acceptodds
Under review as a conference paper at ICLR 2027

From Chains to DAGs: Probing the Graph Structure of Reasoning in LLMs

Abstract

Large language models often externalize reasoning as linear chains, but many reasoning problems can be more naturally modeled as directed acyclic graphs (DAGs), where intermediate conclusions branch, merge, and are reused. Whether such graph structure is reflected in model internals remains unclear. We introduce Reasoning DAG Probing to answer this question. We formalize each premise, intermediate conclusion, and final answer as a DAG node, and train lightweight probes to predict node depth, pairwise distance, and adjacency from LLM hidden states. Across logical, mathematical, and code reasoning benchmarks, we find that DAG structure is meaningfully encoded in LLM representations: it peaks in intermediate layers, with later layers preferentially recover deeper nodes and longer-range reasoning dependencies, and becomes stronger with model scale. DAG structure also progressively emerges during autoregressive generation in ways that anticipate answer correctness, while steering activations toward the ground-truth DAG structure can convert some originally incorrect generations into correct ones. These results suggest that LLM reasoning is not purely sequential, but reflects measurable internal graph structure.

Primary area:
interpretability and explainable AI
Keywords:
graph-structured reasoning, layerwise representation analysis, structual probes, causal intervention

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Ta9VOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Reasoning DAG Probing, which probes LLM hidden states for node depth, pairwise distance, and adjacency in externally constructed reasoning DAGs. Across several tasks and models, the authors report that DAG-related information is most recoverable in intermediate layers, increases with model scale, and tends to emerge during generation. They also show that activation steering can improve some initially incorrect generations.

Critical strengths
  1. The paper presents a clear and potentially useful framework for studying structured reasoning beyond linear chain-of-thought.
  2. The experiments cover multiple tasks, model scales, layers, and generation stages.
  3. The layerwise and generation-time analyses provide informative empirical patterns.
Critical weaknesses
  1. Recovering DAG labels from hidden states does not show that the model actually reasons according to the DAG. In ProofWriter, early nodes may systematically occur earlier in the input, while final conclusions occur later. Position, wording, or formatting cues could therefore predict depth without understanding dependencies. The authors should repeat the experiments after randomizing node order and controlling such superficial cues.

  2. Some DAG annotations, especially for GSM8K, are generated by an LLM rather than directly provided by the dataset. The paper lacks sufficient human alignment, inter-annotator agreement, or annotation-quality analysis to establish that these DAGs faithfully represent the intended reasoning process.

Questions for authors

Same as weakness

3 comments

Thank you for your constructive review and for recognizing the quality of our paper. Your comments helped us identify where the paper would benefit from clearer explanation and additional validation statistics.

W1: Recoverability does not imply actual use.

You are right that probing alone does not prove that the model uses the recovered DAG structure. This is why we include generation-time steering as a causal test of behavioral relevance in Section 3.7. In addition to Section 3.7, we added matched controls to isolate whether the effect is DAG-specific: random norm-matched steering and shuffled-DAG-label steering.

cutoffsnorm budget(s)no steerrandom norm-matchedshuffled DAG labelsgold DAG
3840.00147 (15/17/15)50 (20/16/14), +6.4%40 (12/14/14), -14.9%63 (21/19/23), +34.0%
384+10240.001,0.000550 (10/24/16), +6.4%39 (15/14/10), -17.0%71 (23/26/22), +51.1%
384+1024+20480.001,0.0005,0.00151 (12/24/15), +8.5%40 (12/14/14), -14.9%73 (24/25/24), +55.3%

Random norm-matched perturbations stay close to no-steer, suggesting same-sized activation changes are not sufficient to reliably improve nor damage generation. Shuffled-DAG-label steering is consistently worse than no-steer, indicating that corrupted structural targets can harm generation. Gold-DAG steering improves substantially, supporting the view that the recovered DAG structure is behaviorally relevant to the model's reasoning rather than only passively decodable.

W1: ProofWriter input order and superficial cues.

Your concern about superficial cues is reasonable, but the input-order cue is limited in our setting. The theory contains facts and rules, while sentences directly in the input are either distractors outside the gold DAG or peripheral premise nodes. Deeper intermediate conclusions and the final reasoning structure are not provided; they must be inferred by the model. Thus, sentence position in the theory does not directly reveal the node's reasoning depth. For wording and formatting cues, we include a lexical-only bag-of-words baseline. It remains substantially below the full contextual probe, suggesting shallow textual cues alone do not explain the recoverable DAG structure.

