acceptodds
Under review as a conference paper at ICLR 2027

BRIDGE: BRANCH-GUIDED RESIDUAL INTERVENTIONS VIA DYNAMIC GEOMETRY

Abstract

Representation intervention controls the behavior of large language models (LLMs) at inference time without updating parameters. Existing methods typically choose intermediate or late layers empirically as the source of intervention signals. However, we find that early-layer signals in fact achieve higher coverage of correct answers, but many irrelevant signals hinder direct use for effective intervention. To address this challenge, we propose BRIDGE (Branch-guided Residual Interventions via Dynamic GEometry). Unlike prior methods, BRIDGE extracts intervention signals by identifying the useful early branches and preserves general capabilities through a minimal, norm-preserving edit. Specifically, Early Branch Identification (EBI) projects an early-layer hidden state through the output head to select first-token anchors, extends them into short natural branches, and learns coverage-based utilities to identify the branches worth strengthening. Dynamics-aware Geometric Editing (DGE) then specifies a target utility gain, solves for the smallest norm-preserving residual edit that achieves it, and keeps the edit only when a forward pass confirms a positive realized gain. Across five Llama and Qwen models from 3B to 32B, BRIDGE consistently improves truthfulness, fairness, and safety alignment while preserving general capabilities, showing that properly identified early branches provide effective intervention signals for controllable LLM generation. Our code is available at https://anonymous.4open.science/r/Bridge-1FC8.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
activation steering, LLM alignment

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_2LhSOpenReview ↗
Rating
2
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
3(+1)
Read review
Paper summary

BRIDGE combines early-layer branch identification with inference-time activation editing. A supervised scorer predicts utilities for short continuations anchored by early-layer predictions. These utilities guide small, norm-preserving edits at later layers, which are accepted after forward verification. Experiments evaluate truthfulness, bias-related accuracy, safety refusal, and general capabilities across five Llama and Qwen models.

Critical strengths
  1. Using prompt-specific branches to construct intervention objectives is an interesting approach. Combining branch valuation with controlled, forward-verified edits is reasonably motivated.

  2. The experiments cover two model families and scales up to Qwen3-32B, with reported improvements across several target benchmarks. Including general-capability evaluations and inference costs is useful.

Critical weaknesses
  1. The contribution of early-layer signals is not adequately isolated. Figure 2 compares learned, EBI-filtered early signals against raw middle- and late-layer signals. It does not establish whether early signals remain preferable when other layers receive the same branch construction, utility learning, and editing procedure. The “w/o EBI” replacement also changes both the signal source and extraction mechanism.

  2. The practical advantage over simpler alternatives remains unclear. BRIDGE requires supervised scorer training, branch generation, gradient computation, and forward verification, with appreciable inference overhead. Comparisons against data-matched lightweight fine-tuning and reranking using the same branches and scorer would help establish when internal-state editing is worthwhile and what benefits it provides beyond candidate selection.

  3. The alignment evaluation supports a narrower conclusion than the overall framing suggests. TruthfulQA and BBQ primarily assess multiple-choice performance, while safety evaluation measures refusal on harmful requests without a corresponding assessment of benign helpfulness or over-refusal. Using the same judge for safety-training labels and final evaluation also leaves uncertainty about evaluator-specific effects.

Questions for authors
  1. Would the early-layer advantage persist if middle- and late-layer variants used the same branch-utility learning and DGE procedure, with comparable supervision and tuning budgets?

  2. How does BRIDGE compare with lightweight fine-tuning under the same supervision budget, and with reranking the same branches using the learned scorer under a comparable inference budget?

Reviewer_hmQiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Inference-time representational interventions for LLMs are typically done in intermediate-to-late layers, because these layers are empirically found to have more easily extractable intervention signals. However, this paper argues that edits in early layers can result in higher coverage answers, whereas edits in later layers result in mode-collapse. Thus, the authors propose BRIDGE to perform early layer edits. BRIDGE is evaluated on 3 benchmarks evaluating truthfulness, social bias, and refusal of harmful requests. BRIDGE outperforms other methods in all three of these settings across two model families while preserving the model's general capabilities.

Critical strengths
  • (S1) Interesting method with compelling motivation. The Introduction motivates why inference-time interventions may be better done in early layers. However, most prior works focus on middle-to-late layers, primarily because it is easier to cleanly extract an intervention signal (more likely to be linearly separable). BRIDGE seems like a reasonable method for such early-layer inference-time interventions.

