acceptodds
Under review as a conference paper at ICLR 2027

Do What You Say: Towards Verifiable Reasoning for Scientific Editing

Abstract

Language models increasingly produce plans or rationales before their final outputs to guide generation and support supervision. When these traces are used for auditing or process supervision, a mismatch between their stated actions and the final output creates a say–do gap. Scientific editing provides a concrete testbed for this problem because domain tools can check both the requested outcome and the realized input–output transformation. Yet free-form traces make such comparison difficult because the claimed edit must first be interpreted from prose. We introduce TRACE, a structured reasoning-supervision framework that makes the declared edit explicit and directly checkable. TRACE represents each edit as a typed, replayable Goal–Local Edit–Safeguard contract constructed from task-computable facts. Through supervised fine-tuning, models learn to generate this contract before answering, while outcome-based reinforcement learning further improves task performance. Across three molecular editing benchmarks, TRACE consistently outperforms matched answer-only and free-form reasoning baselines, with gains that persist across model scales, backbones, and RL optimizers. On TOMG with Qwen3-1.7B, TRACE improves the average success–similarity score by over matched answer-only training. TRACE also makes agreement between stated and realized edits directly auditable. Controlled interventions provide evidence that the input-specific Local Edit carries most of the action-relevant signal, while the full contract performs best. The resulting verifier further improves say–do consistency through process supervision and enables reference-answer-free Best-of- selection that improves both editing performance and trace–answer consistency. Beyond molecular editing, TRACE extends to RNA codon optimization and reaction repair, demonstrating broader applicability to scientific editing.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, verifiable reasoning, structured reasoning supervision, trace-answer consistency, scientific editing, molecular editing

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_AXUJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes TRACE, a structured reasoning framework for scientific editing. It represents the editing process using Goal, Local Edit, and Safeguard, allowing the stated edits to be checked against the final outputs. The method uses SFT with structured traces followed by outcome-based GRPO. Experiments on molecular editing and other scientific tasks show consistent improvements over answer-only and free-form reasoning baselines.

Critical strengths
  1. The paper addresses an important problem: the gap between what a model says and what it actually does.
  2. The structured edit representation is simple and useful. It also allows automatic verification without reference answers at inference time.
  3. The experiments are comprehensive, covering multiple tasks, model sizes, and training settings. The ablation and intervention studies provide useful evidence for the role of Local Edit.
Critical weaknesses
  1. The novelty could be better justified. The framework combines structured supervision, domain-specific verification, and existing training methods, but the distinction from prior structured reasoning approaches is not fully clear.
  2. TRACE provides additional edit information during SFT compared with the answer-only baseline. It is unclear how much of the improvement comes from the extra supervision and how much comes from the structured reasoning format itself.
  3. The verification relies on task-specific tools and predefined edit rules. Its applicability to more open-ended scientific reasoning tasks remains uncertain.
Questions for authors
  1. Can the authors further separate the benefits of additional edit supervision from those of the structured TRACE format?
  2. How well would TRACE work on tasks where edit operations cannot be easily extracted or verified?
  3. Does higher say-do consistency always lead to better task performance, especially on more difficult editing tasks?
  4. How does the additional training and inference cost of TRACE compare with the answer-only baseline?
Reviewer_b9bJOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TRACE asks a scientific editing model to state its goal, local edits, and safeguards in a structured contract before producing an answer. Training contracts are extracted from input–reference pairs with task-specific tools; supervised training is followed by outcome-based reinforcement learning. At test time, the contract can be checked against the input and output without a reference answer. The evaluation separates completeness, claim correctness, and agreement with the realized edit. Experiments cover molecular editing and other executable editing tasks, with narrative controls and interventions on the contracts. Task performance improves, although the results also show that better answers need not come with more faithful contracts.

Critical strengths
  1. An edit can satisfy the task while the model gives an incorrect account of what it changed. TRACE makes this failure observable through executable checks of explicit commitments, without relying on a language-model judge. I see practical value in that distinction: a user can check which promised changes actually occurred, even when the usual task metric reports success.
  2. The matched narrative and narrative-plus-operation controls make the evaluation more convincing, as do the component ablations and contract interventions. I also found the process-reward results useful: improving consistency does not reliably improve task outcomes. These experiments help explain what the audit signal captures and where its usefulness ends.
Critical weaknesses

The closest control, narrative+ops, reaches 0.5686 average SR×Sim compared with 0.5762 for TRACE. That is a small difference, and the reported variation is across decoding seeds from a single training run. TRACE also produces longer traces. The training-time comparison shows a cost increase over answer-only training, but does not include narrative+ops. The evidence for a stable performance gain at comparable cost is therefore limited.

Questions for authors
  1. Does the difference between TRACE and narrative+ops in Table 14, 0.5762 versus 0.5686, persist across training runs? What are their total training and inference costs under the same evaluation conditions?
  2. In the Structure tasks, the requested operation and preservation constraints can be checked directly from the input and output. On the same candidate pools, how does selection by D compare with selection using those task checks? What additional information does the contract provide for choosing an answer?
  3. Figure 10 shows an output with D = 1.0 that still contains an incorrect claim detected by K. How often do high-D outputs contain such errors? Among the candidates selected by D, what fraction have fully correct claims, and how does this change when K is also considered?
Reviewer_35wJOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TRACE constructs Goal–Local Edit–Safeguard contracts from input/reference edits, supervises their generation before the answer, and subsequently applies outcome-only reinforcement learning. Task-specific experiments evaluate editing outcomes and trace–answer agreement on molecular benchmarks, codon optimization, and reaction repair, with separate process-reward and candidate-selection analyses.

Critical strengths

The results support a useful combination of editing performance and executable auditing. On Qwen3-1.7B, Table 1 reports a mean success–similarity score of 0.5762 versus 0.5165 for answer-only SFT+GRPO using shared answer pairs and outcome rewards. Reported molecular joint-score averages also improve across the additional model and benchmark families, alongside success–preservation tradeoffs. The contract makes declared edits directly checkable, while Tables 11–12 distinguish execution consistency from task success. Training ablations in Table 20 support the usefulness of input-matched Local Edit supervision. These empirical gains and the auditable representation support acceptance.

Critical weaknesses

The inference-time mechanism evidence remains conditional on the forced-prefix protocol. Appendix C.5.3 reports free-generation SR of 0.810 and 0.747 for MolOpt and Structure, versus 0.697 and 0.480 under clean forced-prefix re-decoding. Table 21 shows clear degradation from removing Local Edit, but its Local Edit-swap 95% confidence intervals include zero in both families. The evidence therefore more clearly establishes dependence on the field within this protocol than precise edit-content dependence during ordinary generation. This limits the mechanistic interpretation without invalidating the main editing results.

Questions for authors

Clarification: What explains the free-generation versus clean forced-prefix SR gap in Appendix C.5.3? Please clarify whether the rates use the same prompts and checkpoint, and how prefix serialization and answer sampling differ. This determines how strongly Table 21 supports conclusions about edit-content dependence during ordinary generation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.