acceptodds
Under review as a conference paper at ICLR 2027

Locally Sound, Globally Insufficient: The Local-Global Gap in Multi-Hop Reasoning

Abstract

Reliable multi-hop reasoning requires more than locally supported steps: a trace can be sound at every reasoning step yet still fail to answer the question as a whole. We call this failure regime the *local–global gap* (LGG), in which the trace is **locally sound yet globally insufficient**. Local soundness requires each step to be supported by the available evidence and preceding steps, whereas global sufficiency requires the reasoning trace to align with the question and establish the submitted answer. In a human-adjudicated diagnostic of 2,598 responses across three multi-hop QA benchmarks and three models, we find that the LGG occurs in every benchmark–model combination and accounts for nearly half of globally insufficient responses overall. However, conventional faithfulness verifiers that check claims against the evidence largely miss these failures: at thresholds retaining at least 95% of reliable traces, recall for LGG cases is substantially lower than that for locally unsound traces. To address these failures, we formalize three dependencies for reliable reasoning: evidence-to-step support, question-to-trace alignment, and trace-to-answer closure. Instead of post-hoc diagnosis, we introduce **E-Closure** to supervise these dependencies during training, combining generation supervision on supported original and counterfactual responses with bidirectional switching constraints. Averaged over three benchmarks and three backbones, existing fine-tuning baselines improve accuracy and local soundness over the base models, but at the cost of global sufficiency. **E-Closure** improves both: among all fine-tuned methods, it achieves the highest average accuracy (92.8%) and trace reliability (89.0%) while yielding the lowest LGG rate (6.2%). Data and code are available anonymously at [https://anonymous.4open.science/r/lgg/](https://anonymous.4open.science/r/lgg/).

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-hop reasoning, chain-of-thought(CoT), reasoning evaluation, reasoning reliability, local–global gap, counterfactual learning, preference learning

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_K7sXOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the relationship between the soundness of model’s evidential reasoning trace at an individual level (local), and the evidence set’s support for a final answer (global). The authors analyze a gap between locally sound evidence trace and support for final correct answer (Local-Global Gap), finding existing faithfulness verifiers are insufficient to capture such gaps. The analysis motivate a training method E-CLOSURE, which generates evidence- and question- edited augmentations to enhance model reasoning which reduces the LGG and improves model performance on multi-hop QA tasks.

Critical strengths
  • The authors’ effort to formalize the problem through defining local soundness, global sufficiency and LGG is appreciated. It leads to some interesting observations such as that in regular SFT, local gains can accompany global degradation.
  • The proposed learning method’s three objectives are well-motivated by the analysis (evidence-to-step support, question-to-trace alignment, and trace-to-answer closure)
  • Human annotation study to measure how often locally sound traces are globally insufficient, across three benchmarks and three models, is valuable.
Critical weaknesses
  • Models are tested with thinking disabled: The primary problem the paper’s analysis reveals, and addresses, are path deviation and incomplete reasoning, but I have doubts these are significant sources of failures in models with reasoning post-training, whereas the model evaluates models without thinking enabled.
  • Generalizability beyond multil-hop QA: The generality of the contribution is limited to multi-hop QA, and it is unclear whether the LGG is similarly prevalent and measurable in more complex agentic QA or problem solving scenarios involving tools or longer task horizons.
  • GPT-5.6 Sol checks local soundness, alignment, and closure when constructing the training traces, then grades the trained models on the same criteria, so there may be some model bias in the results.
Questions for authors
  • Could the authors test with thinking enabled for models?
  • Answers to generalizability concerns above?
  • Why is backbone’s LGG much lower on Musique?
  • Answer decoupling means the trace reached the right answer, but the model submits a different one. Would a simpler fix without training, such as a second pass that reads the trace and extracts the answer it reaches, resolve these cases as well as E-CLOSURE does?
2 comments
AuthorsOpenReview ↗

Response to Reviewer K7sX (1/2)

Dear Reviewer K7sX,

We thank you for taking the time to review our work. We especially appreciate your positive assessment of our formalization of LGG, the motivation for our three training objectives, and the value of our diagnostic study. Below, we address your concerns about evaluation with thinking enabled, generalizability beyond multi-hop QA, and potential evaluator bias, along with your questions about MuSiQue and simpler inference-time alternatives.

 


Weakness 1 and Question 1

Weakness 1: Models are tested with thinking disabled: The primary problem the paper’s analysis reveals, and addresses, are path deviation and incomplete reasoning, but I have doubts these are significant sources of failures in models with reasoning post-training, whereas the model evaluates models without thinking enabled.

Question 1: Could the authors test with thinking enabled for models?

Our Response

We thank the reviewer for suggesting this additional evaluation. We test the untuned Qwen3-8B with thinking enabled on all 866 evaluation questions. We retain E-Closure in its original non-thinking configuration and reuse its existing three-seed results.

All configurations are evaluated using the same deterministic answer-matching accuracy, with equal weighting across the three datasets:

Configuration2WikiMultiHopQAMuSiQueHotpotQAMean ACC
Backbone, thinking disabled92.45%50.20%84.98%75.88%
Backbone, thinking enabled100.00%58.04%86.79%81.61%
E-Closure, thinking disabled99.52%76.60%87.29%87.80%

Enabling thinking improves the backbone’s mean accuracy from 75.88% to 81.61%. Nevertheless, E-Closure retains a 6.2-percentage-point advantage in mean accuracy, including an 18.6-point advantage on MuSiQue.

These results show that the overall accuracy advantage persists when comparing against a thinking-enabled backbone. We will include the results and clarify their scope in the revised manuscript.

 


Weakness 2 and Question 2

Generalizability beyond multil-hop QA: The generality of the contribution is limited to multi-hop QA, and it is unclear whether the LGG is similarly prevalent and measurable in more complex agentic QA or problem solving scenarios involving tools or longer task horizons.

Answers to generalizability concerns above?

Our Response

We agree that extending these findings to agentic QA, tool use, and longer task horizons requires further validation. We chose multi-hop QA because explicit questions, supplied evidence, and inspectable reasoning traces allow us to separately measure evidence support, question alignment, and answer closure.

