Pith. sign in

REVIEW 4 major objections 4 minor 15 references

Deductive Logic in Language Models: Horizontal vs Vertical Reasoning

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper establishes that Chain-of-Thought supervision enables a tiny attention-only transformer to learn genuine deductive rules, implemented by induction-head circuits.

desk verdict A solid empirical result about CoT enabling rule-learning in a tiny transformer, wrapped in a plausible-but-unproven induction-head circuit story and an abstract that promises vertical-reasoning experiments that are absent. read the letter →

arxiv 2510.09340 v2 pith:VD6MKHPY submitted 2025-10-10 cs.AI cs.CL

Davide Maltoni , Matteo Ferrara This is my paper
classification cs.AIcs.CL
keywords deductivereasoninglanguagemodelsmechanisticinterpretabilityinductionheadschain-of-thoughttransformercircuitslogicalinferencesynthetictask
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 shows that a 144,000-parameter, two-layer attention-only transformer, trained from scratch on a symbolic deductive reasoning task, can learn the underlying inference rules rather than statistical shortcuts when supervised with Chain-of-Thought. By decoding internal attention states, the authors identify three circuits—rule completion, rule chaining, and final decision—each built on induction-head-like copy-search-retrieve operations. They also show that without Chain-of-Thought the model memorizes the training set or exploits dataset biases, so the supervision is what forces genuine rule learning. A secondary contribution is a truncated-pseudoinverse technique that decodes queries, keys, and values back into token space, offering a new lens for mechanistic interpretability.

What carries the argument

The load-bearing mechanism is the induction head: an attention pattern that completes a sequence [A], [B] ... [A] with [B] by copying the query token, searching for a prior position where that token appears as a key, and releasing the associated value. In this model, the same mechanism implements both rule completion (finding A→X for a given A) and rule chaining (linking the tail of one implication to the head of the next). A second piece of machinery is the truncated-pseudoinverse decoder, which retro-projects query, key, and value vectors from their subspaces back into the residual-stream space, allowing them to be read out as tokens; this is what makes the circuit-level explanation visibl

What would settle it

Patch or ablate the specific attention links identified as rule-completion (links 11-15) and rule-chaining (links 31-34) circuits in a converged model: if the model's output chain does not change in the expected way (e.g., breaking exactly at the patched step), or if changing the key token at the source position does not alter the generated rule completion, then the circuit-level account is not causally correct.

Watch

Extended reading notes

Core claim

The paper's central claim is that a very simple, non-pretrained language model, when trained with Chain-of-Thought prompting, learns to solve multi-step deductive inference by implementing the actual inference rules. On the task of deciding whether a query implication follows from five given implications, the model achieves ~100% accuracy and generalizes across a generation space of hundreds of billions of examples. Mechanistic analysis shows the inference unfolds in three stages—completing the current rule, chaining rules, and making the final yes/no decision—carried out by attention circuits that behave like induction heads: they copy a literal, search for its occurrence in a rule head, an

Load-bearing premise

The explanation relies on the assumption that the truncated pseudoinverse decoding, with hand-tuned singular-value thresholds, faithfully recovers the token-level information the attention mechanism actually uses—if the decoded queries, keys, and values are artifacts of the chosen thresholds, the identified circuits are not the cause of the behavior.

Editorial extensions

If this is right

  • A two-layer attention-only transformer with a single head is sufficient to implement multi-step deductive inference, provided the residual stream is wide enough.
  • Chain-of-Thought supervision is causal for rule learning: with binary labels only, the model memorizes and fails to generalize.
  • Induction heads, previously linked to in-context learning and text completion, also serve as the computational primitive for logical rule completion and chaining.
  • The emergence of induction heads coincides with sharp generalization jumps, and the final decision circuit forms later than the rule circuits, giving a staged picture of capability acquisition.
  • The truncated pseudoinverse decoding method may generalize as an interpretability tool for inspecting query/key/value semantics in other transformers.

