acceptodds
Under review as a conference paper at ICLR 2027

Construct Before Compose: Preserving Conditional Structure in Representation-Based Bias Mitigation

Abstract

Social biases in large language models can vary across combinations of demographic attributes, yet existing representation-based interventions often construct directions from isolated attributes and attempt to recover such variation only during composition. However, marginal direction construction can discard cross-attribute information before intervention is applied. Hence, we propose Construct Before Compose (CBC), a condition-aware framework that constructs attribute directions while preserving cross-attribute conditions and then applies composition mechanisms. We evaluate CBC on race–gender, age–gender, and age–race settings across four instruction-tuned models from three model families. Compared with single-attribute construction under the same composition rule, CBC improves the mean conditional bias reduction from 7.70% to 12.13% across the three attribute pairs. Further analyses show that mitigation on marginal attributes does not reliably transfer to conditional targets, and that conditional evaluation is necessary to reveal residual slice-level disparities and cross-attribute interaction effects. Across six general-language benchmarks, CBC maintains comparable utility, with a four-model macro-average accuracy change of -0.30 percentage points relative to no intervention. These results establish condition-aware direction construction as a fundamental design principle for reliable representation-based intervention in large language models.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
social biases, large language models, representation engineering

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_arGXOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

CBC builds condition-indexed steering directions before composition. Across three attribute pairs and four models, CBC+APC increases mean conditional bias reduction from 7.70% to 12.13%.

Critical strengths
  1. Separating direction construction from composition is a useful framing.

  2. Paired uncertainty, multiple splits, and negative results are reported.

  3. The analysis distinguishes marginal mitigation, conditional attenuation, and interaction reduction.

Critical weaknesses
  1. CBC changes prompt view, scoring target, direction count, and parameter sharing together. This supports a comparison of complete strategies, but not attribution of the gain specifically to condition indexing.

  2. Magnitude-based WMD need not identify the interaction term, and mean interaction reduction is lower than Single+APC's (3.26% versus 4.99%). This limits the mechanism claim without invalidating conditional attenuation.

  3. Results use one synthetic lexical benchmark and a fixed layer/strength; age–race gains are uncertain. General-language utility checks do not establish that attenuation reflects selective debiasing rather than flattening of relative candidate probabilities.

Questions for authors
  1. Can matched direction-count and prompt-view ablations isolate condition indexing?

  2. How do relative candidate probabilities and top-1 predictions change?

  3. Are gains robust to layer/strength choices and another bias evaluation?

1 comment
AuthorsOpenReview ↗

CBC Attribution, Candidate Behavior, Layer and Strength Sensitivity, Transfer, and Utility

Thank you for the recommendations on matched attribution, relative probabilities, hyperparameter sensitivity, and external validity. The revision now frames the main result as a comparison of the complete Single and CBC construction settings and adds candidate-probability, layer/strength, GlobalBias, and task-level utility results.

  1. Method attribution: The main experiment jointly changes condition indexing, prompt view, scoring target, bank capacity, and parameter sharing. The revision therefore uses the complete construction setting as the comparison unit and removes attribution of the 4.44-point difference to any single module. The six-composer fixed-bank control measures sensitivity to composition, and the two-composer GlobalBias matrix further quantifies the construction-by-composition interaction.

  2. Relative probabilities and Top-1 changes: CBC+APC relative to No-Intervention yields a 19.00% Top-1 flip rate (95% CI [17.49, 20.55]) and mean total variation of 0.1409 (95% CI [0.1352, 0.1467]). CBC+APC relative to Single+APC yields a 24.16% Top-1 flip rate (95% CI [22.40, 25.92]) and mean total variation of 0.1874 (95% CI [0.1808, 0.1940]). The intervention thus changes both candidate ranking and relative probability allocation.

  3. Layer and strength sensitivity: The retained five-checkpoint race–gender grid evaluates absolute blocks 16, 20, and 24 at strength 1.0. Mean CBC+APC Cross4 is 8.22%, 9.11%, and 6.20%, respectively, with observed minima of −0.10%, 2.57%, and −4.16%. At block 20, strengths 0.5, 1.0, and 1.5 yield means of 6.58%, 9.11%, and 10.45%, respectively, with observed minima of 2.41%, 2.57%, and 2.65%. Cross4 remains positive across the tested strengths at block 20. Changing the layer produces checkpoint-level failures at blocks 16 and 24. The historical DAC-Cross4 label in the archived grid corresponds to APC-Cross4 in the revision.