W2: GSM8K annotation quality.

Your concern about LLM-generated GSM8K DAG annotations is valid. Among 8,792 GSM8K datapoints, we randomly sampled 800 for manual validation. 744 annotations were directly correct; the remaining 56 differed from our validators' initial expected decomposition but were still self-coherent and correct reasoning DAGs for the solution. This suggests the GSM8K DAG annotations are reliable for our analysis, and we will report these validation statistics in the appendix.

In the uploaded PDF, we prioritized figure updates for visual inspection. All reported analyses, clarifications, and validation statistics will be incorporated into the final revision.

Follow-up on Input-Order Controls (W1).

While we previously argued that input-order cues are limited because intermediate conclusions and the final reasoning structure are not provided in the theory, we wanted to address your concern as directly as possible. We therefore added a randomized-order control: for each ProofWriter example, we shuffled complete theory sentences, recomputed hidden states from the shuffled input, retrained and re-evaluated the same set of probes. We report layers 24-28 because they are the strongest region for Qwen3-14B in the original setting.

layeroriginal DAG-F1randomized DAG-F1original depth spearmanrandomized depth spearmanoriginal distance spearmanrandomized distance spearmanoriginal adjacency F1randomized adjacency F1
240.6310.6240.8940.8780.6870.6790.6670.665
250.6380.6350.9050.8960.7010.6980.6680.669
260.6350.6270.9020.8850.6960.6810.6670.667
270.6370.6290.9040.8910.6990.6860.6670.665
280.6300.6220.9030.8870.6860.6760.6570.653

DAG-F1 and all three probe-specific metrics remain very similar after randomization. Given that the node-only, bag-of-words, and label-shuffled baselines reported in the paper are much lower, this suggests that the recovered structure is not explained by input sentence order.

Reviewer_Ta9VOpenReview ↗

Response to authors

I have no further concerns and will keep my score!

Reviewer_ZZr5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In the present paper, the authors introduce "Reasoning DAG Probing", a way to investigate whether LLM representations encode "correct" dependencies between reasoning steps. Using reference graphs derived from annotated proofs or worked solutions, they train probes to predict three structural properties: node depth, pairwise distance, and adjacency, with which they can reconstruct the model-inferred graph over the supplied nodes. The analysis considers different scenarios across logical, mathematical, and code tasks, examining recoverability across model layers and along generation progress. They also show that activation steering using reference-graph supervision can improve some initially unsuccessful continuations.

Critical strengths
  • I find it really compelling to ask whether model representations encode dependencies between reasoning steps beyond their linear textual presentation.

  • The framework provides a concrete way to measure reference-graph recoverability across model layers, generation progress, and model sizes, with experiments covering logic, mathematics, and code.

  • The generation-time association between graph recoverability and successful answers is interesting, particularly the increase observed before thinking models produce their final answers, and a generalization to problems that lack a ground truth would revolutionize training of reasoning models.

  • The activation-steering experiment goes beyond observational probing, providing promising evidence of a causal link between decoded structural information and accuracy.

Critical weaknesses
  • The paper addresses an interesting question, but the scope of the method is not made clear. Reconstruction seems to require supplied reference nodes, with graphs obtained from annotated proofs or worked solutions. This is different from discovering a reasoning graph for an unsolved problem, and this limits the broader interpretation and practical applicability of the results.

  • The probing setup does not clearly separate an internally constructed reasoning graph from a representation of supplied statements placed into context. Appended-node probing encodes reference statements through separate node-specific inputs, and includes correct intermediate conclusions. So, yes, there is decodable structural information, but the authors do not demonstrate that the reconstructed graph corresponds to the computation used by the model.

  • The interpretation of reconstruction quality remains limited. DAG-F1 does not establish whether the recovered graph supports a valid derivation, or how particular reconstruction errors relate to answer correctness. Moreover, matching a single reference graph can penalize alternative valid proofs with different intermediate steps or dependencies.

  • The evidence for depth-dependent emergence in Figure 3 is difficult to assess. The main depth probe is trained through ranking, while the figure reports absolute depth error without explaining calibration. Moreover, adjacency AP is independently normalized at each depth-gap, so absolute performance and the magnitude of improvements are hardly readable.