Reading between the lines

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

  • If induction heads are the core primitive, then increasing chain length or rule complexity may require either more layers or additional heads to host more induction circuits; this could be tested by scaling the current task.
  • The finding that CoT acts as curriculum in the vertical setting suggests that hidden-layer progressions (not just output tokens) may be exploitable as a training signal for tasks where explicit step-by-step outputs are unavailable.
  • The bias-exploitation result (the model counting distinct literals) implies that synthetic reasoning benchmarks must include negative controls where such counting shortcuts are impossible, otherwise accuracy gains may reflect dataset structure rather than reasoning.
  • A directly checkable prediction: because the paper claims a single head can host multiple induction circuits, models with fewer heads (given wider residual streams) should still solve similar chaining tasks; if they cannot, the single-head hosting claim is wrong.
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

4 major / 4 minor

Summary. The paper studies a two-layer, single-head, attention-only transformer (NanoGPT variant, 144K parameters) trained from scratch on a synthetic deductive task: given five symbolic implications and a query implication, the model must output a chain-of-thought sequence and a final 0/1 verdict. The authors report near-100% full-sequence accuracy on a held-out validation set, with a generation space of roughly 4.7×10^11 instances, and argue that CoT supervision enables the model to learn the inference rules rather than memorize or exploit dataset statistics. The main mechanistic claim is that the inference is implemented by induction-head-like circuits: layer-1 attention copies tokens, and the single layer-1→layer-2 head performs rule completion, rule chaining, and the final decision, according to attention visualizations, LogitLens, and a newly introduced truncated-pseudoinverse decoding of queries, keys, and values. A checkpoint analysis (§5.2) relates accuracy jumps to the formation of these circuits. The paper also frames the contribution around a horizontal/vertical reasoning distinction, but the vertical side is not experimentally developed.

Significance. If the mechanistic account were established, this would be a useful contribution: it gives a concrete low-level account of how a minimally sized transformer can implement symbolic inference, introduces a potentially reusable interpretability tool (truncated pseudoinverse), and uses a well-controlled synthetic setting with a genuinely large sample space and reproducible code. The empirical generalization result is convincing: 3,072 training examples cannot cover 4.7×10^11 instances, and full-sequence accuracy on a disjoint validation split is strong evidence against memorization. However, the circuit-level story is currently supported only by passive, threshold-dependent inspection, and there is a potential flaw in the positive-example generation. The paper is honest about several limitations, but those limitations directly affect the central mechanistic claim. The learning and generalization result stands; the mechanistic explanation needs substantially stronger evidence.

major comments (4)
  1. [Appendix A / §3.1] The positive-example construction is not well-defined. If the coin toss in Step 4 replaces the head l_b of l_b→l_{b+1} with l_{m+2}, the modified rule is l_{m+2}→l_{b+1}; then no chain from q0=l_1 to l_{m+2} exists, so setting q1=l_{m+2} does not yield a logical consequence. Only the tail-replacement case yields a path. The text says the conversion is deterministic without restricting the coin-toss outcome. Please clarify whether head-replacement cases are discarded or handled differently, and verify a sample of generated labels with an independent solver.
  2. [§4.2–4.4, Appendix B, §5] The central mechanistic claim rests on truncated-pseudoinverse decoding with thresholds chosen by inspection (W_Q/W_V 0.75–0.85, W_K 0.95–0.99) and no causal validation. Because W_Q, W_K, and W_V are full-rank square matrices, exact inversion would recover the original residual stream; the truncation is what creates the token-level readings. The paper concedes in Section 5 that expected tokens may not appear in top-2 or top-1 after retro-projection, so the correspondence between decoded Q/K/V and the proposed algorithm could be an artifact of threshold choice. No activation patching or ablation is performed on links 11–15, 31–34, 51, or 61–65. I recommend adding a threshold-sensitivity analysis over a plausible range and an intervention test (e.g., patching a specific attention link) to show the identified circuits are causally necessary.
  3. [§4 (opening), §5] The circuit description is based on a single guiding example. The statement in Section 5 that “the same behavior occurs across other positive and negative examples” is not accompanied by quantitative evidence. Show, for a random sample of examples, how often the decoded queries/keys/values match the expected token, how often the identified links have high attention strength, and whether the same three-stage decomposition holds. Without this, the claim that a single attention head hosts multiple distinct induction-head circuits (Section 6) is unsupported.
  4. [Abstract, §5.1, title] The title and the abstract promise a comparison between horizontal and vertical reasoning and mention a second task (root-to-leaf navigation in binary trees). The main text contains only the implication-chain task; vertical reasoning appears only as a speculative paragraph in §5.1, and no binary-tree experiments are reported. Either add the missing experiments/analyses or revise the title and abstract to match the actual scope.
minor comments (4)
  1. [§2.2] Typo: “Trasformer-based” should be “Transformer-based”.
  2. [Appendix B] The notation for the truncated pseudoinverse is slightly under-specified. Define W_k = U_k Σ_k V_k^T explicitly before writing W_k^+ = V_k Σ_k^+ U_k^T.
  3. [§5.1] The phrase “with some surprise” is informal; consider replacing it with neutral wording.
  4. [§3.3] The 20% convergence rate for the simplified architecture is worth a brief discussion in terms of whether the reported mechanism is representative of all converged runs; the authors state that they examined additional models, but a quantitative summary would strengthen the paper.