  4. External transfer GlobalBias compares APC and static equal composition with the source banks fixed. Across five checkpoints, the mean under APC is 4.06% for Single and −2.43% for CBC. Under static equal weighting, the corresponding means are 0.60% and 4.99%. With static equal weighting, CBC exceeds Single on Llama-3-8B and Mistral-7B and trails it on the other three checkpoints. This matrix shows how transfer varies with construction, composition, checkpoint, and evaluation distribution.

  5. General-language utility: Across six tasks and four models, the macro changes relative to No-Intervention are −0.463 points for Single and −0.299 points for CBC. CBC decreases by 2.02 points on TruthfulQA and averages a +0.045-point change across the other five tasks. The revision reports both model-level and task-level results so that bias attenuation and general-language utility can be assessed under the same experimental accounting.

Reviewer_5dxYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Construct Before Compose (CBC), a bias-mitigation framework that constructs condition-specific representation directions and combines them for intervention. Experiments show improved conditional preference-contrast attenuation in some settings, with relatively small average utility losses. However, the effectiveness varies across attribute pairs, and further validation is needed to establish the proposed mechanism.

Critical strengths

The paper identifies an intuitive limitation of marginal bias mitigation: average preferences may fail to capture differences in preferences across attribute conditions.

The paper separates the construction of conditional directions from their composition and evaluates multiple composition strategies, including static, semantic, and activation-based approaches, providing a broad overview of methodological approaches.

Critical weaknesses

The experimental setup is insufficiently covered and improperly placed. Section 3.1 does not explain what si or ti is. In short, clarity issues obscure understanding of the section, which appears to be the paper's core contribution.

The experimental setup is highly constrained and assumes the attributes always have only two possibilities. A true intersectional setup must treat race as a set and not limit it to White vs. Black. Although the limitations acknowledge this, can the setup be extended to a set with more than two possibilities? If so, why did the authors consider only two possibilities per attribute and not generalize the methods section? Acknowledging that limitation in the methods section would help. Also, the methods section should be free of experimental-setup limitations. Decoupling both would be appreciated.

The paper lacks a clear discussion of how CBC is novel and beneficial compared with existing methods that handle intersectional bias. Currently, the benefits are assumed and not intuitively clear.

Eq. 3: "preference change across conditioning values is identical" is it an assumption or guaranteed by the setup?

The method uses dataset-derived weights twice: once to compute WMD and once in APC. The authors can discuss how it doesn't introduce cascading interpretation issues.

The paper describes signed preference scores, and later switches to absolute scores. It is unclear how the direction of preference can be discarded without affecting the intervention effect. Introducing a sign later does not quite address the concern. The authors must consider signed alternatives and report in detail how the current method does not yield a rather confined effect. Dropping signed preference must be justified.

Also, if CBC+APC performs better, the improvement cannot be attributed uniquely to preserving "conditional structure". It might partly result from differences in the construction procedure, among other possibilities.

Questions for authors

Why does CBC+APC produce substantial improvements for race–gender and age–gender, but almost no improvement for age–race? Do the authors suspect any reason?

Do the authors think that the observations made in the synthetic controlled scenario reflect a real-life scenario?

1 comment
AuthorsOpenReview ↗

CBC Construction Scope, Interaction Algebra, and Fixed-Bank Composition Controls

Thank you for the specific questions about generality, the algebraic definition, the relationship between WMD and APC weighting, and the age-pair boundary. The revision separates the general construction interface from the present binary instance and adds fixed-bank composer controls and full checkpoint-level results.

  1. Method generality and current experimental scope: A CBC direction bank is indexed by intervention target and conditioning value. The race–gender, age–gender, and age–race experiments instantiate four conditional targets per pair. For attributes with more than two category values, the same interface extends the index set. The current experiments use binary lexical values so that every condition enters a complete paired evaluation. The method and limitations now place white/black, he/she, and young/old within this bounded lexical scope.

  2. Shared interaction coordinate qq: The shared interaction coordinate follows algebraically from one 2×22 \times 2 table of candidate log probabilities. For the conditional-contrast vector δ\boldsymbol{\delta}, we have

δ=(mT+q,mT−q,mS+q,mS−q)⊤.\boldsymbol{\delta} = (m_T + q, m_T - q, m_S + q, m_S - q)^\top.

This equality follows directly from the four table entries and requires no additional model-behavior assumption. Marginal averages identify mTm_T and mSm_S while leaving qq unidentified. Recovering all four conditional contrasts therefore also requires qq.