  • (S2) Strong benchmark performance. BRIDGE outperforms other inference-time approaches on 3 settings across two model families (and preserves general capabilities). The experiments in Section 4.3 nicely validate the methodological components of BRIDGE and analyze the benefits of BRIDGE vs. CAA.

Critical weaknesses
  • (S1) No cross-dataset transfer experiment. BRIDGE is an early-layer edit method. It's possible that early-layer methods have poorer generalization compared to intermediate/late layer methods (such as the ones compared to in this work). Thus, the paper should evaluate if BRIDGE overfits. This is a critical experiment to do, because any practically useful inference-time edit method should generalize to other datasets (e.g., ideally, I don't need to label correct/incorrect for every new jailbreak method/dataset I want to apply BRIDGE to).

This weakness is important enough that it has significantly influenced my rating. If the authors can either (1) conduct such an experiment and/or (2) explain why BRIDGE conceptually would generalize to other datasets, I may be compelled to raise my rating. If conducting an experiment, consider the following setting (if it feels reasonable to the authors): use a source dataset (e.g., Sorry-Bench) as the training + validation dataset, and test on some other target jailbreak dataset. Evaluate BRIDGE and the other methods.

  • (S2) Computation time. BRIDGE seems to be twice as slow as other inference-time methods. One of the practical interests in inference-time methods is that they are lightweight. If an inference-time method results in a significant slow-down, model providers may just opt to use guardrail models instead. It's unclear to me how addressable this weakness is, but it seems like a limitation worth clearly noting in the Conclusion.
Questions for authors
  • (Q1) Can Table 4 also include "Vanilla" efficiency metrics? This gives a baseline for comparison and allows us to understand how much latency BRIDGE is adding on top of the original model with no inference-time edits.

  • (Q2) When reading the main paper, I was wondering how layers are selected for BRIDGE. It feels to me that this detail is important and was buried in the Appendix (C.1). I think it's worth noting in the main text. Relatedly and importantly, I was unable to determine how layers are selected for the other methods? Can that be clarified?

  • (Q3) Does BRIDGE work for intermediate/later layers of the model as well? It would be interesting to know if BRIDGE offers a consistent improvement for intervention steering compared to other works across all layers. Perhaps the authors could extend Appendix C.1 (Fig 4) for additional layers?

Reviewer_2AhgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BRIDGE, a parameter-free inference-time framework for improving LLM alignment. It combines Early Branch Identification (EBI), which scores candidate continuations from early-layer representations, with Dynamics-aware Geometric Editing (DGE), which applies norm-preserving interventions guided by predicted utility. Experiments on five Llama and Qwen models (3B–32B) demonstrate improvements over five baselines on TruthfulQA, BBQ, and SORRY-Bench. The authors also evaluate general capability preservation on MMLU and ARC, alongside ablation, sensitivity, and efficiency analyses.

Critical strengths
  1. Well-motivated and principled intervention design: BRIDGE explores how early-layer representations can support LLM alignment. Unlike conventional steering methods, EBI explicitly evaluates candidate branches using answer coverage, while DGE applies constrained, norm-preserving updates with forward verification. Together, these components provide a coherent approach to inference-time representation intervention.

  2. Consistent improvements across models and benchmarks: BRIDGE achieves strong results across five Llama and Qwen models and multiple alignment benchmarks. For example, it improves TruthfulQA MC1 on Qwen3-14B from 68.85% to 81.96% over the strongest baseline, while largely preserving MMLU and ARC performance. Ablation studies, sensitivity analyses, and efficiency evaluations further support the empirical findings.

Critical weaknesses
  1. The current ablation studies do not sufficiently isolate the contribution of early-layer branch identification and geometric editing: The current experiments do not adequately isolate the contributions of EBI and DGE. In Figure 2(a), EBI-enhanced early-layer representations are compared with raw middle- and late-layer representations, making it difficult to determine whether the improvements result from early-layer information or the learned branch scoring mechanism. Similarly, the DGE ablation uses direct additive intervention without matching the edit magnitude, leaving the benefits of geometric optimization unclear. More controlled experiments, such as applying EBI across different layers and comparing DGE with norm-matched alternatives, are needed to validate these claims.

  2. Limited evidence for generalizable alignment improvements: The evaluation primarily demonstrates benchmark-specific improvements rather than broader alignment generalization. First, BRIDGE trains dataset-specific utility scorers and evaluates them on the same benchmark distributions, without cross-dataset validation. Second, the fairness evaluation only considers the disambiguated BBQ subset and reports accuracy, providing limited evidence of actual bias reduction. Third, SORRY-Bench uses the same Mistral-7B judge for scorer training and evaluation, introducing potential evaluation bias. Moreover, increased refusal rates may not indicate improved safety without measuring over-refusal on harmless requests.

