acceptodds
Under review as a conference paper at ICLR 2027

ARC: Learning Selective Answer Repair from Matched Outcomes

Abstract

Small language models offer efficient inference, but errors in their initial answers can limit their reliability on reasoning tasks. Self-refinement and external tools can help correct these errors, although applying repairs to every answer adds computation and can damage correct answers. The challenge is to predict which repair, if any, will improve an answer before executing it. We introduce ARC (Anchor-conditioned Repair Controller), which learns this decision from matched repair outcomes. During training, we apply all available repairs to the same initial output, called the Anchor, and use their final-answer scores to supervise retention and repair selection. At deployment, an output-conditioned hierarchical router reads the question, generated answer, and public reasoning, then retains the Anchor or executes one repair with the base model fixed. On Qwen3-8B across five reasoning task families, the main evaluation yields a mean task score of 0.6915, a 6.75% improvement over retaining initial answers. It uses 0.155 additional LLM calls and 0.348 s of recorded LLM request time per example, reducing these costs by 84.5% and 90.9%, respectively, compared with one self-refinement call per answer. A separate deterministic evaluation over three seeds yields nearly equal mean in-distribution scores with 24.8% fewer additional calls than routing without the generated answer or reasoning. Experiments with the other evaluated base models also show improvements over answer retention and question-only routing. Our code is available at https://anonymous.4open.science/r/ARC-474B/.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Small Language Models, Selective Answer Repair, Repair Routing, Outcome Supervision, Efficient Inference

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_xfJNOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes ARC, a controller that learns when and which repair to apply to an initial answer from a frozen small LLM. All available repairs are applied to the same Anchor offline, providing matched supervision. At deployment, a hierarchical DeBERTa-v3-small router conditioned on the question, answer, and reasoning retains the Anchor or executes one repair.

Critical strengths
  1. Efficient inference is an important direction.
  2. The method is clear, straightforward, and easy to follow.
Critical weaknesses
  1. The empirical gains are modest, and several simpler variants match or exceed full ARC. The results do not support the design claims.

    (a) In Figure 2, the Flat Router stays within ~0.5 points of ARC at matched budgets, and even scores slightly higher with fewer calls.

    (b) In Table 2, 4/5 of 7 variants outperform full ARC.

    (c) The simpler Family-Aware When-Only reaches 0.6997 at 0.1442 calls, surpassing ARC's 0.6886 at 0.1426 (Figure 4).

  2. The presentation of results is confusing and not organized clearly. Key evidence is scattered across the main text and multiple appendices, and different tables and figures draw on different runs.

  3. There are internal inconsistencies between results. the ARC vs. Flat Router gap is 2.23 points in Table 2 but shrinks to ~0.5 points in Figure 2. Similarly, ARC outperforms Question-Only by 1.51 points in Table 1, while the gap vanishes in the deterministic evaluation (0.6929 vs. 0.6926).

  4. The router must be retrained whenever the base model changes, which limits practical utility. It would substantially strengthen the paper to include transfer experiments or experiments with larger base models.

  5. An experiment on iterative/multi-round repair would clarify whether the framework extends beyond the single-step setting.

  6. It would be better to clarify that "availability" (line 122) refers to the actions' availability. Introduce Ĝ (line 209) explicitly as the predicted gain. Gives precise definitions of the metrics at their first occurrence.

Questions for authors

See weaknesses

Reviewer_w7N3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ARC, a controller that decides whether to retain an initial language-model answer or apply one of several repair procedures. It collects offline outcomes by applying available repairs to the same initial answer, then trains an output-conditioned hierarchical router using action-selection, gain-prediction, ranking, and auxiliary objectives. Experiments across four small language models and five reasoning task families demonstrate improvements over retaining initial answers, with relatively few additional model calls.