Within this setting, our contribution combines a validated diagnosis with an effective training approach. The diagnostic study of 2,598 responses across three benchmarks and three backbones identifies LGG in 46.3% of globally insufficient responses. Guided by this diagnosis, E-Closure improves trace reliability rate over RM by 7.8 percentage points across nine settings. These results establish the practical importance of the distinction and the value of supervising the affected dependencies.

The distinction also provides a concrete starting point for evaluating agentic QA. A step supported by tool observations does not by itself establish that the agent is pursuing the right goal or has justified its final answer. This motivates checking evidence support, task alignment, and answer closure jointly as agents gather and use information. We will clarify the scope of our evidence and identify validation in these broader settings as future work.

 


Continued in the next response...

AuthorsOpenReview ↗

Response to Reviewer K7sX (2/2)

Weakness 3

GPT-5.6 Sol checks local soundness, alignment, and closure when constructing the training traces, then grades the trained models on the same criteria, so there may be some model bias in the results.

Our Response

To assess this potential evaluator bias, we additionally evaluate Opus 5.5 (High), from a different model family than the GPT-based data constructor, while keeping the annotation guidelines fixed.

We evaluate the same 360 responses from 90 questions used in our human audit. The sample contains 10 questions per dataset–backbone setting, shared across four methods, with all checkpoints using training seed 42.

MethodOriginal judge TRR (%)Human TRR (%)Opus TRR (%)Opus LGG (%)
SFT83.3382.2283.3312.22
CF-SFT82.2277.7880.0011.11
RM82.2281.1180.0012.22
E-Closure86.6784.4485.564.44

E-Closure achieves the highest TRR under all three label sources. Under Opus, its TRR advantage over both CF-SFT and RM is 5.56 percentage points, and it has the lowest LGG rate.

These results provide evidence that the observed reliability gains persist beyond the original evaluator family on this audit sample, consistent with the human judgments. We will include this additional evaluation in the revised manuscript.

 


Question 3

Why is backbone’s LGG much lower on Musique?

Our Response

The backbone’s lower LGG coexists with substantially lower local soundness. Locally unsound responses are excluded from LGG, so a lower LGG rate alone does not indicate more reliable reasoning. On MuSiQue, the changes from the backbone to E-Closure are:

BackboneLSR (%)TRR (%)LGG (%)
Qwen3-8B43.1 → 84.232.2 → 65.911.0 → 18.3
Qwen3-32B58.4 → 91.452.5 → 79.25.9 → 12.2
Ministral-3-8B30.2 → 90.224.7 → 71.65.5 → 18.6

Here, the improvement in LSR exceeds the improvement in TRR for all three backbones. By definition, LGG = LSR − TRR, so LGG increases even though the proportion of fully reliable responses rises substantially. The lower backbone LGG should therefore be interpreted alongside its much larger proportion of locally unsound responses.

 


Question 4

Answer decoupling means the trace reached the right answer, but the model submits a different one. Would a simpler fix without training, such as a second pass that reads the trace and extracts the answer it reaches, resolve these cases as well as E-CLOSURE does?

Our Response

We thank the reviewer for suggesting this simple alternative. We tested one-pass answer re-extraction with Qwen3-8B on the same 866 evaluation questions. The untuned model receives the original question and reasoning trace, with the submitted final-answer field removed, and extracts the answer supported by the trace.

We evaluate the overall effect using the same deterministic answer-matching rule for all methods:

Method2WikiMultiHopQAMuSiQueHotpotQAMean ACC
Untuned backbone92.45%50.20%84.98%75.88%
Backbone + answer re-extraction81.65%43.92%80.48%68.69%
E-Closure99.52%76.60%87.29%87.80%

The mean gives equal weight to each dataset, and E-Closure uses its existing three-seed results.

Re-extraction lowers mean accuracy by 7.2 percentage points relative to the backbone and remains 19.1 percentage points below E-Closure. Thus, the tested second-pass procedure does not provide a comparable alternative in overall answer accuracy.

 


We thank the reviewer again for taking the time to provide valuable feedback on our work. We would be glad to further engage in discussion on any additional concerns.

Best regards,

The Authors

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

The paper introduce the notion of a local-global gap (LGG) in multi-hop reasoning by LLMs, using 3 datasets (HotpotQA, MuSiQue, 2WikiMultiHopQA) and 3 models (Qwen3-8B/32B and Ministral-3-8B) to measure it. The LGG hones in on reasoning traces which are locally sound (each reasoning step makes sense), but globally goes wrong (e.g., by outputting a final answer which is disconnected from the actual reasoning).

They find that there is a non-trivial amount of LGG in these settings, and "traditional" reasoning verifiers tend to miss this. While this local-global distinction is a bit artificial, due to the way these benchmarks are formulated, the paper uses it as inspiration for a more robust SFT recipe, creating counterfactual gold reasoning traces (using a strong GPT model) to counter the LGG tendencies. This shows substantial improvements over baseline SFT approaches.

Critical strengths

Careful measurements of the LGG effect, using human judgment. This motivates the proposed training approach.

Using their recipe for creating constrastive training data, the paper demonstrates strong relative improvements in results.

Critical weaknesses

The local-global distinction seems quite arbitrary, and somewhat an artifact of the task/prompt structures. Ideally the final answer (and premise retrieval) should just be another "local" reasoning step, so complete and sound local reasoning would imply global correctness.

The applicability of this method to general LLM training is not clear. Will this effect go away with more/better data and/or larger models, or should this be incorporated into general post training recipes in some way? I feel this lingering question could limit the general interest in this approach.

Questions for authors

N/A

4 comments
AuthorsOpenReview ↗

Response to Reviewer ZsPX (1/1)

Dear Reviewer ZsPX,

We thank the reviewer for taking the time to review our work. We especially appreciate your positive assessment of our diagnostic study and the strong improvements achieved by our training approach. Below, we address your concerns about the local–global distinction and the broader applicability of our method to LLM training.

 


Weakness 1.1

The local-global distinction seems quite arbitrary, and somewhat an artifact of the task/prompt structures.

Our Response

We use a conventional CoT format, asking for step-by-step reasoning followed by an answer, with the same prompt across all three datasets and backbones. Our diagnostic therefore examines ordinary CoT generation on established QA benchmarks.