  • The generation-time analyses do not adequately distinguish structural refinement from increased coverage of reference nodes. Native-span scores can improve as more nodes are mentioned, while appended-node probes are retrained at each cutoff. The results do not establish that an initial internal graph is progressively corrected, or that DAG recoverability provides an early correctness signal.

  • Steering provides promising evidence of behavioral relevance, but the causal interpretation remains incomplete. The intervention uses the test example’s reference graph, and the absence of norm-matched random interventions, incorrect-graph targets, and component ablations makes it difficult to attribute the improvement specifically to better DAG organization. Its applicability without reference supervision is also unclear.

  • The proposed independence from particular verbalizations is not directly tested. Paraphrasing and scrambling reasoning order could help establish that the method recovers stable dependencies rather than properties of the selected textual realization.

  • In general, important implementation details are insufficiently specified, including native-span alignment, probe hyperparameters, treatment of missing nodes, and selection of layers and reconstruction thresholds.

Questions for authors
  • Can the authors provide a worked example showing the probe inputs, reference nodes, extracted token spans, and reconstructed graph? Does the “full theory” include the original query?

  • Can the evaluation accommodate more than one correct reasoning graph?

  • What does a given DAG-F1 imply about reasoning quality? Can the authors relate particular reconstruction errors to answer correctness?

  • Does Figure 3(a) use a regression probe or calibrated ranking scores? For Figure 3(b), please report unnormalized AP and uncertainty alongside the within-group normalization.

  • How are reference nodes aligned with generated text, particularly in cases where statements could be paraphrased, incomplete, or contradicted? How are missing nodes treated?

  • Can the authors distinguish refinement of previously represented dependencies from the appearance of additional nodes and later mentions?

  • How well does an early graph-based signal predict eventual correctness on held-out examples, beyond trace length, node coverage, and confidence? What signal could be used when no reference graph is available?

  • Are recovered dependencies stable under paraphrasing and reordering independent reasoning steps? Even in the controlled setting, could graph supervision improve a smaller model beyond matched token-level CoT supervision?

  • Could a similar approach be used in a masked diffusion setting, to probe the order in which the reasoning could take place?

5 comments
AuthorsOpenReview ↗

Response 1 of 2

Thank you very much for writing such a detailed review. Your feedback helped us substantially refine the paper, especially on separating structural refinement from node coverage, further isolating DAG-specific causal effects in steering, and relating particular reconstruction errors to answer correctness.

W1: Worked solution needed instead of discovering an unsolved problem.

You are right that our setup uses worked solutions rather than directly discovering reasoning graphs for unsolved problems. Still, we view this as a useful milestone toward open-ended graph discovery. First, showing DAG structure in LLM reasoning gives later interpretability work concrete internal signals to study, including recurring graph patterns and how graph-like reasoning might be steered without a reference graph. Second, our method can build supervised data aligning internal reasoning with reference DAGs, which may help train or evaluate more structured and interpretable reasoning.

W2: Appended nodes do not reflect native computation by the model.

We include native-span probing for this exact concern: instead of appending reference nodes, we probe node mentions in the model's own generated trace. This tests whether DAG structure is recoverable from native generation. We still include appended-node probing because it studies representations before decoding begins and ensures every reference node can be evaluated at every cutoff, reducing coverage confounds in generation-time comparisons.

W3/Q3: Interpretation of reconstruction quality.

Your point is valid: DAG-F1 is a structural diagnostic, not a certificate that the model followed a valid proof. Still, it is statistically informative: correct traces show sustained increases during decoding while incorrect traces remain comparatively flat. To relate specific reconstruction errors to correctness, we added test-set case studies in Appendix Figures 19 and 20 comparing the gold DAG, reconstructed native-span DAG, and generated CoT for two failed traces.

Figure 19 is an edge-error case: all gold proof nodes are present, but the reconstructed DAG misses dependency edges needed for the final conclusion. Figure 20 is a missing-node case: an intermediate gold node is absent, and the observed reconstructed graph misses the downstream proof step. In both cases, the graph error matches the CoT error: the model rejects the answer because it treats the missing dependency or intermediate fact as unavailable.

W3/Q2: More than one correct reasoning graph.

This question is meaningful. For constrained datasets such as ProofWriter and GSM8K, we did not observe substantially different but equally valid DAGs. We agree that open-ended tasks such as TACO can admit multiple valid DAGs. For evaluation, we use a deterministic reference DAG so the metric remains interpretable and comparable. Still, DAGs are less tied to surface form than CoT: one dependency graph can be verbalized through many chains of thought. Thus, a canonical DAG is already more invariant than matching a particular textual rationale.