Circularity Check

1 steps flagged · score 2.0 of 10

Core learning claim is self-contained; circuit explanation is partly self-confirming due to post hoc pseudoinverse thresholds and manual ranking adjustment.

  1. other [Section 4.4 (final decision) and Appendix B (truncated pseudoinverse)]
    "Although F does not appear among the top-2 decoded tokens from the residual stream, an extra manual inspection confirmed that it takes the third place in the ranking. ... since a projection through a full-rank matrix does not reduce dimensionality, all the information, even if rearranged, is preserved (i.e., retro-projection using the exact inverse would recover the original residual stream content)."

    The circuit claim that link 51 copies the query tail F is supported by a decoding result that is manually searched until the expected token is found, rather than by a pre-specified criterion; had the top-2 result been taken at face value, the claim would have failed. Because W_Q, W_K, and W_V are full-rank, the exact pseudoinverse recovers the original residual stream, so the token-level readings are produced by the truncation; the thresholds (0.75-0.85 for W_Q/W_V, 0.95-0.99 for W_K) are selected post hoc without an independent, pre-registered criterion. The decoded Q/K/V evidence is therefore made consistent with the expected induction-head story, making the circuit interpretation partly self-confirming rather than an independent test.

full rationale

The paper's core learning/generalization claim is self-contained and does not rely on the interpretability machinery: a 144K-parameter, 2-layer attention-only LM trained with CoT reaches ~100% full-sequence accuracy on a held-out validation set drawn from a ~4.7e11-instance generation space, while the no-CoT control memorizes. That accuracy claim is measured independently of the pseudoinverse decoding and of the induction-head labels, so the central 'model learned the inference rules' result is not circular. The mechanistic part is weaker: decoded Q/K/V content is produced by a truncated pseudoinverse with hand-tuned thresholds, and the paper concedes that expected tokens may not appear in top-2 residuals or top-1 retro-projections. In one concrete case, the expected token was accepted only after an extra manual inspection below the top-2 ranking. This is a self-confirming tendency in the circuit evidence, but it is an interpretability-validity limitation rather than a derivation that reduces to its inputs. Self-citations (Maltoni and Ferrara 2024) are contrastive and not load-bearing. Overall: no significant circularity; score 2 reflects the minor post hoc adjustment in the circuit evidence while the main learning claim remains independent.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central learning result depends only on standard training and architecture choices. The interpretability claims, however, rest on hand-tuned pseudoinverse thresholds and arbitrary attention-strength cutoffs, which are essentially free parameters for the analysis. The invented 'multiple circuits in one head' construct lacks independent causal validation.