The local–global distinction concerns what is checked, rather than how the response is formatted: whether the statements are evidence-supported, whether the trace addresses the question, and whether it establishes the submitted answer. Prompt design may affect how often these failures occur, but they concern the content of the reasoning, not merely compliance with an output format.

 


Weakness 1.2

Ideally the final answer (and premise retrieval) should just be another "local" reasoning step, so complete and sound local reasoning would imply global correctness.

Our Response

Even when the final answer is treated as the last reasoning step, a complete and sound local reasoning does not necessarily imply global correctness. Consider a simplified example of path deviation, a type of global failure identified in our diagnostic study:

Question: In which country was the author of novel N born?

Evidence:
[1] N was written by Ada.
[2] Ada was born in Paris, France.
[3] Ada works in Germany.

Reasoning:
N was written by Ada [1]
→ Ada works in Germany [3]
→ The author’s country of work is Germany.
→ Final answer: Germany.

Every step is supported by the evidence or preceding steps, and the trace is complete. However, the question asks for the country of birth, whose gold answer is France. In our diagnostic study, this global failure accounts for 35.2% of LGG cases.

Detecting this error requires checking question alignment, beyond the support of individual steps.

 


Weakness 2

The applicability of this method to general LLM training is not clear. Will this effect go away with more/better data and/or larger models, or should this be incorporated into general post training recipes in some way? I feel this lingering question could limit the general interest in this approach.

Our Response

Our results show that improving answer accuracy does not automatically eliminate LGG. Conventional training methods, such as SFT and CF-SFT, improve accuracy and local soundness while increasing LGG relative to the untuned backbones. The gap also remains present across all three tested backbones, including Qwen3-8B and Qwen3-32B.

The comparison with CF-SFT provides evidence about training design: it uses the same original and counterfactual generation anchors as E-Closure, yet E-Closure improves trace reliability rate by 5.9% across the nine settings. This supports supplementing generation supervision with explicit supervision of evidence support, question alignment, and trace–answer closure. For post-training tasks with available evidence and reasoning traces, this is a concrete training recommendation supported by our experiments.

The broader contribution is to identify dependencies that ordinary training improvements can leave unresolved and demonstrate one effective way to supervise them.

 


We thank the reviewer again for taking the time to provide valuable feedback on our work. We would be glad to further engage in discussion on any additional concerns.

Best regards,

The Authors

AuthorsOpenReview ↗

Follow-up to Reviewer ZsPX

Dear Reviewer ZsPX,

We understand this is a busy period and thank you again for your review. As the discussion period closes soon, we would like to check whether our response has addressed your concerns, including the distinction between local and global failures and whether the gap persists with larger models. We have also uploaded a revised version of the paper. If so, we would be grateful if you would consider updating your score. We are happy to clarify anything that remains.

Best regards,

The authors

Thank you for the responses, they are appreciated. My overall assessment of the work stays the same.

In terms of the global/local distinction and the example above, I was imagining an "unbroken" sequence of step (each step being "easily" verifiable) sort of like:

Question: In which country was the author of novel N born?

Evidence:
[1] N was written by Ada.
[2] Ada was born in Paris, France.
[3] Ada works in Germany.

Reasoning:
[4] Ada works in Germany [1]
[5] The author’s country of work is Germany. [1,4]
[6] The author of novel N was born in Germany [1,5]  [incorrect logic]

I realize that having a fully qualified "final answer statement" can be tricky, especially given how these task outputs are generally formulated, but that seems more like a syntactic issue rather than a deeper global/local distinction.

AuthorsOpenReview ↗

Response to Reviewer ZsPX

Dear Reviewer ZsPX,

Thank you for the concrete example. We would like to clarify two points.

1 - In multi-hop QA, the answer is often a short, definite value, such as an entity, a date, or yes or no, so that it can be compared with the gold answer. The model needs to output a final answer in addition to the trace. We agree that if the answer were written as a final answer statement, closure could be judged as the local soundness of that statement. In practice, however, the model still has to output the short answer, and the link from the trace to that answer remains. This link does fail: answer decoupling, where the short answer disagrees with the trace, accounts for 69 of the 145 LGG cases (47.6%).

2 - Alignment cannot be reduced in this way. Among the 2,598 annotated responses, alignment fails in 190 (7.3%), and path deviation accounts for 51 of the 145 LGG cases (35.2%), where every step is supported by the evidence. These failures are often triggered by distractor passages that lead the model to a different target. In your example, [3] is such a distractor, and following it means confusing where the author works with where she was born. A trace can do this without a single wrong step:

1. N was written by Ada. [1]
2. Ada works in Germany. [3]
3. The author of novel N works in Germany. [Steps 1, 2]

Answer: Germany

Step 3 is already a fully qualified final statement, and every step is supported. The error is that the question asks where the author was born. A model that drifts to the wrong target writes the drifted target into its final statement too, so the error becomes visible only when the trace is compared with the question.

Best regards,

The authors

Reviewer_9ud6OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces the local–global gap /LGG: reasoning whose steps are evidence-supported but fail to address the question or justify the answer. Human evaluation identified 145 LGG cases among 2,598 responses, accounting for 46.3% of globally insufficient responses. Existing evidence verifiers detect at most 11% while retaining 95% of reliable traces. The proposed E-CLOSURE objective combines original and counterfactual generation anchors with bidirectional margins for evidence–step, question–trace, and trace–answer dependencies. Across 9 settings, it achieves 92.8% accuracy, 89.0% trace reliability, and a 6.2% LGG rate, outperforming SFT, CF-SFT, and response-margin baselines.

Critical strengths
  • The human study is well designed. It turns an intuitive failure mode into a measured one with confidence intervals.

  • Effective methods. SFT and CF-SFT raise accuracy and local soundness but roughly double the LGG rate, from ~5% to ~10%.

  • Solid ablation study. Removing any dependency margin, or the reverse direction, lowers TRR by 4–7 points, with paired CIs above zero. This supports the claim that each component contributes.

