PaperlayerView exact revision on arXiv

Source-linked AI summary

Does Latent Chain-of-Thought Actually Reason? A Single, Unifying Taxonomy for Latent Reasoning

Jingjie Qiu, Sheldon Yu, Xunyi Jiang, Jiajin Liu, Zihan Huang, Zhendong Chu, Tong Yu, Julian McAuley, Junda Wu

arXiv:2610.21182v1cs.AI

TL;DR

Readable or accurate outputs do not establish that a latent chain is used to reach an answer. The paper combines interventions, probes, and trajectory analysis across tasks and recipes, finding that chain use, content, and dynamics vary independently enough to require separate tests.

  • Problem

    Task accuracy and decodable intermediate states do not show whether a latent chain contributes to forming the answer.

    Source: §1 · p. 1

  • Method

    Across tasks and recipes, the paper combines chain deletion and replacement with probes and trajectory analysis to test use, content, and state dynamics.

    Source: §1 · p. 2, §4 · p. 5

  • Results

    On logical tasks, unsupervised chains can carry answers readable before the chain begins, while per-step supervision can install ordered steps and chain dependence; tested arithmetic chains depend on their contents.

    Source: §1 · p. 2

  • Takeaways & Limitations

    A chain’s readable content, causal use, and state movement are distinct properties, so claims about latent reasoning should specify which property they test.

    Source: §1 · p. 3, §8 · p. 10

    Comparisons involving CODI change both mechanism and supervision, and the two backbones also differ in scale and instruction tuning.

    Source: §3.2 · p. 4

Abstract

from arXiv · show

Latent chain-of-thought replaces written reasoning steps with continuous hidden states, giving models additional steps to compute before answering. It remains unclear what computation these steps actually perform. High task accuracy and decodable intermediate answers do not establish that a model uses its latent chain to reach an answer. We study this question across twenty-one model-task configurations covering three training recipes, three reasoning tasks, and two backbones. We combine chain deletion and content replacement with probes and trajectory analysis to test whether the chain affects the answer, what it encodes, and how its states change. On logical tasks, COCONUT often carries an answer that is already readable before the chain begins. Its states settle early, and deleting the chain leaves predictions almost unchanged. Per-step supervision produces an ordered layout of reasoning steps and can make the model depend on the chain. Yet even an ordered chain can act as a placeholder, with deletion hurting performance while content replacement leaves answers unchanged. On arithmetic tasks, all tested recipes show dependence on chain content. These results show why readable latent states alone are insufficient evidence of reasoning. We introduce a taxonomy that separates what a chain represents from how the model uses it. Seen this way, earlier reports no longer compete: placeholder tokens, collapsed superposition and step-dependent causal roles are each consistent with a different kind of chain, each obtained on a different task, training recipe or measurement. Once a study says which property it tested, whether the model uses the chain, what the chain holds, or how its state moves, claims about latent reasoning can be compared.

1 INTRODUCTION

The paper asks whether latent chains help construct answers or merely follow information already present, since accuracy and readable states alone cannot establish that the chain is used. It tests chain use, encoded content, and state dynamics across tasks and training recipes, then unifies the resulting behaviors in a taxonomy.

1 INTRODUCTION

  • Deleting unsupervised chains changes at most three logical-task answers but hurts arithmetic accuracy, while per-step supervision can establish chain reliance.On one logical-task cell, deletion hurts but replacements do not change answers, revealing placeholder-like reliance on positions rather than content.Source: §1 · p. 2
  • On logical tasks, unsupervised chains can carry answers readable before the chain begins, while per-step supervision orders reasoning steps across slots.The ordered layout can still serve only as extra positions under tested interventions.Source: §1 · p. 2
  • Without per-step supervision, logical-task chains settle within the first slots; per-step chains continue changing to the last slot, with movement toward the answer on only one task.Mid-chain edits show that answer revision varies by backbone, distinguishing revision from state movement.Source: §1 · pp. 2–3
  • The paper combines causal interventions, content probes, and trajectory analysis to distinguish what a chain represents from how the model uses it.Its taxonomy treats placeholder behavior, collapsed superposition, and step-dependent causal roles as different chain types associated with different tasks, recipes, or measurements.Source: Figure 1 · p. 2, §1 · p. 3

2 RELATED WORK

Prior work studies latent chains through their training methods, readable representations, and causal effects. These approaches measure different properties, motivating analyses that connect state content and trajectories to interventions on the answer.

