acceptodds
Under review as a conference paper at ICLR 2027

When Representations Collapse in Logical Reasoning: Diagnosing and Mitigating Cascade Failures in LLMs

Abstract

Large language models (LLMs) achieve strong performance on logical reasoning benchmarks, yet remain brittle as proofs become deeper, irrelevant premises accumulate, or evaluation conditions shift. We investigate this brittleness at the level of internal representations. To this end, we introduce Inferential Information Flow (IIF), a reasoning-aware diagnostic framework that measures, layer by layer, whether a model preserves selectivity toward proof-relevant evidence over distractors. Across three model families and three logical reasoning benchmarks, IIF reveals a broad but heterogeneous failure pattern: proof-selective structure often strengthens at intermediate depth and subsequently degrades, while some settings instead fail to form strong selectivity at the query interface. We term the former cascade IIF collapse. Causal interventions further show that some degraded value-channel representations can become harmful to downstream reasoning rather than merely uninformative. Motivated by this diagnosis, we introduce FRGI (Fact-Rule Grouped Intervention), a training-free, proof-free inference-time method that uses query-conditioned saliency to select candidate fact and rule spans and an Unlabeled Collapse Proxy (UCP) to locate intervention layers. FRGI applies localized fact-rule and value-channel restoration without retraining, proof annotations, or prompt modification. Across ProverQA, PrOntoQA, and ProofWriter, FRGI consistently improves accuracy over representative prompting, self-consistency, symbolic-reasoning, and activation-intervention baselines. These results show that representation-level failures in logical reasoning can be diagnosed and partially repaired through targeted inference-time intervention.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Deep Learning, Logical Reasoning, LLM, Collapse

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_k9G8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed Inferential Information Flow, a reasoning-aware diagnostic framework to mesures whether a large language model preserves selectivity toward proof-relevant evidence during the reasoning process. It demonstrated that the selectivity are emphasized in the middle stage but degraded afterward. With this insight, the authors proposed FRGI, i.e., factor-rule grouped intervention method to select evidences and rules adaptively follows a 'what-where-how' design. Experimental results on three typical QA datasets under three LLMs demonstrates that with the assistance of FRGI, the reasoning accuracy is substantially improved.

Critical strengths

This paper is well organized and well motivated. The proposed FRGI which does not require re-training and proof annotations is interesting.

Critical weaknesses
  1. Are the five indicators, i.e., NCE-IIF, CLUB-stype probe, ValueIIF, ICA, and TripletAcc, independent and truly complementary? Would they be biased? They are proxy diagnostics inspired by mutual information. The direct information flow and the actual reasoning process are still uncertain.

  2. The above five indicators can demonstrates the similarity, ranking of the proof-relevant information. However, representations collapse cannot be truly proved by merely these indicators.

  3. It is suggested to add ablation studies on the reasoning process with or without FRGI and test whether the IIF and its five indicators are truly improved through FRGI.

  4. Experiments on only three logical reasoning benchmarks is limited. More experiments on wider range of reasoning benchmarks should be given.

Questions for authors

How do the authors validate the IIF truly capture reasoning-relevant information loss and the proposed FRGI improves both these indicators and reasoning performance?

1 comment
AuthorsOpenReview ↗

Response to Reviewer

Thank you for asking questions. We address the four concerns below and report the paired intervention results in Q1.

W1. Complementarity and interpretation of the diagnostics.

The five diagnostics capture different aspects of proof-evidence selectivity using a shared set of proof-span annotations. NCE-IIF measures how much contrastive score mass is assigned to proof spans. The CLUB-style probe measures how well similarity scores predict proof-span membership. ValueIIF measures selectivity in the value channel, ICA measures alignment with the centroids of proof spans and distractors, and TripletAcc measures how often a proof span ranks above a distractor.

Figure 8 shows that agreement among these diagnostics varies across backbones. On Mistral-7B, pairwise correlations across layers are at least 0.71. On Qwen3-8B, the correlation between ValueIIF and ICA is 0.06. The diagnostics thus provide complementary views of the representations within a shared evaluation setup. Their scores are interpreted relative to the annotated proof. When multiple proofs are valid, a span outside the annotated proof may still support another derivation.

W2. What the layerwise decline establishes.