Critical weaknesses
  • LGG covers only ~5.6% of all responses, and E-CLOSURE's average LGG rate (6.2%) is still above the untuned backbone's (5.4%). On MuSiQue it is markedly worse (12–19% vs. 5–11%). The method mainly repairs damage caused by fine-tuning rather than closing the gap relative to the base model.

  • 2WikiMultiHopQA and HotpotQA are near ceiling for all fine-tuned methods, so most of the signal comes from MuSiQue.

  • The baselines are evidence-entailment metrics that were never designed to check alignment or closure. Missing comparisons include a holistic LLM-judge or reasoning-chain evaluators (e.g., ROSCOE, ReCEval, RACE) and a simple re-derive the answer from the trace check, which is especially relevant because answer decoupling makes up 47.6% of LGG cases.

  • Limited evidence of generalization. Switching gains (e.g., +69.6 points on C over RM) use pairs from the same intervention pipeline and the training margins for evaluation, demonstrating objective optimization rather than broader reasoning generalization.

Questions for authors
  1. How would a trivial inference-time fix compare with E-CLOSURE on answer decoupling? For example, re-extract the final answer from the trace, or add a closure-check-and-regenerate step.
  2. How much of the gain holds up under an evaluator from a different model family than the one that built the training data? Please report main results with an independent judge, or with human labels on a larger sample.
  3. Why does E-CLOSURE raise LGG on MuSiQue relative to the backbone in all three models, and what do those residual cases look like?
  4. Did the alignment/closure judges in Table 1 (gpt-5.6-luna) share prompts or guidelines with the annotators? Could that favor them over the evidence-support verifiers?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 9ud6 (1/3)

Dear Reviewer 9ud6,

We thank you for taking the time to review our work. We especially appreciate your positive assessment of our diagnostic study, the effectiveness of our training approach, and the solid ablation results. Below, we address your concerns about improvements over the untuned backbones, comparisons with baselines, and the independence and broader relevance of our evaluation.

 


Weakness 1

LGG covers only ~5.6% of all responses, and E-CLOSURE's average LGG rate (6.2%) is still above the untuned backbone's (5.4%). On MuSiQue it is markedly worse (12–19% vs. 5–11%). The method mainly repairs damage caused by fine-tuning rather than closing the gap relative to the base model.

Our Response

LGG is a strictly defined failure category requiring both local soundness and global insufficiency. Locally unsound responses are excluded, so a lower LGG rate alone does not imply more reliable reasoning. Although LGG accounts for 5.6% of all diagnostic responses, it constitutes 46.3% of globally insufficient responses.

We therefore also report trace reliability rate (TRR), the proportion of responses that are both locally sound and globally sufficient. Across the nine settings, E-Closure improves TRR from 70.0% to 89.0%, global sufficiency rate (GSR) from 86.8% to 91.5%, and ACC from 79.8% to 92.8% relative to the untuned backbones. The corresponding changes on MuSiQue are:

MuSiQue backboneLSR (%)TRR (%)GSR (%)LGG (%)
Qwen3-8B43.1 → 84.232.2 → 65.971.8 → 72.911.0 → 18.3
Qwen3-32B58.4 → 91.452.5 → 79.279.6 → 82.25.9 → 12.2
Ministral-3-8B30.2 → 90.224.7 → 71.665.5 → 76.65.5 → 18.6

On MuSiQue, local soundness improves more than TRR. Since LGG = LSR − TRR, this explains why LGG rises alongside substantial improvements in reliable reasoning. All three models also improve in global sufficiency. These gains over the untuned backbones, together with reduced LGG relative to the fine-tuned baselines, demonstrate that the method’s value extends beyond repairing degradation caused by fine-tuning.

 


Weakness 2

2WikiMultiHopQA and HotpotQA are near ceiling for all fine-tuned methods, so most of the signal comes from MuSiQue.

Our Response

We agree that near-ceiling performance limits the room for improvement, particularly on 2WikiMultiHopQA. However, high answer accuracy does not imply that reasoning reliability has reached its ceiling. On HotpotQA, E-Closure improves TRR by 6.4–17.5 percentage points over the untuned backbones and 1.9–5.3 percentage points over RM:

HotpotQA backboneUntuned TRR (%)RM TRR (%)E-Closure TRR (%)
Qwen3-8B85.088.393.6
Qwen3-32B91.994.598.3
Ministral-3-8B76.391.993.8

E-Closure also reduces LGG relative to RM in all three settings. Thus, while MuSiQue contributes the largest gains, HotpotQA also shows consistent improvements in reasoning reliability across all three backbones.

 


Weakness 3

The baselines are evidence-entailment metrics that were never designed to check alignment or closure. Missing comparisons include a holistic LLM-judge or reasoning-chain evaluators (e.g., ROSCOE, ReCEval, RACE) and a simple re-derive the answer from the trace check, which is especially relevant because answer decoupling makes up 47.6% of LGG cases.

Our Response

We agree that evidence-entailment metrics were not designed to assess question alignment or answer closure. Our diagnostic uses them to examine whether evidence support alone is sufficient to identify reliable reasoning.

Importantly, Table 1 already includes an LLM-based question-alignment judge and a trace-to-answer closure detector, alongside the evidence-support baselines. At thresholds retaining 95% of reliable traces, these two detectors identify 42.1% and 42.8% of LGG cases, respectively, compared with at most 11.0% for the evidence-support detectors. Their strengths also match the targeted dependencies: the alignment judge detects 66.7% of path deviations, while the closure detector detects 56.5% of answer-decoupling cases.

These results support the value of checking alignment and closure alongside evidence support. We agree that comparisons with holistic reasoning evaluators would broaden this detector study.

We also evaluate the suggested answer-re-extraction baseline using Qwen3-8B. Mean accuracy under deterministic answer matching is 68.69%, compared with 75.88% for the untuned backbone and 87.80% for E-Closure. Thus, the evaluated re-extraction procedure does not recover our method’s overall accuracy gains. We provide the protocol and dataset-level results in our response to Question 1.

 


Continued in the next response...

AuthorsOpenReview ↗

Response to Reviewer 9ud6 (2/3)

Weakness 4

Limited evidence of generalization. Switching gains (e.g., +69.6 points on C over RM) use pairs from the same intervention pipeline and the training margins for evaluation, demonstrating objective optimization rather than broader reasoning generalization.