2 RELATED WORK

  • Latent-chain methods vary in how they create continuous states, from recycled hidden states to answer-position alignment and direct supervision of individual slots.Comparing COCONUT with its per-step supervised variant holds the inference mechanism fixed.Source: §2 · p. 3
  • Probing and trajectory analyses reveal information or structure in latent states, but predictability alone does not show that the model uses that information.Prior findings include candidate superposition collapsing after fine-tuning (Rizvi-Martel et al., 2026).Source: §2 · p. 3
  • Prior causal studies report placeholder behavior and different roles across steps, while related work finds observable patterns in controls and influence concentrated in small state subspaces.The paper pairs content and trajectory measurements with interventions on the same model-task configurations.Source: §2 · pp. 3–4

3 EXPERIMENTAL SETUP

The experiments model latent reasoning as continuous hidden-state slots fed back before answer decoding, and compare three training recipes across two backbones and three reasoning tasks. A fixed held-out evaluation and complementary interventions, probes, and trajectory tests support the comparisons.

3.1 LATENT-CHAIN FORMULATION

  • Each slot’s hidden state is computed from the question and earlier slots, and the model decodes an answer from the question and full chain.The model-specific mapping from hidden state to next input is identity for COCONUT and includes a projection network for CODI.Source: §3.1 · p. 4

3.2 MODELS, TASKS, AND CONTROLS

  • The study compares COCONUT, CODI, and per-step-supervised SIM-COT-COCONUT on proof search, syllogisms, and arithmetic using GPT-2 and Llama-3.2-1B-Instruct.COCONUT versus SIM-COT-COCONUT holds inference mechanism and slot budget fixed; CODI comparisons change both mechanism and supervision, and backbone differences are interpreted cautiously.Source: §3.2 · p. 4

3.3 EVALUATION PROTOCOL

  • Each model-task cell uses the same 300 held-out problems, with parameters fixed while deletion, content substitution, or mid-chain edits alter latent computation.Probes use problem-separated folds and permuted-label nulls; trajectory statistics use specified nulls, and each test reports a minimum detectable effect.Source: §3.3 · pp. 4–5, Table 1 · p. 5

4 RQ1: DOES THE MODEL RELY ON THE LATENT CHAIN, ON ITS POSITIONS OR ON ITS CONTENTS?

Chain dependence varies by task, supervision, and backbone: deletion and content replacement distinguish reliance on positions from sensitivity to supplied vectors. In particular, arithmetic results show content dependence, while some logical-task chains behave as placeholders.

