acceptodds
Under review as a conference paper at ICLR 2027

DACO: DISCRIMINABILITY-AWARE CONTENT OPTIMIZATION FOR AI-GENERATED CONTENT

Abstract

Large language model (LLM) evaluators increasingly rank and score AI-generated or AI-rewritten text. When powerful generative models rewrite content at scale, different texts can become more similar in wording, organization, and presentation. We call this effect content homogenization. We ask whether homogenization changes the internal features used by an evaluator even when its accuracy remains similar, and whether features that remain important after homogenization can guide content optimization. We study these questions through representation analysis using Principal Component Analysis (PCA) components of pairwise hidden-state differences to identify features that predict evaluator preferences. We evaluate three open-weight evaluators (Gemma 4 12B, Ministral 3 14B, and Phi-4 14B) across human-preference judgments, assistant-response ranking, and hiring. We further test generalization with Llama 4 Scout (109B total, 17B active), Llama-3.1-405B-Instruct, Gemini 3.6 Flash, with an additional blind human evaluation. Homogenization changes evaluator accuracy by only about one percentage point while substantially reorganizing which PCA components predict evaluator decisions. This shows that stable accuracy can hide changes in the features associated with evaluator decisions. Held-out functional validation further shows that components that remain important after homogenization can provide useful guidance for content optimization. Based on this finding, we introduce Discriminability-Aware Content Optimization (DACO), a representation-guided content optimization method that uses validated guidance for a single guided rewrite while preserving source content. DACO has zero protected-literal failures while preserving source content. These results show that representation analysis can reveal changes in evaluator features that accuracy alone misses, while also providing a practical basis for content optimization.

Primary area:
interpretability and explainable AI
Keywords:
large language models, LLM-as-a-judge, model evaluation, model interpretability, representation learning

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_6c96OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates how AI-generated content homogenization affects the behavior and internal representations of LLM evaluators. The authors analyze pairwise hidden-state differences using PCA and ridge regression to identify components associated with evaluator preferences before and after LLM rewriting. They find that evaluator accuracy remains largely stable despite substantial changes in the predictive importance of representation components. Building on this analysis, the paper proposes DACO, a representation-guided content optimization method that converts selected components into actionable rewriting guidance. Experiments across human-preference, assistant-response ranking, and hiring tasks evaluate these representation shifts, their causal relevance, and DACO's effectiveness against rewriting baselines, with additional cross-model and human evaluations.

Critical strengths

I think the paper addresses a timely and relevant problem at the intersection of AI-assisted writing and LLM-based evaluation. The experimental setup is reasonably comprehensive, covering multiple evaluation domains and model families. The authors also incorporate thoughtful experimental controls, including position-bias mitigation, held-out data splits, and explicit protected-literal preservation checks. I also appreciate the author provides detailed implementation and statistical reporting, supporting reproducibility.

Critical weaknesses

1. Insufficient validation of the content homogenization setting.

The paper operationalizes content homogenization through standardized LLM rewriting, but provides limited evidence that the resulting texts actually become less distinguishable. Although TF-IDF similarity is mentioned as a diagnostic, systematic measurements of homogenization are not reported, and there is no controlled comparison against general rewriting without homogenization. Consequently, the observed representation shifts could reflect rewriting-induced distribution changes rather than homogenization specifically. Furthermore, the original preference labels in SHP and OASST1 are reused after rewriting, even though changes in writing quality or presentation could alter human preferences without changing factual content. Without validating label consistency, the claim that evaluator accuracy remains stable under homogenization is also difficult to interpret.

2. Predictive importance shifts do not establish changes in evaluator decision mechanisms.

The paper interprets changes in PCA component importance as evidence that homogenization alters the features associated with evaluator decisions. However, permutation importance is distribution-dependent: changes in feature variance or correlations can reorder importance rankings even when the evaluator's underlying decision rule remains unchanged. Using a shared PCA basis does not eliminate this ambiguity. Although the authors attempt causal validation, none of the 12 component interventions achieves significance after multiple-testing correction. Thus, the experiments establish shifts in predictive associations but provide insufficient evidence that the evaluator actually relies on different decision-relevant features. This limits the interpretation of the paper's central representation-level finding.