We interpret the post-peak decline as reduced separation between annotated proof spans and distractors in the measured representations. Appendix C.4 shows this decline in raw ICA and TripletAcc profiles before aggregation, and Appendix D.5 examines it directly in hidden-state and value space. These results support a claim about evidence selectivity at the measured layers. The model's actual use of that evidence during deduction is a separate question.

W3. Ablation with and without FRGI.

We recomputed all five diagnostics on the same ProverQA examples under vanilla inference and full FRGI, using the same proof-span annotations and readout layers. Q1 reports the paired comparison alongside task accuracy.

W4. Benchmark coverage and transfer to FOLIO.

The main FRGI accuracy evaluation covers ProverQA, PrOntoQA, and ProofWriter. Appendix D.5 extends the raw evidence-separation analysis to JustLogic and LogicSkills, bringing the diagnostic evaluation to five benchmarks.

To evaluate FRGI on an additional reasoning benchmark, we ran Qwen3-8B on the combined FOLIO training and validation splits (N=1,208N=1{,}208) using greedy decoding. We used each premise as a candidate evidence span and assigned fact and rule roles directly from the premise text.

MethodCorrect / NNFOLIO accuracy (%)
Vanilla630 / 1,20852.15
Zero-shot CoT669 / 1,20855.38
ZEROSUP (gate bypassed, diagnostic control)605 / 1,20850.08
Full FRGI682 / 1,20856.46

Full FRGI reaches 56.46% accuracy, improving over vanilla inference by 4.31 percentage points and zero-shot CoT by 1.08 points. The cascade and value hooks apply updates on 28.48% and 25.17% of inputs, respectively.

Q1. How do we validate what the IIF diagnostics capture, and does FRGI improve them alongside task accuracy?

We use annotated proof spans and distractors to test whether the representations distinguish evidence relevant to the annotated deduction. To assess whether FRGI improves that distinction, we evaluate Qwen3-8B on the same ProverQA examples with and without FRGI under beam search (k=4k=4). The table reports diagnostic scores averaged over layers 24 and 28. Gold proof-span annotations are used only for evaluation.

MethodNCE-IIFCLUB-styleValueIIFICATripletAccAccuracy (%)
Vanilla0.1820.2240.1510.1420.54338.70
Full FRGI0.4730.5210.4580.4150.82465.60

Full FRGI improves all five diagnostics at the measured readout layers. For example, TripletAcc rises from 0.543 to 0.824 on a 0–1 scale, indicating that proof spans rank above distractors more often. Task accuracy rises from 38.70% to 65.60% on the same examples. In this setting, the accuracy gain is accompanied by improved selectivity for annotated proof evidence across all five diagnostics.

Thank you again for your feedback. We hope the new ablation and FOLIO results address your questions about what the diagnostics capture and how FRGI transfers. We welcome any follow-up questions and would appreciate your reconsidering the rating.

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

This paper introduces Inferential Information Flow (IIF) to analyze how LLMs preserve proof-relevant evidence across Transformer layers. Based on observed degradation in evidence selectivity, the authors propose FRGI, a training-free inference-time intervention method. Experiments across three logical reasoning benchmarks and multiple model families demonstrate consistent accuracy improvements.

Critical strengths

1.The paper provides useful insights into how proof-relevant information evolves across Transformer layers. 2. FRGI offers a practical inference-time intervention strategy, with consistent improvements across different models and benchmarks.

Critical weaknesses
  1. Under Eq. (2), the conditional mutual information becomes layer-invariant when query–span similarities are distinct. Therefore, the observed changes in proxy metrics do not directly establish the claimed IIF collapse.

  2. While the components improve accuracy, the paper does not directly demonstrate that FRGI restores degraded proof-relevant information. The substantial gains from simple activation suppression also leave open whether the improvements arise from collapse-specific repair or general activation-editing effects.

Questions for authors
  1. How do the authors reconcile the formal IIF definition in Eq. (2) with the observed layerwise collapse?
  2. Can the authors provide direct evidence that FRGI restores degraded proof-relevant representations, rather than merely improving prediction accuracy?
  3. Can the authors clarify whether the performance gains are specifically attributable to collapse repair, given the substantial improvements from simple activation suppression?
3 comments
AuthorsOpenReview ↗

Response to Reviewer

Thank you for raising these questions about the definition of IIF and the effect of FRGI on proof-evidence selectivity.