4 RQ1: DOES THE MODEL RELY ON THE LATENT CHAIN, ON ITS POSITIONS OR ON ITS CONTENTS?

  • Deletion removes both chain content and computation steps, while replacement keeps positions; pairing them distinguishes content-sensitive chains from placeholder-like patterns.Inert means at most three answers change after deletion; consulted means deletion reduces accuracy beyond the cell’s minimum detectable effect.Source: §4 · pp. 5–6
  • Deleting GPT-2 COCONUT leaves ProsQA accuracy at 0.973 but reduces GSM8k accuracy from 0.333 to 0.103, and the logical-versus-arithmetic split also appears on the other backbone.CODI depends on its chain on GSM8k but has little deletion effect on PrOntoQA; ProsQA behavior differs by backbone.Source: §4 · p. 6
  • Per-step supervision can establish chain dependence where the COCONUT base has little deletion effect: GPT-2 SIM-COT-COCONUT ProsQA accuracy falls from 0.993 to 0.800.The continued-training control remains inert, supporting an effect of per-step loss beyond extra training; the selected Llama-3.2-1B-Instruct PrOntoQA checkpoint is an exception.Source: §4 · p. 6
  • All tested GSM8k recipes on both backbones respond to content replacement; across eight content-used cells, donor replacement is more harmful than deletion in five.For GPT-2 COCONUT/GSM8k, the accuracy change is −0.307 under donor replacement versus −0.230 under deletion.Source: §4 · p. 6
  • On Llama-3.2-1B-Instruct ProsQA, CODI and per-step cells lose performance under deletion but show no answer changes under tested replacements, including a different-answer donor at twice natural norm.The question remains accessible during replacement, so the result does not imply later computation uses no problem information.Source: §4 · p. 6

5 RQ2: WHAT IS WRITTEN IN THE LATENT CHAIN, AND DOES THE MODEL USE IT?

The probes distinguish what latent chains encode from whether the model uses them: unsupervised chains can carry an answer already readable before the chain, while per-step supervision distributes ordered reasoning steps across slots.

RQ2: What is written in the latent chain?

  • The study contrasts superposition, ordered stepwise reasoning, and answers computed during the question pass as alternative accounts of chain content.The probe setup reads final-layer profiles by slot and checks readability against a permuted-label null and a cell-specific minimum detectable effect.Source: §5 · p. 7
  • Unsupervised chains carry answers readable before the chain begins, while per-step supervision makes answer readability build monotonically across slots.The pre-chain answer is also readable on both independent COCONUT trainings; the Llama-3.2-1B-Instruct COCONUT projection artifact has a ridge-probe readout of 0.760 at slot 0.Source: §5 · p. 7
  • Per-step supervision places step k in slot k, but the same ordered layout can occur in a chain whose contents are replaceable without changing the answer.The training-time control without per-step loss shows no ordered path, while the Llama-3.2-1B-Instruct SIM-COT-COCONUT/ProsQA chain has the layout despite replaceable contents.Source: §5 · p. 8
  • Unsupervised chains store readable reasoning steps uniformly across slots rather than step by step; no ProsQA cell holds both candidate answers at once above the 30% bar.The highest both-readable rate in either probe arm is 0.113; answer-aligned chains hold little of the path.Source: §5 · p. 8
  • Across the results, unsupervised chains carry a finished answer, whereas per-step supervision lays the reasoning path out one step per slot.The probe targets include final-answer concepts, gold path steps on ProsQA, and answer-magnitude bins on GSM8k; probe access does not establish that the model uses the information to answer.Source: §5 · pp. 7–8

6 RQ3: WHAT IS TRANSMITTED ACROSS THE LATENT SLOTS, AND CAN A SETTLED STATE BE REVISED?

The study tracks state movement and tests whether mid-chain edits revise answers. On logical tasks, settling and continued movement differ by training recipe, while edit responses differ by backbone.

RQ3: What is transmitted across the latent slots?

  • Step size is cosine distance between consecutive states; the figure uses a shared color scale in which darker cells indicate larger steps.It displays each grid cell for GPT-2 and Llama-3.2-1B-Instruct using the layer-mean state.Source: §6 · p. 8, Figure 3 · p. 8

RQ3: Can a settled state be revised?

  • The edit test retains the original problem’s first i slots and uses a different-answer twin for the remaining computation.For recycling models, the driver splices in the twin’s recorded suffix; CODI recomputes later slots. The adaptive rate includes problems solved on the twin from scratch.Source: §6 · p. 8

RQ3: What is transmitted across the latent slots?

  • On logical tasks, unsupervised and answer-aligned chains settle early, while per-step chains keep moving to the final slot and move toward the answer only on ProsQA.In 11 of 15 cells with a usable answer direction, each step’s along-answer component is at or below its null; ProsQA is the sole per-step exception confirmed under both probe arms. The Llama-3.2-1B-Instruct per-step GSM8k twin also settles early.Source: Figure 3 · p. 8, §6 · p. 9

RQ3: Can a settled state be revised?