Our Response

We agree that switching gains alone do not establish broad reasoning generalization. This test checks whether the model learns the intended dependencies.

Our main evidence comes from ordinary generation on 866 held-out questions from three benchmarks’ official validation/test splits, while training uses their official training splits. Models generate complete traces and answers from the original questions and evidence. We evaluate answer correctness and reasoning reliability without counterfactual pairs or training-margin scores.

Across three datasets and three backbones, E-Closure achieves the following average gains:

ComparatorACC gain (pp)TRR gain (pp)
SFT+5.0+7.1
CF-SFT+4.1+5.9
RM+2.3+7.8

These results show that the benefits extend beyond the switching test to ordinary generation on unseen questions. They support generalization within the three evaluated QA benchmarks.

 


Question 1

How would a trivial inference-time fix compare with E-CLOSURE on answer decoupling? For example, re-extract the final answer from the trace, or add a closure-check-and-regenerate step.

Our Response

We thank the reviewer for suggesting this practical comparison. We evaluate a one-pass answer-re-extraction baseline using Qwen3-8B on the same 866 evaluation questions. For each original backbone response, we provide the question and reasoning trace to the same untuned model and ask it to extract the supported final answer. The original final-answer field is removed, and no reference answer is provided.

All methods use the same deterministic answer-matching rule from Appendix D.5.3. We compare against the existing E-Closure results averaged over three training seeds; the mean gives equal weight to each dataset.

Method2WikiMultiHopQAMuSiQueHotpotQAMean ACC
Untuned backbone92.45%50.20%84.98%75.88%
Backbone + answer re-extraction81.65%43.92%80.48%68.69%
E-Closure99.52%76.60%87.29%87.80%

Answer re-extraction reduces mean accuracy by 7.2 percentage points relative to the original backbone and remains 19.1 percentage points below E-Closure. Thus, the evaluated one-pass re-extraction procedure does not provide a comparable alternative to E-Closure in overall answer accuracy.

 


Question 2

How much of the gain holds up under an evaluator from a different model family than the one that built the training data? Please report main results with an independent judge, or with human labels on a larger sample.

Our Response

We thank the reviewer for suggesting an evaluation with a different model family. We additionally evaluate Opus 5.5 (High) using the same frozen annotation guidelines as the original evaluator and human audit.

The evaluation covers the same 90 questions and 360 responses used in our human audit, with 10 questions per dataset–backbone setting. Each question is assigned to one backbone, and all four methods use checkpoints from training seed 42. The three label sources therefore assess identical responses.

MethodOriginal judge TRR (%)Human TRR (%)Opus TRR (%)Opus LGG (%)
SFT83.3382.2283.3312.22
CF-SFT82.2277.7880.0011.11
RM82.2281.1180.0012.22
E-Closure86.6784.4485.564.44

Under Opus, E-Closure achieves the highest TRR and lowest LGG rate. Its TRR advantage over both CF-SFT and RM is 5.56 percentage points, compared with 4.44 percentage points under the original evaluator on the same sample. Answer accuracy is 92.22%, tied with RM.

Thus, the observed gains in reasoning reliability persist under an evaluator from a different model family, consistent with the direction of the human-audit results. We will include this additional evaluation in the revised manuscript, identifying its scope as the shared audit sample.

 


Continued in the next response...

AuthorsOpenReview ↗

Response to Reviewer 9ud6 (3/3)

Question 3

Why does E-CLOSURE raise LGG on MuSiQue relative to the backbone in all three models, and what do those residual cases look like?

Our Response

1. Why LGG increases.

On MuSiQue, E-Closure improves TRR by 26.7–46.9 percentage points, while local soundness improves even more in each backbone. Since LGG = LSR − TRR, these gains can coexist with a higher LGG rate.

To examine the change directly, we paired outputs on 255 questions across three backbones and three training seeds. The 337 newly occurring LGG responses originate from:

Backbone responseNewly occurring LGG responses
Locally unsound284 (84.3%)
Fully reliable53 (15.7%)

Thus, most newly occurring LGG responses gain local soundness, although the 53 transitions from fully reliable responses represent genuine regressions.

2. What the residual cases look like.

The 375 residual LGG responses have the following composition. Counts refer to question–backbone–seed combinations.

Failure typeCount (%)Concrete example
Incomplete reasoning191 (50.9%)Lists Helvetia’s county and Tucson’s council term but omits the link between Tucson and that county.
Path deviation156 (41.6%)Answers a question about a record label’s owner using its founder, James Conkling.
Answer decoupling28 (7.5%)States that San Diego ranks third in the trace but submits “first.”

Missing reasoning links and target deviations account for 92.5% of residual LGG. These findings locate the remaining challenge: stronger evidence grounding does not always produce a complete, correctly targeted chain, despite substantial gains in overall reliability. We will include this paired analysis and the examples in the revised manuscript.

 


Question 4

Did the alignment/closure judges in Table 1 (gpt-5.6-luna) share prompts or guidelines with the annotators? Could that favor them over the evidence-support verifiers?

Our Response

The alignment judge and human annotators share the operational definition of question alignment: whether the trace addresses the entity, relation, and constraints requested by the question. This ensures that both assess the same property.

We also clarify that the closure detector uses T5-XXL entailment, not GPT-5.6 Luna (Appendix A.5). It treats the trace as the premise and the submitted answer to the question as the hypothesis. Meanwhile, our RAGAS-style and VeriScore-style evidence-support baselines also use GPT-5.6 Luna, so the alignment judge does not have an exclusive advantage in the evaluator backbone.

The alignment judge explicitly targets the alignment criterion. Accordingly, Table 1 compares the coverage of different verification targets, rather than establishing a general ranking of evaluator quality. All detectors assess the same responses, with recall measured at thresholds retaining 95% of reliable traces. The results support checking question alignment and answer closure alongside evidence support. We will clarify the shared definition and this interpretation in the revised manuscript.

 


We thank the reviewer again for taking the time to provide valuable feedback on our work. We would be glad to further engage in discussion on any additional concerns.

