- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper investigates how supervised fine-tuning (SFT) target defects—specifically hallucinated entities outside the candidate set—trigger severe runaway repetition in long structured text generation. The authors demonstrate that this repetition surge is driven by impaired termination control, characterized by a sharp drop in the model's decision to emit list-closing tokens (i.e., ]), rather than an increased tendency for local copying or an inability to execute EOS tokens once closed. Through contrastive activation additions and sparse autoencoders (SAE), they identify a causal, supervision-dependent representation at layer 32 that bidirectionally steers list-closing behavior, substantially mitigating tail repetition without degrading extraction recall.
Critical strengths
-
The paper establishes an original connection between training data curation and generation termination control, challenging the traditional view that data quality merely concerns item-level factuality. It insightfully demonstrates that constraint-violating supervision content can directly undermine the model's high-level stopping preference, proving that content defects induce behavioral breakdown even when structural tokens and EOS positions remain perfectly standard.
-
By modeling repetition through a competing-risk survival framework that decouples reuse exposure from persistence, the authors provide a rigorous and methodologically sound analytical lens. This formulation elegantly resolves the counter-intuitive paradox of surging overall repetition despite a declining local copying tendency, offering a precise understanding of the mechanics behind runaway generation.
-
Through shared-prefix experiments, the study cleanly separates the high-level decision to close a structured list from the mechanical execution of the EOS token, locating the failure strictly at the close-or-continue juncture. The authors then deliver compelling causal evidence through activation steering—validated by random controls, layer sweeps, and SAE feature analysis—solidly linking internal representations at specific layers to generation trajectory recovery.
Critical weaknesses
-
Scope Limited to a Single Task and Model Family: The empirical study is conducted exclusively on Chinese document-level relation extraction using the Qwen3 model family (1.7B, 4B, 8B). While the findings are consistent across model sizes and training seeds, it remains unverified whether this phenomenon generalizes to other structured generation tasks (e.g., code generation, SQL parsing, dense JSON tool calling) or other LLM model architectures (e.g., Llama, Gemma, Mistral).
-
Interaction with Inference-Time Constrained Decoding Unexplored: In production settings, structured output is frequently governed by constrained decoding frameworks (e.g., Outlines, Guidance, SGLang, vLLM JSON-mode). While grammar masking cannot decide semantic stopping (since both ',' and ']' are syntactically valid), discussing or empirically demonstrating how constrained decoding interacts with this internal termination defect would substantially strengthen the paper's engineering and practical impact.
-
Practical Application of the Intervention: The activation pulse intervention is tested at oracle/pre-selected prefix boundaries (natural stops, early points, pre-reuse points). The paper does not propose an online, automated trigger mechanism to deploy this steering vector dynamically during arbitrary open-ended decoding, which slightly restricts its utility as an inference-time mitigation tool.
Questions for authors
Q1: Clarification on Grammar/Constrained Decoding: In real-world deployment of structured JSON extraction, constrained decoding engines enforce schema adherence. When both continuing (e.g., generating {,) and closing (generating ]) are grammatically valid tokens, does constrained decoding alter the relative stop margin between raw and cleaned models, or does it leave the supervision-induced gap intact?
Q2: Generalizability across Model Families: Have the authors observed similar termination degradation and stop-margin collapse on non-Qwen architectures (e.g., Llama-3 or Mistral) when trained on constraint-violating structured targets?
Q3: Deployment of Activation Steering: The single-pulse intervention effectively reduces tail repetition when applied at natural stopping points. Could the authors clarify how one might automatically detect or calibrate the intervention point during continuous runtime decoding without access to reference stopping ground truth?
4 comments
Q3 / W3: online deployment without reference stopping points
The intervention point does not require ground truth. Our anchors are complete-block boundaries, where the next token either closes or continues the list, and these can be detected from the partial JSON during decoding; "natural stop" is only a post-hoc label. We therefore applied the closing shift online at every boundary after the fourth block (App. P.3). We use its logit form because the recalibrated control shows that the direction acts at the decision point as a closing shift of about 1.5 nats (App. P.2). β = 1.5 is that measured shift and β = 2 the raw − cleaned natural-stop margin gap; neither was tuned on generation outcomes. Over all 1,106 records (two seeds):
- Raw model: repetition 33.63% → 17.99% / 15.01% (−15.64 [−17.86, −13.43] / −18.63 [−20.71, −16.50]); context-limit endings 22.51% → 4.70% / 3.16%; pair F1 +1.38 [0.71, 2.00] / +1.65 [1.03, 2.25], with pair precision unchanged. This removes 56–66% of the raw − cleaned repetition gap and 84–92% of the context-limit gap.
