acceptodds
Withdrawn from ICLR 2027

Silent Compliance: Separating Constraint Echoing from Instruction Following

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Modern language models increasingly follow negative instructions well. Yet even a compliant response may restate the rule, announce compliance, or reintroduce the excluded object. Such unnecessary control language can distract users, but prior work largely evaluates only whether constraints are followed. We call this behavior constraint echoing and introduce a framework that separates visible echoing from substantive instruction following. Then we test whether echoing serves as a functional scaffold. Controlled prefix interventions show that retaining an opening echo does not help organize the answer and lowers task completion and constraint satisfaction. We next explore whether echoing is inherited from the base model or shifts at particular post-training transitions. Across sampled release transitions, SFT improves instruction following and is associated with lower echoing, whereas DPO releases exhibit consistently higher echoing. Paired likelihood tests further suggest that models can become more sensitive to a constraint without becoming more likely to echo it. Finally, we introduce Matched SFT, which trains on silent, task-successful targets selected within matched prompts and mitigates constraint echoing at low cost. Together, these results identify an overlooked failure mode, explain its functional and training-stage properties, and provide a practical way to remove it without weakening constraint following.

Primary area:
foundation or frontier models, including LLMs
Keywords:
instruction following, large language models, negative constraints, post-training, preference optimization, evaluation

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_m6aPOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies a phenomena called “constraint echoing”, where LLMs will repeat instructions given in the prompt. They argue that echoing degrades task performance, and introduce a fine-tuning method to limit constraint echoing that relies on matched pairs with and without a negated instruction.

Critical strengths

This question of whether repeating instructions in the prompt is helpful to an LLM is an interesting one, and I think the proposed method of matched training on matched pairs to reduce the echoing effect is likewise interesting and if presented more clearly could be a helpful technique.

Critical weaknesses

The major weakness of this paper is that it is densely written and hard to read. The text reads as heavily AI-assisted, and there are many unhelpful phrases, jargon, etc. that make understanding the experiment details and contributions of the paper difficult. Along with this, the main paper contains very little detail about the actual data used for experiments with only a toy example in Section 3 and a minor explanation in Appendix F mentioning data is derived from WildChat conversations. I would recommend rejection based on the lack of clarity of the paper alone.

The related work is also fairly sparse, and citation practices could be improved. For example Appendix F mentions that they derive their data from WildChat, but Zhao et al.’s WildChat is not cited. Another paper that came to mind as I read that could be worth citing/comparing to is Huang et al. which identify a related problem where previous assistant replies could introduce bugs when a user updates their question after an assistant response.


Huang et al. Do LLMs Benefit From Their Own Words?

Zhao et al. WildChat: 1M ChatGPT Interaction Logs in the Wild

Questions for authors

Can you clarify some of the language choices in this paper? There is a lot of jargon that I think could be clarified, but I will point out a couple that might be a helpful starting point. What is the difference between a "request cluster" (line 196) and a "family" (line 201) for the single vs. multi-turn experiment setup? Is "joint endpoint" (line 182) the same as "joint utility" (lines 214, 321, etc.), and if so can you give some intuition for what this is trying to measure? It seems like if any one of the 4 conditions fails, this score will be 0 and so it's not clear how helpful it is.

Reviewer_Wm4UOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces the concept of "constraint echoing," which occurs when a language model successfully follows a negative constraint (like "don't mention X") but unnecessarily narrates its compliance (e.g., "I will not discuss X"). The authors argue that while standard benchmarks measure whether a rule is followed, they ignore this extraneous visible trace. Through experiments across different model training stages, the paper shows that echoing is distinct from constraint sensitivity and proposes a "Matched SFT" method to teach models "silent compliance" without degrading their ability to follow instructions.

Critical strengths
  • The paper disentangles substantive constraint compliance from the stylistic, user-visible artifact of constraint echoing. By doing so, it identifies a genuine blind spot in standard instruction-following benchmarks that conflate the two.

  • The "replay intervention" provides a highly effective, controlled mechanism to prove that echoing is not a necessary cognitive scaffold. Demonstrating that retaining a naturally occurring opening echo actively degrades subsequent task completion (e.g., reducing joint utility by 22.3 percentage points) grounds the paper's claims in concrete behavioral outcomes.

  • The proposed "Matched SFT" approach, i.e., leveraging permission-flip supervision to teach selective omission, is an actionable solution that reduces unnecessary traces without hurting core task performance.

Critical weaknesses
  • Narrow Experimental Scope: The study evaluates a very specific scenario, focusing almost entirely on actionable negative constraints (e.g., "do not discuss"). The paper does not address how this framework applies to positive or more complex constraints, which are common in real-world applications and would likely have much broader implications for the study's conclusions.

  • Critical methodological details, including prompt clustering, judge instructions, and metric definitions, are heavily relegated to the appendices. This obscures the experimental setup and makes the main text difficult to follow.

  • Over-Reliance on LLM-as-a-Judge: "Unnecessary" language is a highly subjective, semantically demanding categorization. While the authors calibrated on 50 human examples and checked 100 final decisions, the core results rely entirely on automated consensus, which may miss nuanced edge cases.

Questions for authors

While not a strict requirement, it would be highly interesting to see if this phenomenon holds for proprietary SOTA models. I suspect the baseline occurrence of constraint echoing might be much lower in these models. Would you consider reporting results on even a small subset of prompts for top-tier proprietary models to better contextualize the findings?

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

The work studies the "echo" constraint instruction that can be removed and may receive better performance after removing.

Critical strengths
  1. Instruction-following work asks whether a constraint was respected; this paper isolates a different, reader-visible question — whether a fully compliant answer still leaves a removable trace of the rule. The operational test (if a span can be deleted while the answer stays correct, complete, coherent, and format-compliant, it is an echo) is simple, transferable, and easy to apply. A natural objection is that restating a constraint acts as a useful scaffold. The paper tests this directly by replaying naturally occurring opening echoes with the suffix held byte-for-byte fixed.
  2. SFT reduces echoing and improves task behavior, while DPO releases consistently raise echoing (+5.28, +4.68, +3.12 pp across OLMo, Tulu, Zephyr), with both judges agreeing on direction. This is a practically consequential result for anyone choosing a post-training recipe.
Critical weaknesses
  1. The protocol was calibrated on roughly 50 human-labeled examples and spot-checked on 100 sampled decisions per stage. That is reasonable engineering, but C is low-prevalence and semantically demanding, and the authors themselves concede absolute rates should be read as experimental measurements rather than deployment prevalence.
  2. All checkpoint comparisons are against bundled public releases where data, objectives, templates, schedule, and formatting co-vary. The paper is commendably explicit that this is not causal — but that also means the most striking finding (DPO raises echoing) has no identified mechanism.
Questions for authors

Please reconcile "285 usable strata" with the breakdown that 101 of 285 strata end at the opening boundary and only 184 have a nonempty suffix. What fraction of the continuation-only analysis rests on the 184 suffix-bearing cases?

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

Related papers

Discussion (0)

Sign in to comment.