Q1 (W1). How does Eq. (2) relate to the observed layerwise decline?

Eq. (2) defines IIF as I(G(ℓ);Y∣X)I(G^{(\ell)};Y\mid X), where G(ℓ)G^{(\ell)} is the query–span similarity and YY is the proof-usage label. If similarity scores are distinct within each instance, the score identifies the span and its label, so I(G(ℓ);Y∣X)=H(Y∣X)I(G^{(\ell)};Y\mid X)=H(Y\mid X). This quantity can remain constant across layers even when the evidence ranking changes.

Our empirical analysis examines how well the representations separate proof spans from distractors. For example, TripletAcc decreases when distractors overtake proof spans in the similarity ranking, even if the scores remain distinct. This explains how the observed post-peak decline in selectivity can coexist with a layer-invariant value in Eq. (2). We will clarify in §2 that Eq. (2) gives the formal definition of IIF, while the reported collapse refers to the post-peak decline in measured proof-evidence selectivity.

Q2 (W2). Does FRGI recover measured proof-evidence selectivity?

We evaluated Qwen3-8B on the same ProverQA examples across all variants using beam search (k=4k=4). We recomputed the five diagnostics and raw evidence separation δ\delta at layers 24 and 28, both after the layer-10 peak in raw hidden-state separation reported in Appendix D.5. Each diagnostic is averaged over the two readout layers. Gold proof-span annotations were used only for evaluation.

MethodNCE-IIFCLUB-styleValueIIFICATripletAccRaw δ\deltaAccuracy (%)
Vanilla0.1820.2240.1510.1420.5430.11038.70
ZEROSUP0.3150.3680.2840.2560.6850.22861.20
Matched attenuation + cascade0.3410.3920.3150.2730.7120.25362.15
Full FRGI0.4730.5210.4580.4150.8240.38565.60

Full FRGI improves all five diagnostics averaged over the two measured layers. TripletAcc rises from 0.543 to 0.824, indicating that annotated proof spans rank above distractors more often. Raw evidence separation rises from 0.110 to 0.385, alongside an accuracy increase from 38.70% to 65.60% on the same examples.

Q3 (W2). What does the reference update contribute beyond suppression?

ZEROSUP accounts for a substantial share of the improvement, raising accuracy from 38.70% to 61.20% and raw evidence separation from 0.110 to 0.228. To measure what the reference adds, we compare FRGI with a control that uses the same selected spans, target layers, cascade hook, and attenuation schedule βℓ\beta_\ell. Both variants retain (1−βℓ)vi(ℓ)(1-\beta_\ell)v_i^{(\ell)}, and FRGI additionally injects βℓv~i(ℓ)\beta_\ell\tilde v_i^{(\ell)}.

We repeated this comparison at a contiguous block outside the UCP-selected region, chosen in advance using a vanilla profile on a held-out calibration split. The two configurations use the same evaluation examples, selected spans, cascade hook, number of value hooks, and corresponding attenuation coefficients. We measure δ\delta at layers 24 and 28 in both configurations.

Value-hook locationMatched attenuation (Acc. / δ\delta)With reference (Acc. / δ\delta)Accuracy change (pp)Change in δ\delta
UCP-selected layers62.15% / 0.25365.60% / 0.385+3.45+0.132
Control layer block57.85% / 0.21055.24% / 0.182−2.61−0.028

Adding the reference improves both accuracy and post-peak proof-evidence selectivity at the UCP-selected layers, while both measures decrease at the tested control block. This reversal shows that the effect of the reference update depends on its placement in the configurations we tested.

Thank you again for your feedback. We welcome any follow-up questions and would appreciate your reconsidering the rating in light of these clarifications and results.

Thank you for the clarifications and additional experiments, which address some of my concerns. While certain conceptual limitations remain, the response strengthens the paper. I therefore maintain my original rating.

Thank you for your follow-up and for considering our clarifications and additional experiments. We appreciate your feedback and will reflect the remaining conceptual concerns in the manuscript.