Questions for authors
  1. In Figure 2(a), EBI is applied only to early-layer signals, while middle- and late-layer signals are evaluated without the same identification mechanism. Could the authors apply EBI and DGE to middle- and late-layer branches using the same training data, branch count, scorer architecture, and intervention budget? This would help establish whether early layers genuinely provide a superior signal source.

  2. For BBQ, could the authors report results on ambiguous questions using bias-specific metrics? For SORRY-Bench, could they provide results from an independent safety judge and evaluate the false-refusal rate on harmless instructions? These results would materially strengthen the claimed improvements in fairness and safety.

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

The authiors present BRIDGE which constructs short continuations from an early layer, learns which branches to favor, and uses these scores to guide small, norm-preserving hidden-state edits. The reported improvements across five models are substantial. However, the experiments do not yet establish whether these gains specifically require early layer representations and geometric editing, or whether simpler supervised branch selection could achieve similar results.

Critical strengths
  1. The method reports consistent improvements across model families and sizes, with useful comparisons against several intervention methods.

  2. The paper includes component ablations training-data sensitivity and latency measurements. The separation between scorer training and final evaluation is described explicitly.

Critical weaknesses

A. The main claim about early layers is not tested with a sufficiently controlled comparison. fig 2 compares EBI processed early signals with raw middle and late layer signals, changing both the representation source and how it is processed. fig 4 only varies layers 1 to 6. The same branch scoring and editing pipeline should be evaluated using middle and late source layers. Moreover, the quoted answer “coverage” in app. C.1 is a density threshold after layer-specific two-dimensional PCA, pooled across questions. This is not direct evidence that an individual question’s correct answer is more recoverable from early layers.

B, The contribution of geometric editing remains difficult to separate from branch scoring and additional computation. BRIDGE already generates and scores several candidate continuations, yet there is no comparison against selecting or reranking those branches using the same learned scorer. tab 4 reports approximately 2.36 seconds total latency, compared with roughly one second for the baselines. Comparisons with matched supervision and inference budgets are therefore important. The additive edit ablation should also match residual magnitudes; otherwise, its large capability losses do not isolate the benefit of the proposed geometry.

C, The evaluation supports narrower conclusions than the claims of truthfulness, fairness, and safety preservation. Truthfulqa is evaluated through multiple-choice likelihoods, bbq uses only disambiguated-question accuracy, and safety is measured through harmful-request refusal. There is no benign-request over-refusal evaluation. Preserving mmlu and ARC answer rankings does not establish that the model remains helpful during ordinary generation. Free response factuality, bias under ambiguous contexts, and benign-request completion would provide more convincing evidence of behavioral improvement.

Questions for authors

1, Does the advantage persist when EBI and DGE are applied identically to early, middle, and late source layers?

2, Forward verification checks the learned utility over five fixed branches. How often do accepted edits actually improve answer correctness or independently judged response quality?

3, App A describes nine alignment-source–capability-benchmark settings, but tab. 1 gives one result per capability benchmark. How are these combined, and can the results for each source scorer be reported separately?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thank You for the Reviews and Invitation to Continue Discussion

We would like to sincerely thank all reviewers again for their thoughtful reviews and constructive feedback. We genuinely believe that your comments have helped us further strengthen BRIDGE and have encouraged us to examine our method from several additional perspectives.

The additional experiments have given us a more comprehensive understanding of our method. For example, the frozen cross-dataset experiment demonstrates that BRIDGE can transfer from SORRY-Bench to JailbreakBench without target-domain supervision, achieving results close to those of in-domain training. Evaluation with an independent safety judge further supports the robustness of our results. We also examined the actual outcomes of accepted edits, finding that improvements substantially outnumber regressions across our evaluations, while identifying a small proportion of edits that can still negatively affect predictions. We believe these additional analyses have made our empirical findings more complete and our understanding of BRIDGE more nuanced.

We have submitted a shared author response addressing common questions, along with individual responses to all four reviewers. We would be very grateful if you could revisit our responses when time permits. More importantly, we sincerely hope to have the opportunity to continue the discussion, regardless of whether it ultimately leads to any change in the ratings. We warmly welcome any further questions, suggestions, or remaining concerns, and would be more than happy to provide additional clarifications.

Once again, thank you for your time, effort, and valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.