3. Limited evidence of the effectiveness of DACO.

My primary concern is that DACO does not demonstrate a consistent advantage over simpler rewriting baselines. Across the three primary evaluators, none of its paired comparisons shows a statistically significant positive improvement. Its accuracy is only marginally higher than B0 on Gemma, lower on Ministral, and identical on Phi-4. The external evaluations also provide little additional support. Moreover, the functional validation primarily demonstrates movement along selected representation directions rather than improved ranking outcomes. Thus, it remains unclear to me whether the proposed representation-guided approach provides meaningful benefits beyond conventional content rewriting.

Questions for authors

See weakness above.

Reviewer_fbCTOpenReview ↗
Rating
2(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how “content homogenization” (LLM rewriting that makes texts more alike) affects LLM evaluators. It extracts pair-difference hidden states from three open-weight evaluators (Gemma 4 12B, Ministral 3 14B, Phi-4) on SHP, OASST1, and Djinni. It fits PCA + ridge surrogates to the preference margin and compares the permutation importance of PCA components before and after homogenization. Preference accuracy changes by about 1 pp, while component-importance rankings are reordered.

The authors then propose DACO. DACO turns components that keep or gain importance into natural-language guidance, validates the guidance by its held-out projection movement, and applies one guided rewrite. Further evaluations cover Llama-4-Scout as a second homogenizer, Llama-3.1-405B, Gemini-3.6-Flash, RAF, and a two-annotator human study.

Critical strengths
  • The question matters in practice: LLM judges increasingly score AI-polished content, and accuracy alone may hide shifts in what drives their decisions.

  • The experimental hygiene is careful in several places. Both presentation orders are evaluated, the test split is held out from all selection, surrogates must pass a validity gate before interpretation, random orthogonal controls and BH-FDR are used, and protected literals are preserved mechanically (zero failures).

  • Negative results are reported transparently: null causal effects, the failed 405B surrogate, and DACO not beating baselines.

Critical weaknesses
  1. In the paper, the homogenization manipulation is not verified and may contradict the premise. The paper never shows that rewritten pairs become more similar. TF-IDF similarity is mentioned but not reported. The representation-norm ratio is above 1 for every evaluator, including 405B, so pair differences grow after “homogenization.” Without a manipulation check, the effect cannot be attributed to homogenization rather than to rewriting in general.

  2. The causal evidence is null, so the methodological chain is unsupported.

  • None of the 12 interventions is significant, and the 405B surrogate fails (R² < 0).
  • The fallback, functional validation, is weak. It uses n = 12, a threshold of “mean movement > 0” with no CI, and no comparison against random or generic guidance.
  • Only 34 of 96 statements pass. That is below the roughly 50% expected if movement were pure noise, which suggests the guidance is no better than chance.
  1. There is no null baseline for the “representation shift” claim.

  2. DACO shows no benefit. It produces 0 significant improvements over any baseline. It is significantly worse than B0 on Gemma, and below B0 on Ministral accuracy (63.89% vs. 66.67%). And “Remains competitive” overstates a null-to-negative result, and the stated contribution (RQ3) is not supported.

  3. The practical insights are generic and partly contradictory. Insight 2 (“thorough, detailed”) conflicts with Insights 3–5 (“concise,” “brief”). Four of the five insights have no link to a validated component. They read as common writing advice rather than findings derived from representations.

  4. The human evaluation is uninformative.

  5. Several points need clarification:

  • Section 3.3 says Ministral is used “for guidance generation,” but later text says gpt-oss-120b writes the guidance.
  • A1 and A2 prompts refer to “representation properties” without saying how those are turned into text.
  • It is unclear which candidate(s) are rewritten in SHP/OASST1, and why label-agreement accuracy is the right objective for an “optimization” method.
  • The title emphasizes “discriminability,” but DACO does not target discriminability.
Questions for authors
  1. Can you report a direct homogenization measure (TF-IDF or embedding similarity within pairs and across the corpus) before vs. after rewriting? How do you reconcile it with norm ratios above 1?