W4/Q4: Objective misalignment and normalization in Figure 3.

Your point about Figure 3(a) is valid: MAE is not the most natural visualization for a depth probe trained with a ranking objective. We uploaded Figure 3(a'), a rank-based plot measuring whether nodes at each gold depth rank above shallower nodes, matching the depth probe's objective. We also uploaded Figure 3(a"), the same plot for an MSE-trained depth probe. Both show a similar diagonal banding pattern to Figure 3(a), though noisier.

For Figure 3(b), normalization was for visualization, making layerwise changes within each edge-span group easier to inspect. We will report unnormalized adjacency AP in the appendix and clarify the intended reading.

W5/Q6: Whether DAG-F1 reflects structural refinement or node coverage.

Your concern about changing observability is partly addressed by how DAG-F1 is computed. At each cutoff, DAG-F1 uses the induced gold subgraph over aligned/observable nodes: candidate pairs are only among observed nodes, and gold edges count only when both endpoints are observed. Missing future nodes are not false negatives. To further isolate refinement from coverage, we added a fixed-node evaluation: later cutoffs use only the node/edge set already observed at cutoff 256.

cutofforiginal DAG-F1fixed-node DAG-F1
2560.4040.404
3840.5170.455
5120.5420.472
7680.5830.502
10240.6040.517
15360.6140.523
20480.6160.524

Holding the early-observed graph fixed still shows improvement: from 256 to 2048 tokens, fixed-node DAG-F1 rises from 0.404 to 0.524. Thus, sharpening already-visible DAG structure is part of the pattern. Since this increase is smaller than the original DAG-F1 increase, the original DAG-F1 trend likely reflects both better reconstruction of observed structure and new nodes/edges becoming internally observable and organized.

AuthorsOpenReview ↗

Response 2 of 2

W6: Further isolation of steering causal mechanism.

To test whether steering gains come from DAG-specific information, not generic perturbation, we added two matched controls: random norm-matched steering, and shuffled-DAG-label steering, keeping the same pipeline but corrupting target DAG labels.

cutoffsnorm budget(s)no steerrandom norm-matchedshuffled DAG labelsgold DAG
3840.00147 (15/17/15)50 (20/16/14), +6.4%40 (12/14/14), -14.9%63 (21/19/23), +34.0%
384+10240.001,0.000550 (10/24/16), +6.4%39 (15/14/10), -17.0%71 (23/26/22), +51.1%
384+1024+20480.001,0.0005,0.00151 (12/24/15), +8.5%40 (12/14/14), -14.9%73 (24/25/24), +55.3%

Random norm-matched perturbations stay close to no-steer, suggesting same-sized random activation changes are not sufficient to reliably improve nor damage generation. Larger inter-seed fluctuations are expected when random directions help, hurt, or do little. Shuffled-DAG-label steering is consistently worse than no-steer, showing corrupted structural targets make the objective harmful. Unlike gold-DAG steering, the effect does not grow as later cutoffs are added, suggesting repeated interventions accumulate productively only when aligned with the correct DAG.

W6/Q7: Applicability without reference graph.

We are exploring steering without reference supervision, but this paper aims to establish that internal DAG structure is causally relevant during reasoning. We do not present the current method as a test-time strategy for unknown graphs. Instead, reference-supervised steering is a causal-analysis tool for testing whether changing probed DAG structure affects generation.

W7/Q8: Paraphrasing and reordering independent reasoning steps.

We see two interpretations of this concern. If the concern is that statement order in input theory could leak a node's position in the DAG, this cue is limited: inputs contain peripheral facts and rules, while intermediate nodes and final structure are not included and must be inferred. If the concern is reordering generated reasoning, artificially permuting generated steps is not our target; we analyze traces as naturally generated. Stochastic native-span generations partially address both reordering and paraphrasing: the same problem can be reasoned through with different wording/order across runs, yet dependencies remain recoverable. Also, our lexical-only bag-of-words baseline remains well below the contextual probe, suggesting robustness beyond a single textual rationale.

W8/Q5: Implementation details.

