663

    ISSN 2307-387X. doi: 10.1162/tacl_a_00663. URLhttps://doi.org/10.1162/tacl_a_00663. Jiankai Sun, Chuanyang Zheng, Enze Xie, Zhengying Liu, Ruihang Chu, Jianing Qiu, Jiaqi Xu, Mingyu Ding, Hongyang Li, Mengzhe Geng, Yue Wu, Wenhai Wang, Junsong Chen, Zhangyue Yin, Xiaozhe Ren, Jie Fu, Junxian He, Wu Yuan, Qi Liu, Xihui Liu, Yu Li, Hao Dong, Yu Cheng, Ming ...

    doi:10.1162/tacl_a_00663

  8. [15]

    URLhttps: //doi.org/10.1145/3729218

    doi: 10.1145/3729218. URLhttps: //doi.org/10.1145/3729218. Miles Turpin, Julian Michael, Ethan Perez, and Samuel R. Bowman. Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting. InAdvances in Neural Information Processing Systems (NeurIPS),

    doi:10.1145/3729218

Show all 18 references
  1. [16]

    Iñaki Dellibarda Varela, Pablo Romero-Sorozabal, Eduardo Rocon, and Manuel Cebrian

    URLhttps://proceedings.neurips.cc/paper_files/ paper/2024/file/2f6a6317bada76b26a4f61bb70a7db59-Paper-Conference.pdf. Iñaki Dellibarda Varela, Pablo Romero-Sorozabal, Eduardo Rocon, and Manuel Cebrian. Rethinking the illusion of thinking. arXiv:2507.01231, 7

    2024 arXiv

  2. [18]

    ShunyuYao, DianYu, JeffreyZhao, IzhakShafran, ThomasL.Griffiths, YuanCao, andKarthikNarasimhan

    URLhttps: //arxiv.org/abs/2405.03939. ShunyuYao, DianYu, JeffreyZhao, IzhakShafran, ThomasL.Griffiths, YuanCao, andKarthikNarasimhan. Tree of thoughts: Deliberate problem solving with large language models. InAdvances in Neural Infor- mation Processing Systems (NeurIPS), 2023a...

    arXiv

  3. [19]

    URL https://doi.org/10.1109/TPAMI.2025.3637037

    doi: 10.1109/TPAMI.2025.3637037. URL https://doi.org/10.1109/TPAMI.2025.3637037. Dawei Zhu, Nan Yang, Liang Wang, Yifan Song, Wenhao Wu, Furu Wei, and Sujian Li. PoSE: Efficient con- text window extension of LLMs via positional skip-wise training. InThe Twelfth International C...

    2025

  4. [20]

    A Appendix A.1 Holdout patterns Fig

    URLhttps://openreview.net/forum?id=3Z1gxuAQrA. A Appendix A.1 Holdout patterns Fig. 7 shows precisely how we split each task’s strata into training, validation, and test sets for single-task training (Sec. 3.2). The top x-axis shows thelog10 count of unique bracket nestings in...

    2024

  5. [2020]

    doi: 10.1162/tacl_a_ 00306

    ISSN 2307-387X. doi: 10.1162/tacl_a_ 00306. URLhttps://doi.org/10.1162/tacl_a_00306. Jie Huang and Kevin Chen-Chuan Chang. Towards reasoning in large language models: A survey. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki (eds.),Findings of the Association for Comput...

    doi

  6. [2021]

    doi: 10.18653/v1/2021.naacl-main.405

    Association for Com- putational Linguistics. doi: 10.18653/v1/2021.naacl-main.405. URLhttps://aclanthology.org/2021. naacl-main.405/. William Merrill and Ashish Sabharwal. A logic for expressing log-precision transformers. InProceedings of the 37th International Conference on ...

    2021 doi

  7. [2023]

    the illusion of thinking

    Association for Computational Linguis- tics. doi: 10.18653/v1/2023.findings-acl.67. URLhttps://aclanthology.org/2023.findings-acl.67/. 12 Sheraz Khan, Subha Madhavan, and Kannan Natarajan. A comment on "the illusion of thinking": Reframing the reasoning cliff as an agentic gap...

    2023 arXiv

  8. [2024]

    URLhttps://aclanthology.org/2024

    doi: 10.18653/v1/2024.acl-long.65. URLhttps://aclanthology.org/2024. acl-long.65/. Bishwamittra Ghosh, Soumi Das, Till Speicher, Qinyuan Wu, Mohammad Aflah Khan, Deepak Garg, Kr- ishna P. Gummadi, and Evimaria Terzi. Fine-tuning vs. in-context learning in large language models...

    2024 doi

  9. [2025]

    Measuring faithfulness in chain-of- thought reasoning

    Tamera Lanham, Anna Chen, Ansh Radhakrishnan, Benoit Steiner, Carson Denison, Danny Hernandez, Dustin Li, Esin Durmus, Evan Hubinger, Jackson Kernion, et al. Measuring faithfulness in chain-of- thought reasoning. arXiv:2307.13702,

    arXiv

  10. [2026]

    Chu-Cheng Lin, Aaron Jaech, Xin Li, Matthew R

    URL https://arxiv.org/abs/2604.20811. Chu-Cheng Lin, Aaron Jaech, Xin Li, Matthew R. Gormley, and Jason Eisner. Limitations of autoregres- sive models and their alternatives. In Kristina Toutanova, Anna Rumshisky, Luke Zettlemoyer, Dilek Hakkani-Tur, Iz Beltagy, Steven Bethard...

    2021 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.