PaperlayerView exact revision on arXiv

Source-linked AI summary

Evaluating the Semantic Specificity of Representation Steering in Language Models

Zhangdie Yuan, Andreas Vlachos

arXiv:2608.29431v1cs.CL

TL;DR

Standard benchmark gains from representation steering may reflect superficial label overrides rather than repaired reasoning circuits. The paper proposes Cross-Rule Transfer to audit this distinction and finds that late-layer LRS reduces native entailment competence from 99.6% to 40.4%, with complementary controls supporting the label-bias diagnosis.

  • Problem

    Aggregate benchmark accuracy can be misleading because interventions may inject global label-specific logit biases instead of repairing structural reasoning circuits.

  • Method

    Cross-Rule Transfer audits representational interventions on structurally related rule families where models are natively competent, supplemented by four mechanistic and behavioral controls.

  • Results

    99.6% baseline entailment competence falls to 40.4% after late-layer LRS, while controls support functional equivalence to direct logit biasing and label-preference behavior.

  • Takeaways & Limitations

    CRT provides a model-agnostic safeguard for distinguishing genuine reasoning repairs from superficial token-level logit manipulation before deployment.

  • Takeaways & Limitations

    The evaluated CLUTRR, SpartQA, and ProntoQA benchmarks cover stylized structured reasoning rather than open-ended conversational or free-form dialog reasoning.

Abstract

from arXiv · show

Localized Representation Steering (LRS) is widely used to correct reasoning pathologies in large language models. However, standard benchmark evaluations can easily be fooled by superficial label overrides, creating a false impression of reasoning circuit repairs. In this work, we propose Cross-Rule Transfer (CRT), a diagnostic framework that audits representational interventions by evaluating them on rule families where the model is natively competent. Evaluating late-layer LRS for a widespread logical failure, contradiction blindness, reveals that the intervention merely injects a global label bias: applying the steering vector to rules the model already handles correctly (99.6% baseline) degrades performance to 40.4% by forcing false contradiction predictions. We support this diagnosis with four complementary controls (direct logit bias equivalence, control vector label-flipping, cross-model grafting, and early-layer steering checks), providing a rigorous methodology to distinguish genuine reasoning repairs from superficial label overrides.

1 Introduction

The paper argues that benchmark recovery can reflect superficial label biases rather than repaired reasoning circuits. It introduces CRT to test this distinction and finds that late-layer LRS damages competence on rules the model already handles.

  • Late-layer LRS appears to restore contradiction-blind reasoning to 100% benchmark accuracy without side effects on unrelated tasks.
  • Aggregate accuracy can conceal shortcut exploitation, including global label-specific logit biases that override predictions instead of repairing reasoning circuits.
  • CRT evaluates interventions on structurally disjoint rule families where the model already performs well, testing whether native competence is preserved.
  • CRT combines a late-layer logit-bottleneck explanation with a behavioral audit contrasting natively competent Rule A against failing Rule B.
  • 99.6% native competence on entailment rules falls to 40.4% after steering, supporting a superficial label override rather than a reasoning repair.
  • The diagnostic aims to prevent false attribution of reasoning repairs.

2 Related Work

Related work situates the paper at the intersection of representation steering, reasoning-evaluation pitfalls, and audits of model-editing mechanisms. These literatures motivate checking whether behavioral gains reflect structural computation or token-level shortcuts.

  • Representation engineering and activation steering modify intermediate activations with learned vectors to influence attributes such as truthfulness, safety, and reasoning.
  • Contrastive steering vectors are commonly built by difference-in-means or by isolating function-specific directions.
  • Aggregated metrics can mask lexical or template shortcuts that produce high benchmark accuracy despite systematic failures in structural reasoning.
  • Standard OOD evaluations may miss label biases when target label distributions remain identical.
  • Model-editing audits report that some updates behave as localized logit switches, while late-layer representations encode token-level output commitments.