- Cleaned model: pair F1 unchanged (−0.20 [−0.75, 0.38] / +0.09 [−0.45, 0.64]), so the shift can be applied without knowing in advance whether a model has the defect.
- No truncation: at early anchors, where margins lie far below the decision, no condition produces a close.
We also thank the reviewer for the note on the single pulse. With a larger sample, its effect at natural stops is smaller than the 16-anchor estimate (see our response to Reviewer qBAk), which is why acting at every boundary is the deployable form.
We sincerely thank the reviewer again for these insightful questions, which reflect a deep understanding of how structured generation is deployed in practice. They led us to experiments that substantially strengthen the paper’s generality and practical value. A second model family, Gemma3-4B-it, reproduces the termination degradation and the stop-margin gap. Constrained decoding removes out-of-candidate content but leaves most of the termination gap intact. An online closing shift at every block boundary, which needs no reference stopping points, removes most of the excess repetition and context-limit endings without loss of F1. We hope these results address the reviewer’s concerns, and we would welcome any further discussion or suggestions. If the reviewer finds the additions convincing, we would be grateful if this could be reflected in the final assessment. Thank you again for your encouraging and constructive review.
Dear Reviewer 4eR9,
Thank you again for your valuable comments on our paper. As the author-reviewer discussion period ends today, we would like to kindly check whether our response has addressed your concerns. Thanks again for your time and help in improving the quality of this paper.
Best regards, The authors
Thank the authors for the response. Most of my concerns have been addressed, and I have decided to maintain my original score of weak acceptance.
We sincerely thank Reviewer 4eR9 for a thoughtful and encouraging review. We are grateful that the reviewer recognizes "an original connection between training data curation and generation termination control" that challenges "the traditional view that data quality merely concerns item-level factuality"; finds the competing-risk framework "a rigorous and methodologically sound analytical lens" that "elegantly resolves the counter-intuitive paradox" of rising repetition with declining local copying; and considers the shared-prefix and steering results "compelling causal evidence". We also thank the reviewer for rating the soundness without substantive concerns. The three questions concern generality and deployment, and we answer each with a new experiment (new Appendix P, in blue).
Q2 / W1: other model families
We repeated the 4B pipeline with Gemma3-4B-it, using the same supervision files and SFT recipe and Gemma's recommended decoding (T = 1.0, top-p 0.95, top-k 64), just as Qwen3 used its own (App. P.1). The termination degradation and stop-margin collapse reproduce:
Absolute rates are lower for Gemma under its own decoding, so we compare paired differences. One observation is family-specific: fine-tuning on cleaned data lowers Gemma repetition relative to the untuned model, so the untuned-to-cleaned observation of §4.4 applies to Qwen3. Among other structured tasks, JSON tool calling is the closest analogue (a list whose elements must come from a closed set given in the input), and we leave it to future work.
Q1 / W2: constrained decoding
Constrained decoding leaves the supervision-induced gap intact. A grammar mask only removes tokens; at a block boundary both the close and the continue token are grammatical, so the stop margin is unchanged. We verified this with xgrammar on all shared-prefix anchors: under a JSON schema of the target format, both decision tokens are allowed at all 256 (4B) and 255 (8B) reachable anchors, so the natural-stop raw − cleaned margin gap remains −1.98 [−2.20, −1.77] nats at 4B and −1.88 [−2.09, −1.68] at 8B. In generation with vLLM structured outputs (4B, 1,106 records × 2 seeds):
Restricting source and target to each record's candidate list removes every out-of-candidate block and closes the triple-F1 gap (−1.14 → −0.06), yet 86% of the repetition gap remains (29.6% vs. 5.5%). Constrained decoding fixes what is written, not when to stop, which is why we locate the defect in learned termination.