Best regards,

The Authors

AuthorsOpenReview ↗

Follow-up to Reviewer 9ud6

Dear Reviewer 9ud6,

We understand this is a busy period and thank you again for your review. As the discussion period closes soon, we would like to check whether our response has addressed your concerns, including the LGG rate on MuSiQue, the answer re-extraction baseline, and the evaluator check with Opus 5.5. We have also uploaded a revised version of the paper. If so, we would be grateful if you would consider updating your score. We are happy to clarify anything that remains.

Best regards,

The authors

Reviewer_s6aeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the local–global gap in multi-hop QA, where reasoning steps can be individually supported while the trace fails to address the question or establish the final answer. It distinguishes evidence support, question alignment, and answer closure. It proposes E-CLOSURE to supervise these dependencies through counterfactual generation and bidirectional preference margins. Validation includes a human-adjudicated study of 2,598 responses, verifier comparisons. Experiments across three benchmarks and three backbones, with three training seeds, objective ablations, and human audits.

Critical strengths
  1. The diagnostic is carefully constructed and useful. Independent screening, duplicate annotation, and adjudication support the distinction between local soundness and global sufficiency. Finding 145 LGG cases, accounting for 46.3% of globally insufficient responses, establishes that this is a measurable failure category.

  2. The training results show improvements. E-CLOSURE improves average accuracy and trace reliability over RM by 2.3 and 7.8 percentage points, respectively. Component ablations strengthen the evidence that the proposed supervision is useful.

Critical weaknesses
  1. The paper provides a useful diagnosis of the local–global gap, but does not sufficiently establish the advantages of its specific training design over simpler alternatives. Although the ablations support the contribution of individual components, the paper lacks a clear motivation of the method. Given that the method primarily combines counterfactual augmentation with bidirectional preference constraints, its additional value and methodological contribution beyond simpler alternatives remain insufficiently established.

  2. The closure objective rewards answers implied by an edited trace even when that trace contradicts the unchanged original evidence. The answer-switching evaluation rewards the same behavior, so its advantage over RM partly reflects the two objectives’ different preferences under evidence–trace conflict. Although performance improves on ordinary generations, the paper does not examine whether this supervision preserves the ability to correct erroneous intermediate reasoning, leaving its benefits and potential costs incompletely assessed.

Questions for authors
  1. Could the authors explain more clearly how the diagnostic findings motivate the specific method design?

  2. When a supplied trace contradicts valid evidence, should the model follow the trace or correct it? Given that the counterfactual closure objective rewards following the trace, what evidence shows that this supervision preserves the ability to correct erroneous intermediate reasoning?

4 comments
AuthorsOpenReview ↗

Response to Reviewer s6ae (1/2)

Dear Reviewer s6ae,

We thank you for taking the time to review our work. We especially appreciate your positive assessment of our diagnostic study and the improvements in accuracy and trace reliability achieved by our training approach. Below, we clarify how the diagnosis motivates our method, explain its advantages over simpler alternatives, and address your concerns about correcting erroneous intermediate reasoning.

 


Weakness 1.1 and Question 1

Weakness 1.1: The paper provides a useful diagnosis of the local–global gap, but does not sufficiently establish the advantages of its specific training design over simpler alternatives. Although the ablations support the contribution of individual components, the paper lacks a clear motivation of the method.

Question 1: Could the authors explain more clearly how the diagnostic findings motivate the specific method design?

Response

We thank the reviewer for recognizing our diagnostic contribution. Our diagnostic identifies three failure patterns:

1. Evidence-to-step support failures: Reasoning steps contain unsupported claims or invalid inferences.
2. Question-to-trace alignment failures: The trace pursues a target different from the one requested by the question.
3. Trace-to-answer closure failures: The final answer does not follow from the reasoning trace. 

These distinct failure patterns motivate separate constraints on the affected reasoning step, trace, and final answer. We therefore apply dependency-specific supervision, with each constraint targeting the corresponding failure type:

DependencySupervisionTraining objective
Evidence-to-step supportS:D→st(e)S: D \to s_{t(e)}Train the model to prefer reasoning steps supported by the given evidence over unsupported alternatives.
Question-to-trace alignmentT:q→rT: q \to rTrain the model to prefer reasoning traces that address the question over those that pursue a different target.
Trace-to-answer closureC:r→aC: r \to aTrain the model to prefer final answers supported by the reasoning trace over unsupported alternatives.

We use paired original and counterfactual conditions to supervise each dependency in both directions: changing the evidence, question target, or trace should change the preferred step, trace, or answer, respectively. Here, qq, DD, rr, and aa denote the question, evidence, reasoning trace, and final answer; st(e)s_{t(e)} is the first step affected by editing evidence item ee.

 


Weakness 1.2

Given that the method primarily combines counterfactual augmentation with bidirectional preference constraints, its additional value and methodological contribution beyond simpler alternatives remain insufficiently established.

Response

Although counterfactual augmentation and bidirectional preference learning are established techniques, our methodological contribution lies in translating the diagnosed failure modes into separate preference constraints on the affected step, reasoning trace, and final answer.

To assess the added value of this design, we compare against three simpler alternatives: SFT, CF-SFT, and RM.

1. SFT uses only the original generation anchors, without counterfactual augmentation or preference constraints
   
2. CF-SFT is the first simpler alternative. It uses the same original and counterfactual generation anchors without preference constraints.
   
3. RM is the second simpler alternative. It adds bidirectional preferences over complete responses, thus already combining both ingredients highlighted in this concern.

All three methods use the same training data, update budget, and checkpoint-selection rule as E-Closure. Averaged across nine dataset–backbone settings, E-Closure achieves the following gains in answer accuracy (ACC) and trace reliability rate (TRR):

AlternativeACC gain (pp)TRR gain (pp)
SFT+5.0+7.1
CF-SFT+4.1+5.9
RM+2.3+7.8

These comparisons support the added value of dependency-specific supervision over the evaluated simpler alternatives, beyond ordinary supervised fine-tuning, counterfactual augmentation, or full-response preference constraints.

 


Continued in the next response...