3 Behavioral Atlas

The behavioral atlas uses a procedurally generated NLI benchmark with structurally separated rule families to characterize contradiction blindness across models. It finds the pathology across architectures and scales, while the anchor model predicts entailment universally.

  • The anchor model predicts entailment universally, succeeding on forward rules while failing completely on contradiction detection.
  • Asymmetricity-2.0 is a procedurally generated NLI benchmark built from factual premises and relational hypotheses testing entailment and contradiction.
  • The benchmark uses broad relations and entity pairings, with disjoint entity partitions across splits to reduce pretraining contamination and surface-cue matching.
  • Rule families are structurally disjoint: competent families require forward deductions, whereas failing families require detecting relational violations.
  • At 65,536 instances, seven of 18 models have zero contradiction recall and five have near-zero recall.
  • The pathology spans Llama, Gemma, Qwen, and DeepSeek families across models ranging from 0.5B to 9B parameters.

4 Method: Probing & Steering

The method combines cross-layer behavioral probing with localized steering, then audits whether the resulting vector preserves native competence or instead encodes a label preference.

  • Probing: Probe classifiers distinguish successful from failing predictions across layers, but their behavioral targets may reflect output commitments rather than reasoning circuits.Baseline models usually fail contradiction inputs while succeeding on entailments, making probe labels potentially confounded with output labels.
  • Probing: 86.1% and 86.3% late-layer error rates for Llama-3.2-1B-Instruct and Gemma-4-E4B-It indicate systematic cross-layer transfer misclassification.The late-layer error rates exceed 50%, while a nonlinear probe reaches 83.4% ± 0.4% late-layer error.
  • Localization: The study contrasts late-region failure concentration across models, with Llama and Gemma clearest, Qwen showing leakage, and DeepSeek nearly uniform across layers.Layer regions are normalized as early first 20%, mid middle 40%, and late last 40%.
  • Steering: Steering vectors are constructed from mean activation differences and injected into final prediction tokens at target layers with scaling factor α = 1.5.Unconditional steering at α ≥1.5 collapses entailment performance, whereas task-level gating produces zero source-benchmark side effects.
  • Audit design: Because successful contradiction cases and failing cases are nearly confounded with their output labels, the difference vector is expected to capture a label-preference direction.CRT tests this possibility on structurally disjoint rules where native competence should be preserved, and can also audit leakage-mitigation constructions.

5 Auditing Steering via Cross-Rule Transfer

Cross-Rule Transfer audits whether steering vectors repair reasoning circuits by testing them on rule families where the model is already competent. The audit shows late-layer LRS collapses native entailment performance and behaves like a global label override.

  • Diagnostic protocol: CRT evaluates interventions on rule families with high native competence to distinguish circuit repair from global output bias.The protocol applies the vector unconditionally, measures competence loss and false target-label injections, then classifies the intervention by the resulting error pattern.
  • Primary diagnostic: 99.6% baseline entailment accuracy on Llama falls to 40.4% after steering, with 298 false contradiction predictions among 500 items.The model satisfies the native competence threshold of θ ≥95%, making the performance collapse diagnostically meaningful.
  • Diagnostic rationale: Cross-Rule Transfer differs from standard OOD evaluation by testing logically related tasks where the model is natively competent, rather than new prompts for the same target task.Its purpose is falsification of global label-preference overrides, not measurement of ordinary generalization strength.
  • Steering-strength sweep: No steering strength avoids the trade-off: α ≤0.5 is too weak for target recovery, while α ≥1.0 globally overrides competent-rule predictions.Competent-rule accuracy decreases monotonically as α increases, with no intermediate setting that restores contradiction recall without destroying entailment competence.
  • Model eligibility: Qwen is not a valid CRT subject because its 45.2% entailment baseline fails the θ ≥95% competence criterion.Its improved accuracy under steering is retained only as a directional comparison, not as an independent CRT audit.
  • Diagnostic controls: Direct logit biasing matches LRS downstream, while final-token steering reproduces full target recovery and reasoning-token steering produces 0% recovery.These controls support a token-level output override rather than intervention in the internal reasoning trajectory.
  • Diagnostic controls: The DeepSeek-R1-Distill-Qwen3-8B anomaly reaches 94%–100% nominal accuracy but operates as a format-compliance bypass rather than a reasoning repair.The result illustrates how late-layer interventions can force compliant label tokens without correcting logical reasoning.