free parameters (4)
  • truncated pseudoinverse threshold for W_Q = 0.75-0.85
    Hand-tuned so that the decoded query tokens match the expected literals in the circuit analysis (Appendix B, Section 4.2).
  • truncated pseudoinverse threshold for W_V = 0.75-0.85
    Hand-tuned for value decoding; the authors report optimal thresholds in this range for W_V (Appendix B).
  • truncated pseudoinverse threshold for W_K = 0.95-0.99
    Keys require retaining more variance to make decoded keys interpretable; the threshold is chosen per matrix (Appendix B).
  • attention strength threshold for visualization = 0.4 and 0.1
    Arbitrary thresholds used to filter which attention links are displayed in Figures 3-5; these thresholds affect which links are labeled as circuits.
assumptions (4)
  • domain assumption The induction-head mechanism as defined by Olsson et al. applies to single-head attention-only transformers.
    The paper interprets the attention links as induction heads based on the prior framework of Olsson et al. (2022); the authors do not re-derive the mechanism.
  • domain assumption The dataset generation hash prevents any overlap between training and validation, so validation accuracy reflects generalization.
    Appendix A states a hash table is used to avoid duplicates, but no collision probability or hash verification is reported.
  • domain assumption LogitLens decoding of the residual stream at intermediate layers is a faithful proxy for the model's internal computation.
    Section 4.1 acknowledges limitations but relies on LogitLens to identify the tokens represented in the residual stream; if the representations are not linearly decodable, the circuit analysis is unreliable.
  • ad hoc to paper The three-stage decomposition (rule completion, chaining, final decision) is exhaustive and causally necessary for the model's inference.
    The authors introduce this decomposition in Section 4 but do not prove that these stages cover all computation or that each identified link is required for the output.
invented entities (1)
  • Multiple distinct induction-head circuits hosted within a single attention head
    purpose: Explains how one attention head can simultaneously perform rule completion, chaining, and final decision via separate groups of attention links.
    The paper labels groups of links (11-15, 31-34, 61-65) as separate circuits within the same head, but provides no causal evidence that these are independent or necessary; it is a post-hoc interpretation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deductive Logic in Language Models: Horizontal vs Vertical Reasoning." pith.science (2026). https://pith.science/paper/VD6MKHPY

@misc{pith2026251009340,
  author       = {Pith},
  title        = {Pith review of: Deductive Logic in Language Models: Horizontal vs Vertical Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VD6MKHPY}},
  note         = {Machine review of arXiv:2510.09340}
}
read the original abstract

Recent language models exhibit significant logical reasoning abilities, yet the mechanisms supporting deductive inference remain poorly understood. This paper studies small transformer-based language models trained from scratch on multi-step deductive tasks, focusing on the distinction between horizontal reasoning, where intermediate steps are generated autoregressively, and vertical reasoning, where inference unfolds implicitly across layers before the first output token is produced. We analyze two synthetic tasks: logical consequence over chains of symbolic implications and root-to-leaf navigation in binary trees. Mechanistic interpretability reveals that Chain-of-Thought supervision enables models to learn rule-based inference rather than statistical shortcuts. In the horizontal setting, a shallow attention-only model develops interpretable circuits for rule completion, rule chaining, and final decision making, largely implemented through induction-head-like mechanisms. We further introduce a truncated pseudoinverse method to decode the information carried by queries, keys, and values. For vertical reasoning, Chain-of-Thought appears to act less as explicit step-by-step guidance and more as a form of curriculum learning, helping the model acquire increasingly complex reasoning patterns. Without Chain-of-Thought, models tend to memorize or exploit dataset biases. These results provide a low-level account of how transformers can implement deductive reasoning and suggest how Chain-of-Thought may serve different functions in horizontal and vertical reasoning.

Figures

Figures reproduced from arXiv: 2510.09340 by the authors.

Figure 1
Figure 1. Average accuracy over 20 runs reaching convergence. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. An example of output produced by the developed visualization tool. The explanation is in the main text. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Circuits involved in Rule completion. The explanation is in the main text. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Circuits involved in Rule chaining. The explanation is in the main text. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Circuits involved in Start and final decision. The explanation is in the main text. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: The model convergence during a specific run, with identification of points of interest [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 5 linked inside Pith

  1. [1]

    Huang J., Chang K. C. C., Towards Reasoning in Large Language Models: A Survey, inFindings of the Association for Computational Linguistics: ACL 2023, pp. 1049–1065,

    2023

  2. [5]

    Olsson C., Elhage N., Nanda N., Joseph N., DasSarma N., Henighan T., Mann B., Askell A., Bai Y ., Chen A., Conerly T., Drain D., Ganguli D., Hatfield-Dodds Z., Hernandez D., Johnston S., Jones A., Kernion J., Lovitt L., Ndousse K., Amodei D., Brown T., Clark J., Kaplan J., McCandlish S., Olah C., In-context Learning and Induction Heads, Transformer Circui...

