acceptodds
Under review as a conference paper at ICLR 2027

Learning What Not to Say: Conditional Semantic Control in Language Models

Abstract

Language models can follow instructions about what to say, yet remain unreliable at controlling what not to say. We study semantic omission: completing a writing task while keeping an instructed semantic dimension absent from the generated text. Unlike surface-level constraints, semantic omission requires preventing the target semantics from reappearing through paraphrases, implicit references, and rhetorical structures, while preserving the utility of the remaining content. We construct a benchmark of 400 writing tasks across six domains and four writing settings, and systematically evaluate semantic omission across multiple model families. Our diagnosis reveals a behavioral gap between explicit boundary identification and artifact-level omission: models may correctly identify what should be excluded, yet still reintroduce the excluded semantics during generation. Thus, successful boundary identification alone is insufficient for reliable semantic omission. Motivated by this behavioral dissociation, we formulate the repair problem as conditional semantic control and propose SelectOmit, a post-training method that teaches models to selectively control whether a semantic dimension is expressed according to the current instruction. It jointly encourages suppression of the excluded semantics, preservation of the remaining task content, and recovery of the same semantics when explicitly requested. Experiments across multiple model families show that SelectOmit substantially improves semantic omission while preserving task utility, and enables reliable instruction-conditioned inclusion and exclusion of the same semantic content.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Semantic Omission, Negative Instruction Following, Controllable Text Generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_LiLbOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper studies semantic omission, where a model must complete a writing task while keeping a specified semantic dimension out of the output. Paraphrase, implication, contrast and similar rhetorical routes all count as leakage. The authors build a benchmark of 100 LLM-generated scenarios across six domains, each instantiated in four formats (Direct, Open-ended, Rewrite, Multi-turn), for 400 tasks. Outputs are scored by an LLM judge on omission and task fulfillment.

The diagnosis covers five open-weight models in the 4B to 9B range, plus API models in the appendix. Its main findings are three. Failures are common and concentrate in Rewrite Removal. Models that correctly articulate the boundary in a separate probe still leak. Clearer prompts, including an Oracle boundary, reduce but do not remove leakage.

The proposed method, SelectOmit, is a LoRA post-training objective with four terms. These are token-level unlikelihood on annotated leakage spans, cross-entropy on clean responses, a DPO-style preference of clean over utility-degraded responses, and span-level cross-entropy on counterfactual inclusion responses. On Qwen3-4B, valid success rises from 80.0% to 97.5%. Paired include/exclude validity reaches 81.5%, against 62.75% for standard DPO, and four-way compositional success reaches 74%. Results are supported by human validation, ablations and a 50-document authentic-source evaluation.

Critical strengths
  • The paper reframes negative instruction following as instruction-conditioned semantic control. Prior work asks whether a model understands a negated request. This paper asks whether the same model can make the same content absent or present depending only on the current instruction, which is a stricter and more practically meaningful target. Defining success over paired include/exclude conditions makes over-suppression measurable for the first time in this setting. Standard DPO looks strong on exclusion (95.75%) yet fails on inclusion (64.5%).
  • The experiments are comprehensively designed. Each loss term targets one failure mode the formulation exposes. Span-matched Conditional SFT and Positive-matched SFT receive the same inclusion data as SelectOmit, which rules out the simplest explanation that the gain comes from extra counterfactual data. Moreover, blinded human evaluation, authentic-source tests and a stratified 600-response judge validation further strengthen the empirical case.
Critical weaknesses
  • The methodological novelty appears somewhat modest. The conditional-control framing is conceptually close to the bidirectional objective of RePS (line 464), and the paper could explain more clearly why applying it through ordinary instructions calls for a new method. SelectOmit combines four established components with fixed weights, with little analysis of why this combination yields conditional control, and the utility-preference term shows no significant effect (+1.25, CI [−1.00, +3.75]).
  • The evidence for the two central claims could be strengthened. The identification probe is a separate call, and its rates (60.75% to 78.75%) are lower than valid success (80.42% to 95.25%). Without comparing leakage between identified and non-identified items, the knowing/doing dissociation is hard to separate from probe noise. On repair, the exclusion gain over DPO is +1.0 pp with a CI including zero. DPO saw no inclusion pairs, so its weaker inclusion is somewhat expected, and a bidirectional DPO baseline would be more convincing. SelectOmit also lowers inclusion validity from 92.25% (Base) to 83.75%.
Questions for authors
  • How does SelectOmit compare with DPO trained on both exclusion pairs (clean over leaking) and inclusion pairs (target-present over target-absent)?
  • What is the leakage rate among items where the probe failed to identify the boundary? If it is close to the rate among identified items, how do you distinguish a knowing/doing dissociation from a probe that is only weakly coupled to generation?
  • Every trained method lowers inclusion validity below Base. Can you characterize what SelectOmit loses on inclusion, and whether it suppresses neighboring semantics as well?

If the authors can address the weaknesses and questions raised above, I would raise my score.

Reviewer_xEANOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces the task of semantic omission and proposes diagnostic tools and a fine-tuning method that teaches LLMs when and how to omit semantic information appropriately. The experimental results are good evidence for the success of semantic omission but the rational of the proposed method is unknown and might rely on shallow heuristics.

Critical strengths

A very interesting and challenging problem

Critical weaknesses

The rationale for the proposed method lacks sufficient empirical validation and grounding in prior work. In particular, it remains unclear why and how a method based on distributional semantics can capture the inherently pragmatic nature of implicit reference. I appreciate the authors' efforts; however, I am not fully convinced by the central claim unless this conceptual gap is adequately addressed. This is my primary concern.