Reviewer_r3C3OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper first introduces Inferential Information Flow (IIF), a layerwise diagnostic of whether LLM representations keep proof-relevant facts and rules distinguishable from distractors. It is approximated with five complementary metrics, not exact mutual information. On LLaMA-3.1-8B, Mistral-7B, and Qwen3-8B, IIF often rises to a mid-depth peak and then degrades (“cascade IIF collapse”), most clearly on ProverQA and PrOntoQA. On ProofWriter, query-level selectivity often never forms, though rule-level signals still decline after a peak. Based on this, the authors propose FRGI, a training-free, proof-free inference-time method. It selects fact and rule spans via attention saliency, picks layers with an unlabeled collapse proxy (UCP), and partially restores value-channel representations during prefill. FRGI beats prompting, self-consistency, symbolic, and activation-steering baselines in all nine model–benchmark settings, at near-vanilla latency. Ablations show that span targeting, value restoration, and UCP layer selection each matter, and UCP comes within about 2 points of a proof-annotated oracle. Gains shrink with scale: about +27 points at 8B, but only about +5 at 14B and 32B.

Critical strengths

Firstly, this paper identifies an interesting specific phenomenon called 'cascade IIF collapse'. Selectivity between the final query token and each evidence span strengthens up to an intermediate layer, then degrades toward the output. The paper quantifies this as the relative drop from the peak to the mean of the last three layers. For example, on ProverQA the drop is about 60–63% for LLaMA-3.1-8B, Mistral-7B, and Qwen3-8B, which peak at L14, L15, and L19. The implication is that an LLM forms proof-relevant structure. However,it does not preserve it to the final layers, which causes the wrong answer despite valid intermediate reasoning conclusions.

Second, this paper proposes a training-free method for repairing cascade IIF collapse at inference time. It does not change model weights, the prompt, or the architecture. It is proof-free at deployment, so no proof annotations or answer labels are used. This method is suitable for practical inference-time steering of the LLM decoding process.

Critical weaknesses

The central claim of this paper is that the proposed FRGI works by repairing cascade IIF collapse of LLM reasoning. However, in Table 17, zeroing the selected value vectors brings LLaMA’s average from 44.20% (vanilla) to 70.20%. Full FRGI reaches 75.52%. So it looks like most of the gain comes from removing the content of the selected spans, not from restoring a healthy reference. Furthermore, in all four reported case studies (two LLaMA, two Mistral), I notice that UCP selects l*=8 and l_c=3. For these 32-layer models, these are the lower bounds of the search ranges, according to Eq. 25 and Eq. 27 in the appendix. This suggests the selection may always sit at the boundaries of the window, instead of an adaptive selection based on the UCP score. Could it be caused by the UCP score simply keeping rising toward earlier layers? In that case, UCP behaves like a fixed rule (for example, “always use layer 8”) rather than locating the collapse for each model and instance. In this sense, these empirical observations deserve further elaboration.

According to the FRGI's design, this method requires the input to contain identifiable fact and rule spans. Its Unlabeled Collapse Proxy (UCP) also relies on signals such as rule-state attenuation and rule–fact coupling. So this method heavily relies on structured spans, such as clear rule–fact organization. This may limit the applicability of the proposed method in reasoning tasks. For example, could the proposed method be applied to other logic reasoning datasets like FOLIO ?

Questions for authors

In Table 17, could you elaborate on the contribution of using FRGI in improving accuracy? Furthermore, could you provide the distribution of the UCP scores across layers so we have a better idea of why the 4 LLMs choose l*=8 (which is the lower bound of the layer search ranges)?

How would the proposed FRGI method behave over other logical reasoning data like FOLIO?

4 comments
AuthorsOpenReview ↗

Response to Reviewer #1

Thank you for the careful review and for raising these questions. We address them below in order.

Q1 (W1). What does the reference-based update add beyond zero suppression?

We agree that zero suppression accounts for much of the accuracy gain in Table 17. Under beam search (k=4k=4), zeroing the selected value vectors raises the mean accuracy of LLaMA-3.1-8B across ProverQA, PrOntoQA, and ProofWriter from 44.20% to 70.20%. Full FRGI reaches 75.52%, a further gain of 5.32 percentage points. Full FRGI improves over ZEROSUP in all nine model–benchmark pairs.