    2022

  3. [8]

    Ferrando J., Sarti G., Bisazza A., Costa-jussà M. R., A Primer on the Inner Workings of Transformer-based Language Models,arXiv: 2405.00208, 2024 Madsen A., Reddy S., Chandar S., Post-hoc Interpretability for Neural NLP: A Survey, inACM Computing Surveys, vol. 55,

    arXiv 2024

  4. [10]

    Heimersheim S., Nanda N., How to use and interpret activation patching,arXiv: 2404.15255,

    arXiv

  5. [11]

    URL: https://www.alignmentforum.org/posts/z6QQJbtpkEAX3Aojj/interim-research-report- taking-features-out-of-superposition (Accessed: 2025-10-01) Olah C., Mechanistic Interpretability, Variables, and the Importance of Interpretable Bases, Transformer Circuits Thread - https://transformer-circuits.pub/2022/mech-interp-essay,

    2025

  6. [12]

    Elhage N., Nanda N., Olsson C., Henighan T., Joseph N., Mann B., Askell A., Bai Y ., Chen A., Conerly T., DasSarma N., Drain D., Ganguli D., Hatfield-Dodds Z., Hernandez D., Jones A., Kernion J., Lovitt L., Ndousse K., Amodei D., Brown T., Clark J., Kaplan J., McCandlish S., Olah C., A Mathematical Framework for Transformer Circuits,Transformer Circuits T...

    2021

  7. [13]

    URL: https://www.alignmentforum.org/posts/u6KXXmKFbXfWzoAXn/a-circuit-for-python- docstrings-in-a-4-layer-attention-only (Accessed: 2025-10-01) Hanna M., Liu O., Variengien A., How does GPT-2 compute greater-than? interpreting mathe- matical abilities in a pre-trained language model, in37th International Conference on Neural Information Processing Systems (NIPS),

    2025

  8. [14]

    Wei J., Wang X., Schuurmans D., Bosma M., Ichter B., Xia F., Chi E., Le Q., Zhou D., Chain-of- Thought Prompting Elicits Reasoning in Large Language Models,arXiv: 2201.11903,

    arXiv

Show all 15 references
  1. [15]

    URL: https://www.alignmentforum.org/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit- lens (Accessed: 2025-10-01) Maltoni D., Ferrara M., Arithmetic with language models: From memorization to computation, Neural Networks, vol. 179,

    2025

  2. [2020]

    Chen M. K., Zhang X., Tao D., JustLogic: A Comprehensive Benchmark for Evaluating Deduc- tive Reasoning in Large Language Models,arXiv: 2501.14851, 2025 Tafjord O., Dalvi B., Clark P., ProofWriter: Generating Implications, Proofs, and Abductive Statements over Natural Language...

    2025 arXiv

  3. [2021]

    Clark P., Tafjord O., Richardson K., Transformers as Soft Reasoners over Language, in29th International Joint Conference on Artificial Intelligence (IJCAI-2020),

    2020

  4. [2022]

    Vig J., Gehrmann S., Belinkov Y ., Qian S., Nevo D., Singer Y ., Shieber S., Investigating Gender Bias in Language Models Using Causal Mediation Analysis, inAdvances in Neural Informa- tion Processing Systems (NeurIPS 2020), vol. 33, pp. 12388-12401,

    2020

  5. [2023]

    Liu H., Fu Z., Ding M., Ning R., Zhang C., Liu X., Zhang Y ., Logical Reasoning in Large Language Models: A Survey,arXiv: 2502.09100,

    arXiv

  6. [2024]

    Li J., Cao P., Jin Z., Chen Y ., Liu K., Zhao J., MIRAGE: Evaluating and Explaining Inductive Reasoning Process in Language Models,13th International Conference on Learning Repre- sentations (ICLR 2025),

    2025

  7. [2025]

    323–339,

    13 Bowen C., Sætre R., Miyao Y ., A Comprehensive Evaluation of Inductive Reasoning Capabilities and Problem Solving in Large Language Models, inFindings of the Association for Computa- tional Linguistics: EACL 2024, pp. 323–339,

    2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.