Critical strengths
  1. Controlled supervision and evaluation. Applying different repairs to the same initial answer provides a consistent basis for comparing their utility. Shared action outcomes and source-group-based data splits strengthen the internal comparisons.
  2. Useful efficiency gains and transparent reporting. ARC improves answer quality over retention while intervening on a small fraction of examples. The paper also reports deterministic retraining, negative ablations, and fixed-action controls, which help clarify the method’s practical benefits and limitations.
Critical weaknesses
  1. Section 2 distinguishes ARC from selective refinement and learned routing through its matched repair outcomes and heterogeneous action selection, but the experiments do not sufficiently isolate the benefits of this formulation over simpler gain-prediction or gating policies.
  2. The paper does not compare ARC with prompting the same base model to select retention or a repair procedure after inspecting its initial answer. Such a comparison under matched inference budgets is needed to establish the practical advantage of training an additional router.
  3. The method relies on a lightweight encoder to predict repair gains from the question, answer, and reasoning, but provides limited direct evaluation of these predictions. Within-task prediction and failure analyses are needed to determine whether the controller learns repair utility beyond task-specific correlations.
  4. Appendix I.2.1 reports that restricting Qwen ARC to one fixed repair per task family produces almost identical decisions, and Table 39 shows a family-aware binary gate achieving higher scores at similar realized call counts, albeit with privileged family labels. These results leave the additional value of heterogeneous action selection unclear.
  5. The component evidence is mixed. Deterministic Qwen retraining yields nearly identical ID scores for ARC and Question-Only (Appendix F.2, Table 14), while removing the auxiliary objective or SelfRefine improves mean scores in Table 2. Moreover, intervention rates differ across ablations (§7.3), and Table 25 shows comparable ARC and Flat Router scores in part of the cost range, limiting conclusions about the hierarchy’s independent contribution.
  6. Each base model receives a separately trained router, and Hard Mixed only adds fixed distractors to existing ID examples (Appendix A.2). This does not establish transfer to unseen datasets, domains, or base models.
Questions for authors
  1. Controller necessity: How does ARC compare with having the same base model directly choose retention or a repair procedure from its initial answer? Please include routing overhead and compare under matched token or latency budgets.
  2. Prediction reliability: How well do predicted gains track realized gains within each task and action? Can you provide prediction-error, ranking, and failure analyses that distinguish repair-utility prediction from task recognition?
  3. Task–tool associations: Can you add a predicted-family-plus-binary-gate baseline using the same public inputs and comparable training resources as ARC? On which tasks does choosing among multiple repairs provide a measurable advantage over a fixed family-specific repair?
  4. Generalization: Can the trained controller transfer without retraining to an unseen dataset within an existing task family? How robust are its decisions when the base model, input distribution, or repair implementations change?
Reviewer_5uuaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper introduces an Anchor-conditioned Repair Controller, which is a model that decides, given an initial draft of an answer, whether to make a repair and if so, which. This ARC model is trained by applying and scoring all possible repair operations over anchors (draft answers) created by the base model. Evaluating across four 4B-8B models and several task families find that these interventions improve task performance while reducing cost.

The router itself is a hierarchical model for the type of repair (e.g. structured vs. temporal) and training is done by optimizing the path to the best type of repair, as well as a group-level objective of whether the repairs were effective at all. Considering the problem setup, this solution seems effective – analysis and ablations verify that a router model is more effective than question/answer-only inputs to the router as well as an objective that has fewer parts.

However, the improvements are relatively minor given the complexity in curating the data and training the hierarchical router compared to a flat router or even no repairs. Some alternative hypotheses are not really compared against and it is not clear how the repair methods in this paper were chosen. Thus, I’m leaning slightly toward rejection.

Critical strengths

The experiments and ablations are detailed (i.e. 3 random seeds) and support the claim that the addition of an ARC router leads to improved performance with minimal cost -- the perspective of repair cost vs. performance gain (Fig 2) is interesting.

Critical weaknesses

The repair tools need to be defined in advance. In this work, they appear to be custom-chosen to be relevant to the tasks (and the usage of self-refine might even hurt). In practice, the types of repair operations may be more numerous and harder to determine the value of.

Why are we limited to only one self-refinement call per answer? Wouldn’t applying all applicable repairs make more sense? This would also provide a richer space of model types and possible training objectives.

It is interesting that models converge on 2-4 of their favorite tools (3c) and that there is relatively low diversity. I would have expected tasks/datasets to have different tool needs, but there is no breakdown by task type.

Questions for authors

A flat router seems to do almost as well (in Figure 2): is there a way to extend that tradeoff and spend more LLM calls (to get higher performance)? It appears the main draw of the method is a cheap and effective routing method; in that case, the flat router is most of the Pareto frontier at the moment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.