To quantify what the reference contributes at the same level of attenuation, we evaluated Qwen3-8B on ProverQA under beam search (k=4k=4), using the same evaluation set for all variants. We added a matched attenuation + cascade control that retains FRGI's early cascade hook (Appendix E.3.1, Eq. 31) and uses the same selected spans, target layers, and βℓ\beta_\ell schedule. At each selected token position ii in target layer ℓ\ell, both variants retain (1−βℓ)vi(ℓ)(1-\beta_\ell)v_i^{(\ell)}, where vi(ℓ)v_i^{(\ell)} is the current value vector. Full FRGI also adds the reference term βℓv~i(ℓ)\beta_\ell\tilde v_i^{(\ell)}. The reference vector v~i(ℓ)\tilde v_i^{(\ell)} is obtained by projecting the same token's hidden state from the reference layer into the target layer's value space (Appendix E.3.2–E.3.3, Eqs. 33–34).

We measured raw hidden-state separation δ\delta between gold proof spans and distractors at layers 24 and 28 (both after the layer-10 peak reported for Qwen3-8B on ProverQA in Appendix D.5, Fig. 14 and Table 10). We used the same two readout layers for all variants. Gold proof-span annotations were used only to compute this diagnostic.

MethodUpdate at selected value positionsAccuracy (%)δ\delta at layer 24δ\delta at layer 28Mean δ\delta
Vanillavi(ℓ)v_i^{(\ell)}38.700.1320.0880.110
ZEROSUP0061.200.2710.1850.228
Matched attenuation + cascade(1−βℓ)vi(ℓ)(1-\beta_\ell)v_i^{(\ell)}62.150.2880.2180.253
Full FRGI(1−βℓ)vi(ℓ)+βℓv~i(ℓ)(1-\beta_\ell)v_i^{(\ell)}+\beta_\ell\tilde v_i^{(\ell)}65.600.4050.3650.385

Zero suppression improves both accuracy and evidence separation. With the same attenuation schedule and cascade hook, adding the reference raises accuracy from 62.15% to 65.60% (+3.45 percentage points) and mean evidence separation from 0.253 to 0.385 (+0.132). The separation improves at both measured layers. This matched comparison shows that the reference contributes to the accuracy gain and improves proof-evidence selectivity at the two post-peak layers we examined.

Q1.2 (W2). Why do the four case studies select the lower search boundaries?

The four case studies in Figs. 16–19 all select ℓ∗=8\ell^*=8 and ℓc=3\ell_c=3, the lower bounds of the respective search windows for LLaMA-3.1-8B and Mistral-7B. Eq. (25) selects the collapse anchor ℓ∗\ell^* by maximizing the composite UCP score over layers 8–24. Eq. (27) selects the early cascade layer ℓc\ell_c using the rule–fact coupling-weakening score over layers 3–12.

Figure 20 in the revised manuscript shows the median and interquartile range of the composite UCP score on ProverQA alongside the four individual case trajectories. All four cases attain their highest score at layer 8 among the plotted layers. The median score rises from 0.42 at layer 8 to 0.76 at layer 16, its highest value among the plotted layers.

To assess how representative these four cases are, we counted layer selections across the full ProverQA evaluation set. The selector chooses ℓ∗=8\ell^*=8 on 18.34% of inputs for LLaMA-3.1-8B and 21.48% for Mistral-7B. The joint choice (ℓ∗,ℓc)=(8,3)(\ell^*,\ell_c)=(8,3) shown in the four case studies occurs on 6.84% and 8.61% of inputs, respectively.

Then, to measure the effect of adaptive layer selection on accuracy, we compared adaptive UCP with a policy that fixes both layer choices at (8,3)(8,3) on the same ProverQA split.

BackboneFixed (8,3)(8,3) accuracy (%)Adaptive UCP accuracy (%)Gain (pp)
LLaMA-3.1-8B57.4263.24+5.82
Mistral-7B48.3158.67+10.36

These results support the role of UCP as an input-dependent layer selector. The lower-bound pair is uncommon on the evaluation set, and allowing the layer choices to vary improves accuracy over fixing both layers at those bounds.

AuthorsOpenReview ↗

Response to Reviewer #2

Q2 (W2). How does FRGI behave on FOLIO?