AuthorsOpenReview ↗

Response to Reviewer s6ae (2/2)

Weakness 2 and Question 2

Weakness 2: The closure objective rewards answers implied by an edited trace even when that trace contradicts the unchanged original evidence. The answer-switching evaluation rewards the same behavior, so its advantage over RM partly reflects the two objectives’ different preferences under evidence–trace conflict. Although performance improves on ordinary generations, the paper does not examine whether this supervision preserves the ability to correct erroneous intermediate reasoning, leaving its benefits and potential costs incompletely assessed.

 
Question 2: When a supplied trace contradicts valid evidence, should the model follow the trace or correct it? Given that the counterfactual closure objective rewards following the trace, what evidence shows that this supervision preserves the ability to correct erroneous intermediate reasoning?

Response

We thank the reviewer for this constructive suggestion to examine whether closure supervision affects error correction. We address the concern from two perspectives:

1. Method design and ordinary-generation performance.

E-Closure is designed to reduce reasoning errors during generation. The S constraint penalizes unsupported intermediate steps, the T constraint penalizes traces that deviate from the question, and the C constraint penalizes answers not established by the trace.

Across the nine dataset–backbone settings, these objectives yield the following improvements:

MetricUntuned backbonesE-Closure
Local soundness rate (LSR)75.5%95.2%
Question alignment (A)91.9%95.4%
Answer accuracy (ACC)79.8%92.8%

These results show that the full objective produces fewer responses with unsupported reasoning or question misalignment, while improving final-answer accuracy. Our method therefore reduces the occurrence of reasoning failures in ordinary generation, even without an explicit objective for correcting erroneous prefixes.

 

2. Additional evaluation with corrupted reasoning prefixes.

We evaluate Qwen3-8B on 120 held-out questions, comprising 40 questions from each dataset. Each question has a correct prefix and a corrupted prefix that differs only in the last intermediate fact, while the question and valid evidence remain unchanged. The following simplified example illustrates how the prefixes are supplied:

Shared input for both runs
-------------------------
SYSTEM: [Original system prompt, unchanged]

USER:
Question: Which novel's author died earlier, Novel A or Novel B?

Evidence:
[1] Novel A was written by Ada.
[2] Ada died in 1989.
[3] Novel B was written by Ben.
[4] Ben died in 1995.


Run 1: Correct prefix
---------------------
<|im_start|>assistant (prefilled, unfinished):
1. Novel A was written by Ada [1].
2. Ada died in 1989 [2].
[Model continues generating from here]


Run 2: Corrupted prefix
-----------------------
<|im_start|>assistant (prefilled, unfinished):
1. Novel A was written by Ada [1].
2. Ada died in 2008 [2].
[Model continues generating from here]

Each prefix is supplied as the beginning of the assistant’s response, and the model continues under the original prompt without an explicit correction instruction. We average E-Closure and w/o-C over training seeds 42, 43, and 44; the untuned backbone is evaluated once.

MethodCorrect-prefix ACCCorrupted-prefix ACCACC drop (pp)
Backbone75.0%29.2%45.8
w/o-C84.4%33.1%51.4
E-Closure85.8%29.7%56.1

In this controlled test, although E-Closure shows a larger accuracy drop than w/o-C under prefix corruption, the corrupted-prefix ACC comparison does not provide statistically conclusive evidence that the C constraint further impairs recovery from erroneous intermediate reasoning.

Overall, 1. Our method is designed to reduce reasoning errors during generation through dependency-specific supervision, and the main experiments demonstrate this benefit. 2. We will clearly distinguish error prevention from recovery under corrupted prefixes, and incorporate the additional experiment and its findings into the revised manuscript.

 


We thank the reviewer again for taking the time to provide valuable feedback on our work. We would be glad to further engage in discussion on any additional concerns.

Best regards,

The Authors

Weakness 1.1 and Question 1

The question is why the diagnostic findings lead to the training mechanism of paired counterfactual construction with bidirectional preference margins. The expected answer is an explanation of the necessity of this mechanism compared with other methods. The response maps the three failure modes to the three constraints. This mapping explains what the method supervises but does not explain why the method uses this mechanism for supervision.

Weakness 1.2

The concern is the methodological contribution of the method beyond the combination of counterfactual augmentation and bidirectional preference constraints. The expected answer is either an argument for dependency-specific supervision or a comparison with methods outside this line of work. The response reports gains over SFT, CF-SFT, and RM. The original submission contains these comparisons, and all three baselines belong to the same line of methods. These comparisons show the effect of splitting the supervision, but not the necessity of the method relative to other methods.

Weakness 2 and Question 2

The method adds a training term that makes the answer follow the reasoning trace. The question is whether this term weakens the model’s ability to correct a trace that contradicts the evidence. The expected answer is comparing models with and without this term on traces with errors, and a statement that whether the model should follow or correct such a trace. The response reports gains over the backbones and argues that the accuracy difference under corrupted prefixes is not statistically significant. The backbone comparison does not isolate the effect of this term. The authors’ own experiment shows an accuracy drop of 56.1 points with the term and 51.4 points without it. The response does not discuss this result, and it does not say whether the model should follow or correct the trace.

The response does not resolve these concerns, so the rating remains unchanged.

AuthorsOpenReview ↗

Response to Reviewer s6ae

Dear Reviewer s6ae,

Thank you for your reply and the precise follow-up. It makes clear what kind of answer is expected for each point, and we answer in that form below.


Weakness 1.1 and Question 1

Why does the diagnosis lead to paired counterfactuals with bidirectional margins?

Our Response

In an LGG case, every step is supported by the evidence, so checking steps against the evidence or rewarding correct answers does not show what went wrong. The SFT and CF-SFT results confirm this: accuracy and local soundness go up, but global sufficiency does not. What breaks is a dependency. The trace should change when the question changes, and the answer should change when the trace changes. To supervise a dependency, the model must see the same input with one part changed, together with the output that should change with it. That is a counterfactual pair. For example, if the question asks for the actor instead of the director, with the same evidence, the trace should go through Cage instead of Jewison. A one-way preference for the Jewison trace under the director question can be learned without reading the question, simply by always preferring Jewison. Adding the reverse preference under the actor question removes this shortcut, because the only way to satisfy both is to read which relation the question asks for. This is why the margins are paired and bidirectional. SFT, one-way preference learning such as DPO, and outcome rewards do not force the output to change with the input in this way.