  2. For functional validation, what pass rate do random or generic guidance statements achieve under the same >0 criterion and n = 12?

  3. DACO guidance comes from Ministral’s PC2. Why should it transfer to Gemma and Phi-4, whose PCA bases differ? Did you derive evaluator-specific guidance?

  4. In SHP and OASST1, are both candidates rewritten by DACO, or only one? What exactly is being optimized when the metric is agreement with the human label?

  5. Given the null causal results and the null DACO results, which claims do you consider supported, and would you consider reframing the paper as an analysis or negative-results study?

1 comment

Thank you for the detailed response and the additional analyses. I appreciate the transparency, in particular the explicit acknowledgment that TF-IDF does not show lexical convergence, and that no causal claim is made. The clarifications on who generates the guidance, which candidates are rewritten, and the new decision-flip counts are helpful. Most of my concerns have been addressed, and I have raised my score to 2.

By the way, the layout of the experimental results table in the paper could be made clearer.

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

The paper is concerned about homogenization: if all job applicants use ChatGPT to polish their resumes, they will start to sound alike, and the judge might get less accurate in picking out the good candidates. They study two subquestions under this: does homogenization change the judge's decision, and can we create a tool that create resumes that stand out after polishing, despite homogenization?

For the first question, they do the following:

  • Take two original resumes (natural condition)
  • Use gpt-oss-120b to polish them (homogenized condition)
  • Run judge on both, record preference & hidden state difference (one layer)
  • Use PCA to reduce hidden state difference to 512 components
  • Fit linear regression to predict judge preference from 512 components
  • Test the robust components via shuffling
  • Main finding: accuracy barely changes after polishing, but the components that predict the judge’s choice get reshuffled.

For the second question, they propose DACO, which does the following:

  • Pick a component that stays important after polishing
  • Turn that direction into English advice by prompting gpt-oss with positive and negative examples. The advice is roughly "concrete, quantitative, results-oriented details.”
  • Rewrite 12 held-out resumes using the advice and check whether their hidden states move along that direction. Keep advice if yes.
  • Take a polished resume, then rewrite it once more following the advice without inventing facts
  • Compare DACO with other rewriting strategies (plain polish, best-of-4 polish, “make it stand out” prompting, random-direction advice) and with RAF, a separate adversarial ranking attack
  • Main finding: Selected components don't seem to cause the judge’s decisions, and DACO doesn’t beat the simpler rewriting baselines.
Critical strengths
  • I thought the experiment design is genuinely interesting, and quite rigorous. The selection of advice is wrt whether it recovers the hidden state change, not cheating on judge decision.
  • Baselines are strong, in particular, the best-of-4 posh and RAF adversarial attack can be argued as slightly unfair baselines. I appreciate that authors chose these strong baselines.
Critical weaknesses
  • The empirical performance isn't very impressive.
  • The experiment design is quite complicated and involve a lot of design choices. I suspect some specific design choices might have contributed to the lack of empirical performance, assuming the intuitions are correct in the first place.
  • For example, I'm not sure why PCA + linear regression setup is the correct one to validate the first research question. It's unclear why the discovery of influential dimensions despite homogenization should be learned from the hidden state difference.
  • Prompting a model to convert PCA components into "advice" seems unlikely to work, which is perhaps why the advice looks generic. I'd consider some alternative pipeline like JEPA, to find advice that activate judge dimensions that remain robust despite homogenization.
Questions for authors
  • Have you considered directly turning influential dimensions into advice, not through the few-shot examples?
  • Have you considered alternative ways to finding "robust" dimensions despite homogenization, instead of trying to fit the difference?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.