We evaluated Qwen3-8B on the combined FOLIO training and validation splits (N=1,208N=1{,}208) using greedy decoding. We treated each natural-language premise as a candidate evidence span and used the conclusion as the query. Fact and rule roles were assigned from the premise text alone, without gold logical forms or proof annotations during inference. All methods used the same evaluation examples and answer-label mapping.

UCP uses signals derived from the fact–rule partition, and the cascade hook requires at least one selected span from each group (Appendix E.2–E.3). We report accuracy alongside the fraction of inputs on which each hook applied an update.

MethodCorrect / NNAccuracy (%)Cascade updatesValue updates
Vanilla630 / 1,20852.15——
Zero-shot CoT669 / 1,20855.38——
ZEROSUP (gate bypassed, diagnostic control)605 / 1,20850.08—304 (25.17%)
Full FRGI682 / 1,20856.46344 (28.48%)304 (25.17%)

Full FRGI reaches 56.46% accuracy, outperforming vanilla inference by 4.31 percentage points and zero-shot CoT by 1.08 points. The value hook applies updates on 25.17% of inputs. The gate-bypassed ZEROSUP control reaches 50.08%, below vanilla accuracy. These results show that FRGI can improve accuracy on FOLIO using fact and rule roles inferred directly from the premise text.

Thank you again for your review. We welcome any follow-up questions. If these results address your concerns, we would appreciate it if you would reconsider your rating.

Thanks authors for the detailed clarification. I think they are helpful in addressing my previous concerns. I would, therefore, maintain my positive recommendation of this paper.

Thank you for your follow-up. We are glad that our clarifications helped address your concerns.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Manuscript revision record

  • v1: Added Appendix Sections G.3 and G.4 and Figure 20 in the revised manuscript.
AuthorsOpenReview ↗

Summary of the Discussion and Additional Evidence

We thank the reviewers for their engagement and the Area Chair for overseeing the discussion. All three reviewers responded to our rebuttal and maintained their ratings of 3 (Weak Acceptance). All three reviewers engaged positively in the discussion, acknowledging that the clarifications and additional experiments addressed key concerns and strengthened the paper.

The discussion clarified the scope of the diagnostic claims and added evidence for the contribution of FRGI.

1. Evidence selectivity is often formed at intermediate layers and subsequently weakens. Across LLaMA-3.1-8B, Mistral-7B, and Qwen3-8B, the ProverQA diagnostic profiles show approximately 60–63% post-peak declines. The analyses also cover PrOntoQA and ProofWriter, with dataset-dependent behavior, including weak formation of query-level selectivity in some ProofWriter settings. Appendix D.5 extends the raw evidence-separation analysis to JustLogic and LogicSkills. We clarified that the empirical finding concerns measured selectivity for annotated proof evidence and distinguished it from the conditional mutual information in Eq. (2).

2. FRGI improves the measured representations alongside reasoning accuracy. In the paired Qwen3-8B–ProverQA experiment under beam search (k=4k=4), FRGI improves all five diagnostics at the two measured post-peak layers. TripletAcc rises from 0.543 to 0.824, and accuracy rises from 38.70% to 65.60%. With attenuation, selected spans, target layers, and the cascade hook held fixed, adding the reference contributes 3.45 accuracy points and increases raw evidence separation by 0.132. The additional placement control shows that the effect of the reference update depends on the intervention configuration.

3. Adaptive layer selection contributes to FRGI's gains. The joint lower-bound choice (8,3)(8,3) occurs on fewer than 9% of ProverQA inputs for either LLaMA or Mistral. Adaptive UCP improves accuracy over fixing both choices at (8,3)(8,3) by 5.82 and 10.36 percentage points, respectively. Figure 20 provides the score profiles requested during discussion.

4. The FOLIO evaluation provides preliminary evidence of transfer. On the combined FOLIO training and validation splits, Qwen3-8B with FRGI achieves 56.46%, compared with 52.15% for vanilla and 55.38% for zero-shot CoT, using fact and rule roles inferred from premise text.

We have added Appendix Sections G.3 and G.4 and Figure 20 to the revised manuscript. We will carry the clarified interpretation and scope of the findings into the final text. We respectfully ask the Area Chair to consider the additional evidence and the reviewers' follow-up assessments when making the final decision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.