  1. Difference from prior composition methods: CBC differs in its construction target. Marginal methods estimate one direction per attribute and then compose those directions. CBC first estimates directions separately under each value of the other attribute. In the public-method transfer results, FairSteer and DSAS-inspired attain Marginal2 values of 15.32% and 15.93%, with corresponding Cross4 values of 3.55% and 4.11%. The main controlled experiment compares the complete Single and CBC construction settings under the same APC rule.

  2. Relationship between WMD and APC weighting: The weights act at different stages. WMD uses training-set bias magnitudes once to estimate each direction. APC uses the current target activation at inference time to compose an already constructed bank. With the CBC bank fixed, all six composers yield positive four-model Cross4, ranging from 7.01% to 9.24%. GlobalBias further gives CBC-minus-Single differences of −6.49 points under APC and +4.39 points under static equal weighting. Composition is therefore an experimental factor distinct from direction extraction.

  3. Score sign and weak-effect boundary: Magnitude-stratified WMD groups examples by ∣s∣|s|, so reversing the sign leaves its groups and weights unchanged. The revision discusses this property together with the weaker qq-RMS result and defines the evaluated instance as magnitude-based construction. Age–race provides the clearest controlled boundary. Its four-model mean difference is +0.15 points (95% CI [−1.06, 1.36]), and two of the four model-level differences are positive. Qwen2.5-3B has a −2.18-point Pair4 difference and a −22.66% worst CBC slice. The aggregate 4.44-point result is therefore driven primarily by race–gender and age–gender, which the revised main result states directly. GlobalBias broadens the names, descriptors, templates, and checkpoint set and directly evaluates lexical distribution transfer of frozen direction banks. Its dynamic/static ordering shift and checkpoint heterogeneity constitute the external-transfer evidence. Open-ended generation and broader demographic taxonomies define the external-validity boundary.

Reviewer_p16cOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Construct Before Compose (CBC), which constructs attribute directions under cross-attribute conditions before composition. Using weighted mean difference and activation projection composition (APC), CBC is evaluated across four instruction-tuned checkpoints and three attribute pairs. It improves the mean Pair4 reduction from 7.70% to 12.13%, with a general-language macro-average accuracy change of approximately −0.30 percentage points. The paper also examines the differences between single-attribute mitigation, conditional bias attenuation, and interaction reduction.

Critical strengths

1.The distinction between Marginal2, Pair4, and interaction reduction is useful. The experiments show that successful single-attribute mitigation does not reliably transfer to conditional targets, supporting the need for conditional evaluation.

2.The race–gender improvement is supported by multiple random splits and an occupation-grouped split. Positive Cross4 reductions under alternative composers and the small average utility loss provide additional support for the approach.

Critical weaknesses

1.The main comparison does not isolate the contribution of conditional construction. Single+APC and CBC+APC jointly differ in prompt view, scoring target, direction count, and parameter sharing. Although acknowledged, this limits the evidence that preserving cross-attribute conditions is responsible for the gains.

2.The evidence for preserving conditional structure is incomplete. Appendix B.3 shows that the magnitude-based WMD estimator need not identify the interaction component. CBC+APC also achieves less mean interaction reduction than Single+APC (3.26% versus 4.99%). With bias evaluation restricted to WinoGenerated lexical probes, the broader significance of the Pair4 improvement remains uncertain.

Questions for authors

Can you compare averaged preference targets with conditional targets while holding the two-slot prompt view and direction count fixed? This would help clarify the source of the improvement.

Given Appendix B.3 and the weaker interaction reduction, what evidence supports preserving conditional structure beyond improving Pair4? An independent bias evaluation would materially strengthen the assessment.

1 comment
AuthorsOpenReview ↗

Scope of the CBC Method Comparison, Interpretation of Pair4, and Evidence on Candidate Behavior and Transfer

Thank you for focusing on the central identification question: whether the experiment isolates condition indexing or evaluates the complete method, including its prompts, targets, and bank size. We revised the method definition, attribution, and metric interpretation to align each claim with the comparison actually performed.

  1. Scope of the complete-construction effect : Single+APC and CBC+APC share the intervention and evaluation protocol, while prompt view, scoring target, direction count, and parameter sharing change jointly. The revised comparison unit is therefore the complete construction setting. Under the fixed protocol, the equal-pair Pair4 mean increases from 7.70% to 12.13%, a 4.44-point difference. By attribute pair, the differences are +8.42 points for race--gender (95% CI [7.02, 9.92]), +4.74 for age--gender (95% CI [3.50, 5.96]), and +0.15 for age--race (95% CI [-1.06, 1.36]). The revision reports these values as method-level effects whose attribution matches the jointly changed factors.