  • Mid-chain edit responses split by backbone: GPT-2 increasingly retains the original-prefix answer, whereas Llama-3.2-1B-Instruct follows the twin’s edited answer.Adaptive rate is the fraction of answers following the twin; the figure compares GPT-2 (solid) with Llama-3.2-1B-Instruct (dashed).Source: §6 · p. 9, Figure 4 · p. 9

RQ3: What is transmitted across the latent slots?

  • The authors conclude that recurrence is largely unused for serial computation: answer-aligned chains transmit a fixed question-time state, while per-step movement is answer-directed only on ProsQA.The edit battery measures answer response, not whether the model replans along post-edit trajectories.Source: §6 · pp. 8–9

7 DISCUSSION

The discussion combines chain-use interventions, content probes, and state dynamics into a taxonomy that distinguishes six measured chain profiles and leaves unresolved cases unclassified.

7.1 A Taxonomy of Latent Chains

  • The taxonomy assigns use categories from deletion and substitution, then profiles each category using content probes and state dynamics; readability does not override intervention results.Unresolved measurements remain unclassified.Source: §7.1 · p. 9
  • The six profiles distinguish inert carriers or empty chains, bare or laid-out placeholders, and content-used assembling or unstructured chains.Carriers hold a readable answer and settle early; laid-out placeholders have ordered steps and keep changing. “Assembling” does not establish that slots causally correspond to decoded steps.Source: §7.1 · p. 9
  • The grid includes all six taxonomy leaves: both placeholder types occur on Llama-3.2-1B-Instruct/ProsQA, while content-consulted chains are assembling on logical tasks and unstructured on GSM8k.Every profiled inert chain is either a carrier or empty.Source: §7.1 · p. 9
  • The taxonomy predicts intervention behavior beyond the measurements used to define its leaves: carriers resist corruption, placeholders resist even a different-answer donor, and only an assembling chain is helped by donor thoughts.The donor is applied at twice the natural norm.Source: §7.1 · pp. 9–10
  • Separating representation from use reconciles prior reports of placeholder behavior, collapsed superposition, and step-dependent causal roles as findings about different chain types.The paper maps the placeholder signature reported by Zhang et al. (2025), collapsed superposition reported by Rizvi-Martel et al. (2026), and the staged causal structure and non-local routing of Li et al. (2026) to distinct profiles.Source: §7.1 · p. 10
  • For agents using latent chains, readable traces would not certify the computation behind an action, which matters for auditing and security.The discussion names agent applications including acting, recommending, content optimization, and memory management.Source: §7.1 · p. 10

8 CONCLUSION

The conclusion separates reliance, represented content, and state movement as distinct properties of latent chains. It also discloses the paper’s use of generative AI tools in implementation, feedback, and drafting.

8 CONCLUSION

  • Latent-chain claims should specify whether they concern model reliance, chain content, or state movement, because these properties are separate.The conclusion states that chains can carry an answer already readable before they begin, and that per-step supervision keeps states changing on logical tasks, with ProsQA movement directed toward the answer.Source: §8 · p. 10

USE OF LARGE LANGUAGE MODELS

  • Generative AI tools were used to implement analysis code, provide experimental-design and interpretation feedback, and help draft the paper and its figures.The tools also helped draft the paper’s LaTeX.Source: USE OF LARGE LANGUAGE MODELS · p. 10

A EXPERIMENTAL DETAILS

The experimental grid reports its model and checkpoint coverage, then uses slot interventions, probes, and edits to examine latent-chain behavior. The two backbones are reported separately.