Weakness 1.2

Contribution beyond counterfactual augmentation and bidirectional preference

Our Response

We agree that SFT, CF-SFT, and RM come from the same line. They are controls that remove one ingredient at a time. RM keeps the same counterfactual pairs, generation anchors, margin weights, and hyperparameters as E-Closure, with bidirectional margins over complete responses (Appendix D.3). RM is therefore the combination the reviewer describes, and the gap between RM and E-Closure (accuracy 90.5 to 92.8, global sufficiency 84.9 to 91.5, trace reliability 81.2 to 89.0) is what dependency-specific supervision adds beyond it. This is expected, because original and edited responses share most of their tokens, and a whole-response margin spreads its signal over content that should not change. Methods outside this line do not target these dependencies. Outcome rewards give full credit to a correct answer even when the trace does not lead to it, and step-level rewards pass LGG traces by construction, since every step is supported.

More broadly, our contribution is not a new optimization tool. It is (1) a human-adjudicated diagnosis showing that locally sound traces still fail globally, and that these failures sit in two specific dependencies, from question to trace and from trace to answer, and (2) a training design derived from that diagnosis, in which each constraint changes exactly one input and supervises only the output that should change with it. Counterfactual pairs and preference margins are the means. Knowing which dependencies to supervise, and where in the output, is what the diagnosis contributes and what the RM comparison shows to matter.


Weakness 2 and Question 2

Follow or correct, and the 56.1 versus 51.4 result

Our Response

We agree with the reviewer: the model should correct a trace that contradicts the evidence rather than follow it, and the answer should then follow the corrected trace. Our design already requires this. The S term ties every step to the evidence, so correcting an unsupported step happens inside the trace. The C term acts only on the answer once the trace is complete, so the answer follows a trace whose steps are already checked against the evidence. The results support this. Removing S lowers local soundness by 3.70 points (95% CI [1.96, 5.54]), and adding C does not weaken it: local soundness is 93.76% with C and 90.72% without C (Table 3). The corrupted-prefix test inserts an error the model rarely makes on its own. There, the 4.7-point difference between the two drops splits into 1.3 points from higher accuracy with clean prefixes and 3.4 points from lower accuracy with corrupted prefixes, about 4 of 120 questions. We will state in the limitations that E-Closure does not explicitly train recovery from such inserted errors.


We hope these answers address the three points in the form the reviewer asked for. If they do, we would be grateful if the reviewer would consider updating the score. We are happy to answer any further questions.

Best regards,

The authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal and Revision Summary

Dear Reviewers and Area Chair,

We thank the reviewers for their careful feedback and their recognition of our diagnostic study and training approach. We have addressed each comment and question in our individual responses, clarified the scope of our claims, and added experiments and analyses targeting the main concerns (see Rebuttal Summary).

The additional experiments and analyses strengthen the evidence for our findings and clarify their scope. Our central contribution is a diagnostic study of the local–global gap and dependency-specific training that improves ordinary-generation reliability across the evaluated benchmarks. We incorporated these results and clarifications into the revised manuscript (see Revision Summary).

Rebuttal Summary


  1. Conceptual distinction [s6ae, ZsPX]. We clarified how evidence support, question alignment, and answer closure differ, and how the diagnosed failures motivate dependency-specific supervision. Controlled comparisons with SFT, CF-SFT, and RM support the design’s value.

  1. Evaluator independence [9ud6, K7sX]. We added an evaluation using Opus 5.5 (High) on the same 90-question, 360-response human-audit sample. E-Closure achieves 85.56% TRR, compared with 80.00% for both CF-SFT and RM. The observed reliability advantage persists under a different evaluator family and agrees with the direction of the human judgments.

  1. Thinking-enabled inference [K7sX]. We evaluated the untuned Qwen3-8B with thinking enabled on all 866 questions. Under deterministic answer matching, enabling thinking increases the backbone’s average accuracy from 75.88% to 81.61%. E-Closure, evaluated in its original non-thinking configuration, achieves 87.80%. Thus, thinking improves the backbone’s accuracy, while E-Closure retains a substantial 6.2-percentage-point advantage over the thinking-enabled backbone.

  1. Answer re-extraction [9ud6, K7sX]. We tested one-pass answer re-extraction using the untuned Qwen3-8B on the same 866 questions. Mean answer-matching accuracy is 68.69%, compared with 87.80% for E-Closure. The evaluated procedure therefore does not provide a comparable alternative in overall answer accuracy.

  1. Higher LGG on MuSiQue [9ud6, K7sX]. We clarified that LGG = LSR − TRR, so larger gains in LSR than in TRR can produce higher LGG alongside improved overall reliability. Residual-error statistics and concrete examples show that missing reasoning links and target deviations dominate the remaining LGG cases.

  1. Error prevention and correction [s6ae]. We clarified that C-switch evaluates the trace–answer dependency, not error-correction ability. Closure supervision operates jointly with evidence-support and question-alignment constraints. The full objective, including C, improves ordinary-generation local soundness from 75.5% to 95.2% across nine settings. We also tested 120 held-out questions using correct and corrupted reasoning prefixes while keeping the valid evidence unchanged, to assess whether the model could still produce the correct final answer. This additional test highlights the importance of preventing intermediate errors and shows that recovery is already challenging for the untuned backbone.

  1. Broader applicability [9ud6, ZsPX, K7sX]. We clarified why we chose three multi-hop QA benchmarks, whose explicit questions and supplied evidence allow a controlled study of LGG. Our findings also offer useful guidance for agentic QA by highlighting the need to evaluate evidence support, task alignment, and answer closure together.

Revision Summary

1. Updated the main figure (Section 1, P2, Figure 1).

2. Added evaluation with an independent model family (Section D.8, P46–47, Figure 8).

3. Added a thinking-enabled backbone comparison (Section D.9, P47, Figure 9).

4. Added an inference-time answer re-extraction baseline (Section D.10, P48, Figure 10).

5. Added a controlled prefix-corruption experiment (Section D.11, P48–49, Figure 11).

6. Expanded the analysis of residual LGG on MuSiQue (Section D.12, P49–50, Figures 12–13).

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.