  2. Operational meaning of “conditional structure” :CBC directly targets four condition-indexed contrasts rather than the shared interaction coordinate qq. The algebraic decomposition shows that marginal averages identify mTm_T and mSm_S while leaving qq free. The empirical results likewise separate the two targets: CBC+APC raises Cross4 from 0.82% to 9.24%, while its mean interaction-RMS reduction is 3.26%, below the 4.99% obtained by Single+APC. The revision now reports conditional-contrast attenuation and interaction attenuation as separate estimands.

  3. Whether Pair4 changes enter model behavior : We added three complementary results. First, CBC+APC changes the Top-1 candidate on 19.00% of paired templates relative to No-Intervention, with mean total variation 0.1409 over the four-candidate distribution. Pair4 changes therefore accompany changes in candidate decisions. Second, single-attribute methods transfer unevenly: FairSteer and DSAS-inspired reach Marginal2 values of 15.32% and 15.93%, but Cross4 values of only 3.55% and 4.11%, and both are negative on Llama-3-8B. Third, GlobalBias transfers frozen direction banks without fitting on the external data. CBC-minus-Single is -6.49 points under APC and +4.39 points under static equal weighting. This external result attributes transfer behavior jointly to construction, composition, and checkpoint.

  4. Method-presentation revisions: The method now defines the conditional measurement target, direction construction, fixed-bank composition, and projection replacement in sequence, with all scores, condition indices, and composition weights introduced before use. Binary attribute values are presented as the current experimental instance; indexing the direction bank by intervention target and conditioning value is the construction principle. The sign invariance of WMD is stated explicitly in the method and discussion.

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

The paper proposes Construct Before Compose (CBC), a framework that estimates steering directions under different demographic conditions before combining them to mitigate bias in LLMs. Across four models and three demographic attribute pairs, CBC improves average conditional bias reduction over single-attribute construction while maintaining broadly comparable average utility. The paper also shows that reducing single-attribute bias does not reliably reduce conditional disparities, motivating condition-aware construction and evaluation.

Critical strengths
  1. The paper addresses the gap between single-attribute bias mitigation and conditional demographic disparities.
  2. Race–gender improvements are supported across multiple checkpoints and data splits, with small average utility degradation.
  3. The paper distinguishes different forms of bias reduction, acknowledges limitations, and provides detailed supplementary data and code.
Critical weaknesses
  1. CBC changes the prompt format, supervision targets, and direction count simultaneously, making the specific contribution of conditional construction difficult to isolate.
  2. Gains vary across attribute pairs and models; better average conditional scores can coexist with worsened interaction or subgroup disparities.
  3. Evaluation focuses on synthetic occupation templates, leaving effectiveness on independent benchmarks and realistic generation tasks unclear.
  4. The supplementary material contains unresolved age-experiment protocol questions and a scoring bug in the fresh utility-evaluation code.
Questions for authors
  1. Can you isolate conditional construction using identical joint prompts and activations? This would help me understand which changes account for the improvement over the baseline.

  2. Could you clarify which protocol produced the reported age results and whether the possible scoring bug I noticed in the utility evaluation code affects any reported utility scores?

  3. I had trouble interpreting the tied-direction and shuffled-score controls. It would help to see their per-example outputs and a check of their scores under the final protocol.

  4. Have you evaluated CBC on an independent bias benchmark? I’d be interested to see whether the reduction in conditional bias carries over beyond the occupation templates.

  5. The gains vary across attribute pairs and models. Could you walk through one example where CBC improves Pair4 more than it improves the interaction metric, and show where that difference comes from?

2 comments
AuthorsOpenReview ↗

The 4.44-percentage-point gain in mean Pair4 across three attribute pairs reflects the full CBC+APC vs. Single+APC comparison. On GlobalBias, switching from APC to static equal weighting with both banks fixed reverses their average ordering across five checkpoints. We clarify the age evaluation protocols and metric interpretation and report candidate-level diagnostics, worst-slice outcomes, task-level utility, and Qwen2.5-7B results.

Thank you for the detailed suggessions.

  1. Scope of the CBC Gain and Fixed-Bank Composition Controls:The revision reports the 4.44-point gap only as a method-level comparison of the complete construction settings. We added two composition controls. With the CBC direction bank fixed, all six rules yielded positive four-model mean Cross4 (APC 9.24%, static equal weighting 7.58%, TrainBias 7.48%, and the three DynEmb variants 7.01–7.42%). On GlobalBias, with both source banks fixed, changing only the composition rule from APC to static equal weighting shifted the five-checkpoint means from 4.06% to 0.60% for Single and from −2.43% to 4.99% for CBC. The CBC-minus-Single gap moved from −6.49 to +4.39 points, a 10.88-point shift that reversed the ordering. The revision treats construction and composition as distinct, interacting design choices.