6 Discussion

The discussion argues that representation steering should be validated beyond isolated target-task recovery. CRT exposes label-bias failures through black-box tests on natively competent rule families.

  • Implications for model editing: Representation-based corrections should not be deployed without validation because label-bias interventions can behave unpredictably under different label distributions.The discussion highlights the DeepSeek anomaly, where steering bypassed formatting constraints instead of repairing logical reasoning.
  • Validation protocol: CRT is model-agnostic and lightweight because it operates at the evaluation interface without requiring internal parameter access.This makes it applicable to closed-source or API-only models, where unembedding-based geometric checks are unavailable.
  • Diagnostic controls: The shuffle control flips all-entailment predictions to all-contradiction because its vector remains aligned with the label preference direction.Random Gaussian perturbations and equal-norm orthogonal vectors do not alter outputs, ruling out generic noise as the explanation.
  • Validation protocol: The authors recommend pairing reasoning-repair claims with audits on natively competent within-domain rule families.A collapse under unconditional steering should be characterized as label-bias injection rather than a mechanistic reasoning fix.

7 Conclusion

The paper concludes that late-layer representation steering can mimic reasoning repair while acting as superficial logit-bias manipulation. CRT provides a standardized behavioral safeguard against misattributing such effects to structural reasoning changes.

  • Conclusion: Late-layer interventions collapse entailment competence from 99.6% to 40.4% while showing functional equivalence to direct logit biasing.The conclusion frames this as evidence that steering projects directly into vocabulary logits rather than reliably modifying reasoning circuits.
  • Conclusion: CRT is presented as a model-agnostic falsification safeguard for distinguishing structural circuit interventions from token-level logit manipulation.Its purpose is to prevent false attribution of reasoning capabilities in language models.

Limitations

The findings are bounded by biases in the Wikidata-derived benchmark, English-only evaluation, and stylized downstream tasks that may not represent open-ended reasoning.

  • The Asymmetricity-2.0 benchmark uses relational facts extracted from Wikidata, whose geographic, cultural, and historical reporting biases cannot be controlled.
  • The study evaluates only English-language models, templates, and relational logic, limiting direct evidence for multilingual applicability.The authors identify multilingual models and non-English relational logic as necessary next steps.
  • The downstream CLUTRR, SpartQA, and ProntoQA benchmarks cover stylized structured reasoning rather than open-ended conversational or free-form dialogue.Their label-token boundaries are also less representative of unconstrained dialogue settings.

Ethics Statement

The paper argues that representation-steering audits are important for preventing false impressions of reasoning repair or safety alignment, while noting that parameter-level validation remains incomplete.

  • Rigorous validation is presented as necessary because interventions that merely shift label distributions can create silent deployment failures.
  • Five independent diagnostic tests converge on late-layer LRS operating primarily along label-preference directions under greedy decoding in this setting.The tests include competence collapse, DLB equivalence, shuffle-vector label flipping, cross-architecture grafting, and early-layer failure.
  • Cross-Rule Transfer provides behavioral and representational evidence, but complete parameter-level mechanistic validation would require direct causal checks on model weights.
  • The authors position CRT as a general-purpose first line of defense when weight-space verification is impossible for API-based models.

B Asymmetricity-2.0 Benchmark Design & Validation

Asymmetricity-2.0 is a large, procedurally generated binary NLI benchmark built from Wikidata relations, with controlled logical rules, disjoint OOD splits, and validation against shortcut and labeling concerns.