We added key implementation details in Section 2 and Appendix Sections J/K. For native-span alignment, ProofWriter node statements use a simple structured format such as "A is B", so spans are aligned by exact/normalized matching. On 948 Qwen3-4B test traces, we obtain 41,224 aligned spans, and 93.6% of gold nodes are aligned by 4096 tokens. Missing nodes are not false negatives in DAG-F1, as noted in W5/Q6. Layer selection and thresholds are fixed on development data before test evaluation. We will expand details in the final revision for reproducibility.

Q1: A worked example of probing.

Appendix Section C gives theory/query examples. In appended-node probing, the input is the theory plus one reference node, e.g., [full theory] Harry is furry. Other reasoning nodes are not included, so the input jumps from theory to target node. The full theory excludes the original query. In generation-time analysis, the query is included to prompt the model to answer, using Figure 12's template. Reconstructed DAG examples are in Figures 19 and 20.

Q7: Early DAG signals for predicting final correctness.

We ran held-out logistic regression on 948 Qwen3-4B ProofWriter test traces with stratified 5-fold CV to predict final correctness. At each cutoff, trace length is fixed by design, and the control uses aligned/observed node coverage. Adding native-span DAG-F1 improves held-out AUC:

cutoffcoverage AUCcoverage + DAG-F1 AUCdelta
2560.4950.753+0.258
3840.5900.798+0.208
5120.6540.831+0.177
7680.6750.837+0.162
10240.6590.816+0.157

Thus, early graph recoverability predicts correctness beyond how many reference nodes appear, with larger gains at earlier cutoffs.

Q9: Probe reasoning order in masked diffusion setting.

This is an interesting question. We are not experts on masked diffusion models, so we can only offer a tentative view. The framework should apply in principle, but the temporal axis would change: instead of autoregressive token positions, one could probe denoising/refinement steps. This could reveal whether nodes closer to the answer become recoverable earlier/later, and whether dependency edges sharpen as masks are filled. This is a promising future direction.

In the uploaded PDF, we prioritized figure updates for visual inspection. All reported analyses, clarifications, and validation statistics will be incorporated into the final revision.

Reviewer_ZZr5OpenReview ↗

Response to rebuttal

I warmly thank the authors for the effort they put into addressing all the points raised in my review. I think some of the additions and clarifications do improve their manuscript.

I wish it were easier to understand, at least intuitively, what the direction for generalizing their approach could look like, especially to unsupervised cases (no ground-truth reasoning graphs) and to degenerate cases (multiple possible reasoning DAGs). I understand the current work is building some foundation for this research, but it remains slightly unclear to me whether it is truly feasible to expand further with this approach.

Given how important this kind of analysis and steering tool would be if it were general, and given the thorough rebuttal, I will still consider increasing my score at the end of the discussions.

Reviewer_pSCMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper constructs reasoning DAGs for ProofWriter, GSM8K, and TACO and probes frozen LLM representations for node depth, pairwise longest-path distance, and adjacency. Predictions are combined to reconstruct directed graphs. Static probes encode the theory with an appended canonical node statement; generation-time analyses additionally examine native node mentions. The paper reports an intermediate-layer recovery peak, differences by node depth and model scale, and associations between graph recovery and answer correctness. Finally, gold-DAG-guided activation steering on Qwen3-4B increases successful continuations of initially incorrect ProofWriter generations from 47/363 to as many as 73/363 across three replay seeds.

Critical strengths
  1. The relational probing target is informative. Studying adjacency and branching dependencies alongside node depth is a useful extension beyond classifying isolated reasoning variables. ProofWriter provides explicit proof structure, and node-only, lexical, shuffled-label, and alternative depth-objective controls help establish that contextual representations carry graph-related information.

  2. The work goes beyond static decodability. Native-span probing and continuation interventions are worthwhile attempts to connect representational measurements to behavior. Appendix K explicitly rebuilds downstream states and the KV cache after intervention, and it selects intervention hyperparameters on development data. The paired-prefix design and separate reporting for initially correct and incorrect examples make the steering result more interpretable.

Critical weaknesses
  1. The evidence does not yet isolate a graph-specific causal mechanism. Static probing supplies canonical node text and the theory, so successful recovery establishes conditional decodability rather than spontaneous graph construction. Native-span probing improves this connection, but the gold DAG determines both alignment and steering targets. The intervention comparison is against no steering; there are no reported norm-matched random directions, shuffled-DAG targets, or equally informative non-graph supervision controls. Gold-informed perturbations could improve continuation through generic task information rather than the particular dependency structure. Thus, the results support behaviorally useful graph-related signals, but the stronger interpretation that reasoning internally operates by progressively constructing these graphs remains underdetermined.