Questions for authors

1, what are exactly the differences and connections among boundary identification, diagnosing semantic omission and condiational semantic control? what is the rational underlying them in terms of proposing a method about conditional semantica control?

  1. how could we know LLMs exactly know about how to do semantic omission but not superficial heuristics through the statistical association among tokens?

  2. references and rhetorical strucutures are pragmatic knowledge. to what extent the authors believe that LLMs exaclty understand those pragmatic knowledge during the process of semantic omission and how to prove that?

  3. LLMs are skept at distributional semantics, but implicit reference lies in pragmatics, there is a huge gap between LLMs' capabilities and the required linguistic cues of "implicit references". to what extent the authors believe the pragmatic cues of implicit references can be captures by the learning objectives in equaion (6) without a real pragmatic inference process?

Reviewer_vbQ9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies semantic omission where it completes a writing task while excluding a specified semantic dimension. It introduces 400 tasks from 100 scenarios and proposes SELECTOMIT, which combines leakage-span unlikelihood, clean-response supervision, utility preferences, and counterfactual inclusion supervision. Its central evaluation pairs instructions to exclude or include the same semantic content. The results expose a limitation of omission-only evaluation where an adapted model may improve at excluding content while becoming worse at including it when requested.

Critical strengths
  1. The paired evaluation reveals a consequential trade-off. Standard DPO achieves 95.75% exclusion validity but only 64.50% inclusion validity, resulting in 62.75% paired validity. SELECTOMIT reaches 81.50%, compared with 74.00% for Base and span-matched conditional SFT (Table 4). The positive-matched and span-matched SFT controls are useful: supplying inclusion examples or localizing inclusion supervision alone does not reproduce the full result under the tested recipes. The paired gain over Base is +7.50 percentage points, with a scenario-cluster bootstrap 95% interval of [+2.50, +12.50] (Table 23).

  2. Human evaluation and authentic-source transfer corroborate the main finding. Blinded human evaluation reports paired validity of 82% for SELECTOMIT, 74% for Base, and 63% for DPO (Table 16). On 50 authentic-source documents, SELECTOMIT achieves 80% paired validity, versus 60% for Base and 58% for DPO; both comparisons have Holm-adjusted p=.0137 (Tables 18 and 20). These results strengthen the evidence beyond performance on a synthetic benchmark scored by a single automated evaluator.

Critical weaknesses
  1. The method’s advantage is not separated from its richer supervision. SELECTOMIT combines established objectives to address specific failure modes. The important question is whether this combination outperforms a simpler alternative given the same information. Standard DPO receives exclusion preference pairs, whereas SELECTOMIT additionally receives inclusion targets and utility negatives, with a different initialization and optimization recipe (Appendix C.4.1). The existing SFT controls do not answer whether conditional preference training with comparable supervision achieves similar paired behavior.

This comparison is particularly important because the exclusion-only advantage over DPO is just +1.00 point, with a 95% interval of [−1.25, +3.25] (Table 22). The main advantage is conditional behavior, yet the DPO baseline is not trained with the corresponding conditional supervision. A DPO-plus-inclusion comparator using the same utility information and a comparable tuning budget would directly test the incremental value of SELECTOMIT’s objective. The paired uncertainty contrast against span-matched conditional SFT is also missing from Table 23.

  1. The 384-token evaluation does not establish sustained control in longer writing. All evaluated writing is capped at 384 completion tokens (Appendix B.4). This limits the evidence to short-form generation. It does not establish whether the model maintains exclusion across a longer article, avoids reintroducing the target in later paragraphs, or preserves the coverage required by a substantial writing task. The authentic-source evaluation also omits the original word-count requirement (Appendix C.9.2).
    Actual output lengths and truncation rates are needed to interpret the current results. Longer tasks should require complete, substantive responses and evaluate omission together with content coverage and paired validity. Merely increasing the token cap would not resolve this concern if the tasks still elicit short answers.

  2. Inclusion recovery remains incomplete, and its cross-model generality is unclear. Inclusion validity drops from 92.25% for Base to 83.75% for SELECTOMIT. Human evaluation shows the same pattern: 92% versus 84% (Tables 4 and 16). The method improves the inclusion–exclusion trade-off, but does not preserve the base model’s inclusion performance. An analysis of previously successful inclusion cases that become invalid would clarify what is being lost.

The paper is not limited to Qwen3-4B: Table 3 reports omission improvements on Qwen3-8B, Llama-3.1-8B, and Gemma-2-9B, and Appendix B.3 diagnoses stronger API models. However, these results do not replace a cross-model comparison of paired validity and inclusion regression. The detailed conditional-control evidence remains concentrated on Qwen3-4B. This matters when some stronger base checkpoints already perform well—for example, Gemma-2-9B improves from 95.50% to 97.25% on omission validity.

Questions for authors
  1. Does SELECTOMIT retain its paired-validity advantage over DPO plus the same inclusion supervision and utility information, with comparable optimization and hyperparameter-search budgets? This is the comparison most likely to change my assessment of the method contribution.

  2. What are the actual output-length distributions and truncation rates? Does the advantage persist on tasks requiring longer, complete responses, with the original length requirements retained in evaluation?

  3. On stronger open-weight models, does omission-focused adaptation produce comparable inclusion regression, and does SELECTOMIT still improve paired validity over matched baselines?

  4. Which inclusion cases that Base handles correctly become invalid after SELECTOMIT? Please also report the paired uncertainty contrast against span-matched conditional SFT.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.