A.1 MODELS AND THE GRID

  • Table 2 summarizes each model’s slot inputs, supervision, slot budget, and grid coverage.GPT-2 and Llama-3.2-1B-Instruct results are reported separately.Source: Table 2 · p. 13, B RQ1 SUPPLEMENT: DELETION AND… · p. 14
  • The frozen ProsQA prefixes score above published numbers, while the original full split reproduces published accuracy within 0.002 (0.818 against 0.816).The passage attributes the difference to the split, not scoring, and says no within-paper contrast is affected.Source: §A.1 · p. 13

A.2 INTERVENTIONS, PROBES AND EDITS

  • Slot interventions replace the fed vector with norm-matched noise, a dataset-mean vector, or another problem’s vector, with downstream slots recomputed.Mean and donor vectors remain nearly parallel to natural slots (cosine 0.98–1.00); noise is near-orthogonal and matches magnitude only.Source: §A.2 · p. 14
  • Probes use PCA to 30 dimensions and cross-validated logistic regression, with a ProsQA ridge arm; edits splice twin-problem content into the chain.Probe verdicts require agreement between arms on ProsQA; edited problems are restricted to scratch-solvable cases and include answer-preserving controls.Source: §A.2 · p. 14

B.1 REPLICATION AND TRAINING-TIME CONTROLS

Replication and control checks support the reported deletion patterns, while truncation and substitution tests distinguish chain reliance from the effects of replacing its contents.

B.1 REPLICATION AND TRAINING-TIME CONTROLS

  • The second GPT-2 COCONUT training agrees with the primary cell within 0.010 on every condition; the Llama loss-off control is unresolved.The GPT-2 loss-off control is as inert as its continued-training cell.Source: Table 3 · p. 14, §B.1 · p. 15

B.2 TRUNCATION

  • On GPT-2, consulted cells improve monotonically as more slots are retained, and the last two GSM8k slots still carry 24%–32% of the deletion gap.On Llama-3.2-1B-Instruct, GSM8k curves saturate early, consistent with computation finishing rather than thinking hurting accuracy.Source: §B.2 · p. 15, Table 4 · p. 15

B.3 SUBSTITUTION: PER-CELL CONTRASTS

  • Across eighteen cells, donor replacement is worse than deletion in five, not applicable in seven, bounded in three, and better than deletion in three.In content-used GSM8k cells and on PrOntoQA, another problem’s thoughts cost more than none; on per-step ProsQA, donor thoughts beat having none.Source: §B.3 · p. 15, Table 5 · p. 15

B.4 THE DIRECTION-BY-SCALE CONTROL

  • On CODI/ProsQA, donor replacement changes zero answers despite a need for slots on the training manifold; on SIM-COT-COCONUT/PrOntoQA, mean thoughts are more damaging than random ones.For the per-step Llama GSM8k twin, four deltas fall within 0.023, so direction and scale cannot be separated.Source: §B.4 · p. 16, Table 6 · p. 16

C.1 THE ANSWER THROUGH DEPTH AND BY SLOT

The section examines what probes read across slots and layers, how states evolve, and whether mid-chain edits change answers. Its findings distinguish early readability, ongoing accumulation, and backbone-specific editing behavior.

C.1 THE ANSWER THROUGH DEPTH AND BY SLOT

  • 0.760 at the embedding layer, 0.063 at the mid-stack trough, and 0.753 at output show the answer is readable before and after the stack; the unsupervised chain changes little by slot.The supervised chain is unreadable at its first two slots and reaches 0.963 at the last.Source: Figure 6 · p. 16, §C.1 · p. 16

C.3 ACCUMULATION ON GSM8K

  • Answer accumulation appears in every GPT-2 recycling cell (first-to-last gains +0.23 to +0.29, p < 0.0001), and only at the readout’s resolution on the Llama per-step twin.No other Llama cell shows accumulation at that resolution; cross-backbone amounts are not compared.Source: §C.3 · p. 16