  2. Some central layerwise/generation-time interpretations require better-controlled measurements. The main depth probe uses a ranking loss, which constrains ordering rather than absolute depth. Figure 3(a) nevertheless uses absolute MAE to argue that deeper layers preferentially recover deeper nodes; the text does not clarify whether this plot uses a separately calibrated or regression-trained probe. Changes in output scale or offset can change this heatmap without changing ranking quality. Native-span coverage also changes with generation length, and Appendix J trains separate appended-node probes at each cutoff. Fixed-node coverage analyses, explicit alignment accuracy, and calibrated measurements are needed to separate changing readout/coverage from growth of internal structure.

Questions for authors
  1. Causal-control request: Do the continuation gains exceed norm-matched random perturbations, graph-label shuffling, and a non-graph task-informed steering control? Please use the same held-out examples and report paired, example-clustered uncertainty across replay seeds.
  2. Measurement clarification: Which training objective and calibration produce Figure 3(a)'s MAE? Does the depth-by-layer pattern persist with regression probes or a development-only calibration that is fixed before testing?
  3. Generation-time clarification: How does DAG-F1 change when evaluated on a fixed set of aligned nodes and edges across cutoffs? Please specify alignment quality, missing-node treatment, and whether layer selection is fixed on development data. These details determine whether increasing F1 reflects structural emergence or improving observability.
2 comments

We appreciate your constructive feedback, especially on steering controls and the mismatch between training objective and visualization, which helped us refine the paper.

W1/Q1: Further isolation of graph-specific causal mechanism.

We added two matched controls: random norm-matched steering, applying a random update with the same norm budget, and shuffled-DAG-label steering, keeping the same pipeline but corrupting the target DAG labels. All use the same 121 examples, 3 replay seeds, layer 21, cutoffs, and budgets as gold-DAG runs.

cutoffsnorm budget(s)no steerrandom norm-matchedshuffled DAG labelsgold DAG
3840.00147 (15/17/15)50 (20/16/14), +6.4%40 (12/14/14), -14.9%63 (21/19/23), +34.0%
384+10240.001,0.000550 (10/24/16), +6.4%39 (15/14/10), -17.0%71 (23/26/22), +51.1%
384+1024+20480.001,0.0005,0.00151 (12/24/15), +8.5%40 (12/14/14), -14.9%73 (24/25/24), +55.3%

Random norm-matched perturbations stay close to no-steer, suggesting same-sized activation changes are not sufficient to reliably improve nor damage generation. Their larger inter-seed fluctuations are expected when random directions help, hurt, or do little. Shuffled-DAG-label steering is consistently worse than no-steer, indicating the objective becomes harmful with corrupted structural targets. Unlike gold-DAG steering, it does not improve as later cutoffs are added, suggesting repeated interventions accumulate productively only when aligned with the correct DAG structure.

W2/Q2: Rank-based probe training vs. reported MAE in Figure 3(a).

Your point is valid: MAE is not the best visualization for a depth probe trained with a ranking objective, since ranking constrains relative order, not calibrated numeric depth. We uploaded rank-based depth-recoverability plots of the same ordering optimized by the depth-ranking probe, measuring whether nodes at each gold depth are ranked above shallower nodes. We also uploaded one for an MSE-trained depth regression probe. They are inserted below Figure 3(a) as Figure 3(a') and Figure 3(a"), and exclude depth 0 because there are no shallower nodes to compare against.

Both added plots show a similar diagonal banding pattern to Figure 3(a). Because the added metric is a score rather than a loss, high-recoverability bands appear white rather than dark. The bands are noisier than in Figure 3(a), but the same depth-by-layer trend is still clear, suggesting deeper layers preferentially improve the relative ordering of deeper reasoning nodes under both rank-trained and regression-trained depth probes.

The current Figure 3(a)/(a')/(a") layout is temporary for you to view the update. In the final revision, we will reorganize the plots cleanly and ensure the ranking-objective versus MAE-visualization mismatch is resolved.

W2/Q3: Whether DAG-F1 reflects structural emergence or node coverage.

