acceptodds
Withdrawn from ICLR 2027

When Agreeing Context Hurts: Sequence Pairing in Supervised Fine-Tuning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Evidence in a neighboring contract can agree with a target judgment yet make a model less likely to answer it correctly. On ContractNLI, deleting that neighbor's target-consistent evidence corrects many pre-existing errors. We study how training-time sequence pairing shapes this behavior. Across supervised fine-tuning (SFT) arms, we keep complete judgments, answers, and optimizer-update memberships fixed while changing which judgments share a sequence. ProofWriter establishes the controlled comparison: pairing changes delivered answers, complete training layout changes the contrast, and 4B checkpoints selected by single-question accuracy trail alternatives by more than ten points on paired requests over a separate set of SFT-unseen theories. In ContractNLI, the original pairing rule deliberately guaranteed disagreement on the target hypothesis. Before new training, we predicted that weakening this relation would attenuate the observed context and deletion effects. Re-pairing the same judgments raises training agreement from zero to about ; in two independently trained repaired checkpoints, the roughly twenty-point accuracy deficits with agreeing versus conflicting neighbors move near zero, and deleting target-consistent rather than neutral text yields exactly zero change in target accuracy, while document-conditioned paired-task competence is retained. The repair also incurs condition-specific losses. Sequence partners can shape how fine-tuned models use neighboring information, and checkpoint comparisons should reflect the requests expected at use time.

Primary area:
foundation or frontier models, including LLMs
Keywords:
supervised fine-tuning, training data organization, sequence packing, context-dependent learning, evaluation transfer

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_pGDaOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper seems to be about natural-language understanding, but unfortunately my own understanding of the paper cannot surpass this basic level. The abstract is convoluted and obscure, and even after a slow read of the first two pages—including the entire introduction—I still hadn’t learnt about the setting, challenge, or contributions.

The paper is clearly AI-generated, but somehow in a much worse way than one could expect from today’s frontier models.

Critical strengths

None

Critical weaknesses
  • The paper is impossible to understand for somebody not familiar with ContractNLI, ProofWriter, etc. No attempt is being made to introduce the domain, motivate a task, describe challenges, etc.
  • As an example, these are the first sentences of the abstract: “Evidence in a neighboring contract can agree with a target judgment yet make a model less likely to answer it correctly. On ContractNLI, deleting that neighbor’s target-consistent evidence corrects many pre-existing errors.” The abstract contains terms such as “contract”, “ContractNLI”, “optimizer-update memberships”, “ProofWriter”, etc., without any attempt to paint a bigger picture.
  • Overall, this paper remains too arcane for me to make a judgment about its technical content. It is so poorly structured and written that I cannot even grasp its most basic aspects.
Questions for authors

What is the paper about?

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

The paper studies how pairing complete judgments during supervised fine-tuning changes subsequent use of neighboring context. It holds judgment occurrences, answers, and optimizer-update memberships fixed while changing sequence partners. On ContractNLI, it finds that agreeing neighboring evidence can reduce target accuracy. A prospectively motivated re-pairing intervention largely removes this effect while retaining document-conditioned task competence.

Critical strengths

S1: The controlled rewiring design isolates a consequential training choice that is easy to overlook. Keeping complete judgments and their update memberships fixed makes the comparison more informative than changing the training corpus or curriculum wholesale.

S2: The prospective repair is compelling. The authors identify a relation imposed by the original pairing rule, predict the direction of an intervention, and test it in new training realizations. The attenuation of roughly twenty-point context effects provides stronger evidence than a retrospective analysis of unusual model behavior alone. This construction-level finding is the main reason I lean toward acceptance.

Critical weaknesses

W1: The intervention establishes a construction-level effect more clearly than an effect attributable specifically to agreement. Re-pairing changes companion identity, wording, length, and local layout alongside the agreement relation. The original ContractNLI construction also deliberately enforces disagreement, making the broader relevance to ordinary SFT pairing uncertain. The paper acknowledges these boundaries, but they limit how generally the repair can be interpreted.

Questions for authors

Q1: Which part of the evidence distinguishes sensitivity to the agreement relation from sensitivity to the other companion features changed by re-pairing? Please clarify the narrowest causal claim supported by the existing comparisons and which conclusions depend on the deliberately anti-correlated training construction.

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

The paper studies how pairing complete judgments during SFT affects later responses. Using ProofWriter and ContractNLI, it shows that pairing and request format can change checkpoint rankings, and that weakening an imposed disagreement relation reduces harmful neighboring-context effects.

Critical strengths

We appreciate the careful matching of training occurrences and update memberships, the repeated training runs, and the transparent reporting of failed controls and repair costs.

Critical weaknesses
  1. We find the contribution too narrow. The ContractNLI setup makes neighboring contracts disagree with the target hypothesis in every original cross-contract training row. We are not convinced that this establishes a broader insight beyond sensitivity to constructed training correlations. Its relevance to ordinary SFT remains unclear.

  2. The experiments do not fully establish the proposed explanation. Re-pairing changes document identity, wording, length, and local composition alongside agreement. The intended control separating these factors was not successfully constructed.

Questions for authors
  1. Does the agreeing-context effect occur under ordinary random pairing, without deliberately enforcing disagreement?

  2. What evidence separates the effect of agreement from the other changes introduced by re-pairing?

  3. What generalizable conclusion goes beyond learning context correlations and matching training formats to downstream requests?

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

Related papers

Discussion (0)

Sign in to comment.