C.4 LAYOUT AND HELD-OUT CELLS

  • Two answer-aligned cells fail localisation under both probe arms, while their specificity control flips by arm; specificity is therefore unresolved, and “holds little” means little linear accessibility.The replication checkpoint agrees with the primary in both arms.Source: §C.4 · p. 16

C.5 ONE CANDIDATE AT A TIME

  • No cell reaches the pre-set 30% threshold for both ProsQA candidates being readable from the same slot; some cells are not applicable to that question.The probe uses a separate binary classifier for each candidate to avoid softmax-induced anti-correlation.Source: §C.5 · p. 17, Table 10 · p. 18

D.1 STATE-MACHINE SETTINGS; D.2 SETTLING

  • Settling is measured as the last slot where a discrete state label changes; Table 11 reports mean transition slot against a cross-problem null, with a 0.5-slot verdict floor.State labels use a shared random projection and joint clustering across latent and written states.Source: §D.1 · p. 18, Table 11 · p. 18

D.3 CONSECUTIVE-SLOT SIMILARITY

  • On logical tasks, unsupervised chains are nearly stationary while per-step chains keep moving to the last slot; the cosine-based settling criterion misfires and supports no body verdict.On GSM8k, movement differs by backbone: GPT-2 recipes keep moving, while Llama unsupervised and per-step chains settle from the first step.Source: §D.3 · pp. 18–19

D.4 ANSWER ALIGNMENT: FITTING POINT AND NULLS

  • At the last-slot fit, 18 of 69 cell-by-K verdicts move, and 32 of 69 differ between the label-permuted and cross-problem nulls.On SIM-COT-COCONUT/ProsQA, the along-component is 1.94 times its null under the probe of record and 1.23 under ridge.Source: §D.4 · p. 19

D.5 MID-CHAIN EDITS: COMMITMENT AND BACKBONE COMPARISON

  • Adaptation falls monotonically with injection position in every cell; under identical edits, Llama-3.2-1B-Instruct re-computes where GPT-2 holds or corrupts.Every gap in the paired backbone comparison clears its own minimum detectable effect; the recycling-family drift readout is degenerate, so the body reads answer-side rate A(i).Source: §D.5 · p. 19, Table 13 · p. 19, Table 14 · p. 20

E.1 PER-STEP SUPERVISION ON LLAMA-3.2-1B-INSTRUCT

The section reports Llama per-step supervision, a released ten-slot checkpoint, and the classification rules used to assign latent-chain profiles. It also marks limits on interpreting the backbone comparison.

E.1 PER-STEP SUPERVISION ON LLAMA-3.2-1B-INSTRUCT

  • The Llama per-step cells use ProsQA and PrOntoQA training with N=300; the ProsQA placeholder verdict bounds the harness, and the per-step backbone comparison is descriptive only.The paired sets have 53, 64, and 76 usable problems, too few for the Table 14 test; on these sets the Llama twin and control follow edits while the GPT-2 twin falls to 0.250.Source: §E.1 · p. 20, Table 15 · p. 20

E.2 THE RELEASED TEN-SLOT SIM-COT-COCONUT CHECKPOINT

  • Deleting the released ten-slot GSM8k chain flips 122 of 300 answers (C0 0.450, C1 0.047); donor thoughts are worse than deletion (p=0.00006).The checkpoint is kept out of the body because its slot budget differs, and its commitment window spans the whole chain.Source: §E.2 · p. 21

F.1 CLASSIFICATION RULES

  • Deletion separates inert, consulted, and unresolved cells; replacement separates consulted cells into placeholder, content-used, or unresolved, while readability and dynamics profile them.Unresolved measurements are not assigned a taxonomy leaf.Source: §F.1 · p. 21

F.2 CELL ASSIGNMENTS

  • The six leaves distinguish carriers, empty chains, two placeholder profiles, assembling chains, and unstructured content-used chains.Table 16 lists the twenty-one cells by deletion verdict, substitution verdict, profile, and resulting leaf.Source: §F.1 · p. 21, Table 16 · p. 22
Loading 2610.21182v1…