acceptodds
Under review as a conference paper at ICLR 2027

Selective Computation Declines After Short-Answer Continual Fine-Tuning: Elicitation and Target Format

Abstract

Standard continual learning (CL) metrics measure forgetting strictly within the new task stream, leaving a massive blind spot: the degradation of a pretrained model's prior, general capabilities. Here, we uncover a striking, highly selective failure mode induced by continual adaptation. Fine-tuning LLMs on a classification stream severely degrades their accuracy on out-of-distribution queries requiring computation, while leaving retrieval queries almost intact. Concurrently, adaptation triggers a drastic shift in stopping behavior, forcing the model to terminate immediately after brief answers. While this stopping shift and the computation loss appear causally linked, we rigorously decouple them: manipulating only the training target format completely reverses the stopping policy, yet the computational drop remains unchanged. Crucially, the underlying capability is not erased. Given a Chain-of-Thought (CoT) scaffold, the adapted model successfully recovers correct, step-by-step derivations for queries it previously failed to answer autonomously. We conclude that continual adaptation suppresses the model's intrinsic propensity to reason rather than destroying its mechanics. The model remains fully capable but becomes fundamentally unwilling to derive answers—a silent deficit standard CL metrics fail to capture.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Continual Learning, Catastrophic Forgetting, Fine-tuning, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_eBXwOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates whether current Language Model (LM) continual learning evaluations confound the model's knowledge and evaluation protocol, such as response format. In the short-answer setting, the experimental results show that fine-tuning reduces computation accuracy more than retrieval accuracy. The results show that LM evaluation is sensitive to the target format and the evaluation protocol.

Critical strengths
  1. The experiment findings might be robust as experiments were conducted across three random seeds.
Critical weaknesses
  1. The paper is not self-contained. Several terms and experimental details are not included or discussed. Without these details, the experimental results, their implications, and conclusions are unclear to me. For example:
  • There is no formal definition of computation and retrieval accuracy. While the paper defines the evaluation framework in Section 2, it does not clearly discuss the meaning and details of these two accuracies.
  • The paper states that the experiments are conducted on 14 tasks. However, it does not say which tasks are tested in the main text. Without mentioning the datasets, the meaning of the experiment numbers is unclear. Also, the “200-item suite” has not yet been introduced, despite being mentioned in many experimental results.
  • Methods mentioned in the experimental results are not introduced in the main text. For instance, what are the "rational-first training", "label-only," and "prompted" methods mentioned in Table 1 and Figure 2? The main text does not clearly define these methods.
  1. If I understand correctly, this paper argues that the performance decline after fine-tuning reflects both model knowledge forgetting and response-format shift. While this is a known phenomenon (related works cited around Line 90), this paper is basically measuring this phenomenon in one specific scenario: the "short-answer setting". Meanwhile, this paper does not propose a method to mitigate or solve this problem (or, even if the paper does, the current presentation is unclear to me). Therefore, the contribution of this paper is limited.
Questions for authors

My questions are listed in Critical Weakness.

Reviewer_9SB3OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies why LLMs appear to forget after continual fine-tuning on short-answer classification tasks. The question is whether the drop on unrelated held-out questions reflects real capability loss or just a change in how the model responds. The main argument is that fine-tuning on bare labels teaches the model to give an answer and stop. This hurts multi-step computation questions, which need intermediate reasoning, but barely affects factual retrieval. In the authors' view, the ability itself is not lost, it is just no longer used by default, and where the answer sits in the training target decides whether this habit forms. To test this, the authors compare base and fine-tuned models on a self-built set of computation and retrieval questions across several backbones. They ask the base model to give only the answer, prompt the fine-tuned model with few-shot examples, and force it to keep generating after it answers.

Critical strengths

Carefully controlled interventions. Within its setting, the paper isolates one variable at a time.It varies the evaluation protocol and training target while holding the checkpoint fixed, using answer-first, rationale-first, and randomized-position arms with matched rationale content and length. This part of the experimental design is sound and internally consistent, and it allows format effects to be separated from changes in the model’s weights.