Your concern about changing observability is partly addressed by how DAG-F1 is computed. At each cutoff, DAG-F1 uses the induced gold subgraph over aligned/observable nodes: candidate pairs use only observed nodes, and gold edges are included only when both endpoints are observed. Thus, early cutoffs are not penalized for missing future nodes. For the remaining concern that the evaluated node/edge set changes as nodes become observable, we added a fixed-node evaluation: later cutoffs are evaluated only on the node/edge set already observed at an initial cutoff 256.

cutofforiginal DAG-F1fixed-node DAG-F1
2560.4040.404
3840.5170.455
5120.5420.472
7680.5830.502
10240.6040.517
15360.6140.523
20480.6160.524

Holding the early-observed graph fixed still shows steady improvement as the trace grows: from 256 to 2048 tokens, fixed-node DAG-F1 increases from 0.404 to 0.524. Thus, sharpening of already-visible DAG structure is part of the progressive emergence pattern. At the same time, the fixed-node increase is smaller than the original DAG-F1 increase. We therefore clarify that our claim was not that generation-time DAG emergence comes only from sharpening already-visible nodes. Instead, it reflects a combination of effects: existing observed structure can become better reconstructed, and additional reasoning nodes and edges can become internally observable and structurally organized as the trace unfolds.

Q3: Node alignment quality and probing layer selection.

In ProofWriter, node statements are given in a simple structured format; all nodes are "triples" such as "A is B". So alignment is done with exact/normalized match with high quality. On 948 Qwen3-4B test traces, we obtain 41,224 aligned spans, and 93.6% of gold nodes are aligned by 4096 tokens. Layer and threshold selection are fixed on development data before applying to the test evaluation.

In the uploaded PDF, we prioritized figure updates for visual inspection. All reported analyses, clarifications, and validation statistics will be incorporated into the final revision.

Dear Reviewer pSCM,

As we approach the end of the discussion period, we would like to kindly ask whether you have any further feedback on our response, so that we can best resolve your questions and address any remaining concerns.

In summary, we addressed four main points: (1) we added matched steering controls, including random norm-matched steering and shuffled-DAG-label steering, to better isolate the DAG-specific causal effect; (2) we added rank-compatible visualizations for both ranking-trained and regression-trained probes, addressing the mismatch between the ranking-based objective and the original MAE plot; (3) we clarified the missing-node treatment in the computation of DAG-F1, and added a fixed-node analysis to separate structural refinement from increasing node coverage; and (4) we clarified native-span alignment quality and development-set layer/threshold selection.

We would be grateful to know whether these additions resolve your main concerns, or whether there are specific points you would still like us to clarify. Thank you again for your constructive feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Discussion Recap

We thank all reviewers for their constructive feedback and thank the area chairs for coordinating the discussion. We have uploaded a revised version of the paper, with all updates marked in blue. We incorporated the main analyses, clarifications, and validation statistics from the rebuttal into this revision. The major concerns we addressed are summarized below:

  1. Whether steering effects are DAG-specific. We added random norm-matched and shuffled-DAG-label steering controls. Random perturbations stay close to no-steer, while shuffled DAG labels hurt performance.

  2. Whether Figure 3(a) mismatched the probe objective. We added rank-compatible depth-recoverability plots for both the rank-trained probe and a regression-trained probe, showing similar depth-by-layer banding patterns.

  3. Whether generation-time DAG-F1 mainly reflects node coverage. We clarified that DAG-F1 is computed over aligned/observable nodes and added a fixed-node analysis showing that DAG-F1 still improves when the evaluated node/edge set is held fixed.

  4. Whether reconstruction errors relate to answer correctness. We added case studies comparing gold DAGs, reconstructed DAGs, and generated CoTs, showing missing-edge and missing-node errors in reconstructed DAGs that align with local reasoning failures.

  5. Whether early DAG signals predict correctness beyond coverage. We added a held-out logistic-regression analysis showing that native-span DAG-F1 improves correctness prediction over observed-node coverage alone, especially at early cutoffs.

  6. Whether superficial input-order cues explain recoverability. We added a randomized input-order control on ProofWriter; DAG-F1 and probe-specific metrics remain very similar after shuffling complete theory sentences.

  7. Whether GSM8K annotations are reliable. We added manual validation statistics for 800 sampled GSM8K annotations: 744 were directly correct, and the remaining 56 were self-coherent valid DAG formalizations of the worked solution.

  8. Implementation and evaluation clarity. We added details on native-span alignment quality, missing-node treatment, shared native-span probes across token positions, development-set selection, and steering implementation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.