  2. Evaluation Protocols and Results for Age–Gender and Age–Race: The revision specifies the full protocol for both age pairs. Each uses 596 validation templates and three arms (No-Intervention, a two-direction Single bank, and a four-direction CBC bank). The four conditional contrasts combine young/old with he/she for age–gender and with white/black for age–race. Each score is the log probability of the full candidate continuation. Across four models, age–gender shows a stable Pair4 gain (9.82% to 14.56%, +4.74 points, 95% CI [3.50, 5.96]). Age–race is a near-zero boundary case (12.45% to 12.59%, +0.15 points, 95% CI [−1.06, 1.36]).

3.Protocol Differences in the Earlier Tied-Direction and Shuffled-Score Controls:The tied-direction and shuffled-score records cited in the review used an earlier direction-binding and scoring pipeline whose protocol differs from the frozen comparison. To apply a consistent evidential standard, the revised conclusions draw on the unified three-pair experiment, three additional race–gender splits, the six-composer fixed-bank control, candidate-probability diagnostics, and GlobalBias transfer. The earlier records remain archived as protocol-history diagnostics and are excluded from the revised quantitative synthesis.

4.What Pair4 and Interaction RMS Measure:Pair4 aggregates attenuation across four conditional contrasts. The shared two-factor coordinate q measures how one attribute contrast varies with the other attribute. On race–gender, CBC+APC and Single+APC yield Cross4 values of 9.24% and 0.82% and mean q-RMS reductions of 3.26% and 4.99%, respectively. CBC thus achieves greater average conditional-contrast attenuation and a smaller reduction in interaction RMS. The revision reports both estimands and notes that magnitude-stratified WMD groups examples by |s|, discarding score signs. It frames the mechanism in terms of conditional-magnitude construction and interprets Pair4 gains as conditional-contrast attenuation.

5.Candidate Ranking and Probability Changes on WinoGenerated and Transfer to GlobalBias:Candidate-level replay shows that, relative to No-Intervention, CBC+APC changes the Top-1 candidate on 19.00% of paired templates (95% CI [17.49, 20.55]) and shifts relative probability mass across the four candidates (mean total variation 0.1409, 95% CI [0.1352, 0.1467]). Relative to Single+APC, CBC+APC differs in Top-1 on 24.16% of templates (95% CI [22.40, 25.92]), with mean total variation of 0.1874 (95% CI [0.1808, 0.1940]).On GlobalBias, we transfer the frozen WinoGenerated race–gender banks to 730 descriptors, three templates, and 40 gender-by-ethnicity name groups across five checkpoints. The aggregate ordering reverses between APC and static equal weighting, with substantial checkpoint-level heterogeneity.

6.Worst-Slice Results and Task-Level Utility:The revision reports local outcomes alongside averages. On the main race–gender split, the four-model mean worst-slice result is +4.64% for CBC and −6.08% for Single. Outcomes vary across attribute pairs. Qwen2.5-3B’s worst CBC age–race slice reaches −22.66%, and the CBC-minus-Single difference for Llama-3-8B on age–gender is −7.51 points. Across six general-language tasks, the four-model macro changes relative to No-Intervention are −0.463 points for Single and −0.299 for CBC. CBC declines by 2.02 points on TruthfulQA and averages +0.045 points over the other five tasks.

7.Model Selection and Qwen2.5-7B Results:To address model selection, the revision specifies the scope of the four-model main table and retains complete Qwen2.5-7B results in the reproduction materials. On this checkpoint, CBC minus Single is −1.47 points for race–gender, +10.50 for age–gender, and +6.09 for age–race. On GlobalBias, Single and CBC score 19.31% and 21.84% under APC and 19.89% and 7.43% under static equal weighting, respectively. These results reinforce that construction effects depend on the attribute pair, composer, and checkpoint.

Thank you for the detailed rebuttal and additional experiments. However, my main concerns remain regarding the robustness and generalizability of the proposed method. The improvements appear highly dependent on the composition strategy, attribute pair, and model checkpoint, with some settings showing negligible or negative gains. Moreover, the connection between conditional-contrast attenuation and meaningful bias reduction remains unclear. Therefore, I will maintain my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.