Critical weaknesses
  1. Limited novelty. The paper’s central claims are largely known or expected. It is well known that fine-tuning makes a model specialize to the output format of its training data, and that chain-of-thought mainly helps multi-step math and symbolic problems rather than factual recall. Given these facts, it is unsurprising that training on bare labels suppresses intermediate reasoning, that this hurts computation but not retrieval (which never required reasoning), and that adding rationales before the answer in the training targets preserves reasoning.
  2. The experimental design is too simple to support general conclusions. The training setting is narrow: a single stream of 14 label-only classification tasks in one fixed order, LoRA only, and models of at most 14B. The evaluation is small and non-standard, relying on a self-built suite of 100 computation and 100 retrieval questions rather than established benchmarks (e.g., GSM8K or TriviaQA). The two categories also differ in difficulty and answer form, and the difficulty-matched set in Section 3 is not used for the main results. Finally, there are no continual-learning or data-mixing baselines and no non-sequential control, even though Figure 4 shows most of the decline appears after the first task, suggesting that the continual learning setting and background is not essential to the effect.
Questions for authors
  1. As typeset, many entries in Tables 1–3 exceed 100 even though the columns are labelled as percentages or percentage points (e.g., "4027.3" and "1513.5" in Table 1, "4061.7" and "1575.5" in Table 2, "4026.3" and "4032.3" in Table 3), so the main result tables cannot be read as printed. Please correct the tables and confirm every intended value, together with its unit and the number of items and seeds behind it.
  2. Few-shot prompting improves most models, so a gain on the adapted model alone does not show recovered reasoning. What is each base model’s accuracy under the same prompt π⋆? If π⋆ also lifts the base model, the true gap is larger than the reported 50–70% “reduction” suggests. Do the gains come from correct derivations rather than from longer outputs that happen to contain the gold string?
  3. The experiments use a self-built 200-item suite. Do the findings hold for established benchmarks such as GSM8K, and can also work on standard continual-learning baselines?
  4. Which model generated the rationales used in the rationale-first targets? If it is stronger than the fine-tuned model, how much of the smaller decline relative to label-only training comes from distillation rather than from answer position?
Reviewer_JopDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper fine-tunes LLMs on 14 short-answer classification tasks and finds that multi-step computation accuracy drops much more than retrieval accuracy. It then asks whether this is lost capability or partly a change in how the model answers. An answer-only instruction plus a short output cap makes the base model show much of the same pattern, while a development-selected prompt recovers a large fraction of the adapted model’s computation score without changing its weights. The paper also compares answer-only, answer-then-rationale, and rationale-then-answer targets. Rationale-first targets leave a much smaller computation drop. Results cover five backbones and two string-based scoring rules.

Critical strengths
  1. The response-protocol controls are very good. Testing the adapted checkpoint under a different prompt and the base checkpoint under answer-only/capped conditions directly addresses an alternative explanation that many forgetting papers would miss. Reporting both a permissive gold-string score and a stricter answer-span score is also the right choice.

  2. The target-format experiment is more convincing than a generic rationale comparison. Rationale content and length are matched, answer-first is included, and randomizing answer position weakens the rationale-first benefit. The prompt is selected on development data and then tested separately.

Critical weaknesses
  1. It is hard to know how far the result generalizes from the paper’s custom 200-item suite and one stream of 14 classification tasks. The computation/retrieval split is made by construction, and the complete suite is only promised in the supplement. I would like to inspect the items, scoring variants, and category annotations before placing much weight on the size of the reported gap. A replication on an established, independently constructed benchmark would help.

  2. The replication story is uneven. The main Qwen-3B comparison uses three seeds, but the rationale-first experiments on the other 4 backbones use one seed each. Pooling item--seed results is informative, but it does not measure variation across independently trained models. Also, answer position is still tied to different token-level supervision and generation behavior. The randomized-position result is helpful, but it does not completely isolate answer position as the cause.

Questions for authors
  1. Could you release the 200-item suite, construction rubric, gold variants, and item-level model outputs during review? Can the main result be reproduced on a public benchmark using a protocol fixed before evaluation?

  2. Could you add independent seeds for the rationale-first runs on the auxiliary backbones and report checkpoint-level uncertainty? It would also be useful to vary rationale quality or content independently of answer position.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.