  • Benchmark construction: Asymmetricity-2.0 contains over 72 million NLI entries generated from Wikidata relation triples and natural-language templates.It includes both lexicalized and delexicalized variants to test reliance on surface wording.
  • Logical rule families: The benchmark tests symmetric, antisymmetric, and irreflexive relational rules, including forward deductions and contradiction detection.Contradictions include violations such as an entity being married to itself or being its own parent.
  • Validation: Reasoning chains contain one- or two-step structures, and generated contradictions are verified by a deterministic logical rule solver.
  • Distribution shifts: Six OOD splits separate entities, entity-relation pairs, properties, templates, or reasoning-chain lengths between training and evaluation.The property-disjoint split is designed to test semantic generalization.
  • Class balance: Logical templates are balanced 50/50, but instantiated evaluation splits can be approximately 30% contradiction and 70% entailment because of consistency and entity-disjointness constraints.
  • Steering audit: At α = 1.50, unconditional steering collapses downstream transfer to chance because it flips entailment predictions to contradiction.At α ∈[0.25, 1.0], the vector achieves perfect recovery on source and downstream contradiction items.
  • External validation: On MNLI, late-layer steering trained on failures reduces competent transitivity and symmetry splits to 35.4% accuracy, indicating the same override mechanism beyond Asymmetricity-2.0.

N Downstream Transfer and Equivalence Results

Downstream audits show that late-layer LRS and direct logit biasing behave equivalently, with label-bias effects persisting under sampling and matching the residual-stream logit-shift mechanism.

  • Functional equivalence: LRS and DLB show identical downstream accuracy profiles across models and tasks under greedy decoding with explicit label-token outputs.The downstream evaluation covers CLUTRR, SpartQA, and ProntoQA.
  • Gating control: Oracle-gated LRS achieves perfect recovery with zero side effects on non-target rule families, but the table explicitly attributes this behavior to task-level gating.
  • Decoding robustness: Under temperature sampling, LRS and DLB remain tightly aligned with Cohen’s kappa κ ≥0.94, while CRT competence collapses to 38.6% ± 1.2% accuracy.These runs use T = 0.7 and nucleus sampling p = 0.9.
  • Mechanism: The residual-stream intervention adds αv to the hidden state, shifting logits by αvW_U before the final-token argmax.
  • Mechanism: When the steering vector aligns with the contradiction–entailment unembedding difference, representation steering is geometrically equivalent to a static token-level logit bias.The paper identifies cosine similarity with unembedding differences as a direct parameter-level verification.

O The DeepSeek Anomaly

DeepSeek’s apparent downstream recovery is driven by forced output formatting and label injection rather than restored reasoning. Across models, steering produces systematic label flips and weak generalization beyond the source benchmark.

  • DeepSeek anomaly: 94%–100% nominal downstream accuracy in DeepSeek-R1-Distill-Qwen3-8B rises from largely unparsed outputs, but per-instance inspection attributes the gain to forced formatting.The steering vector forces parseable outputs containing the target label token, yielding a +1.000 Δ on CLUTRR and SpartQA.
  • Overall diagnosis: Downstream probes remain weak, chance-like, or near-floor overall, while label-mode columns indicate that accuracy changes arise from systematic label flips rather than improved reasoning.Table 21 reports balanced 50/50 splits with N=100 per task and distinguishes recoveries from side effects.
  • DeepSeek anomaly: The DeepSeek vector acts as a brute-force label injector that bypasses formatting failures, creating a false impression of reasoning recovery.Its imposed label template happens to match the ground truth rather than restoring a reasoning circuit.
  • Cross-model transfer: Δ=0 on CLUTRR for Llama-3.2-1B-Instruct masks a complete inversion from uniformly entailment predictions to uniformly contradiction predictions.Before steering, all entailment items are correct and all contradiction items are wrong; afterward, the reverse pattern appears.
  • Cross-model transfer: Qwen-3.5-9B-Instruct shows a contradiction-bias reversal that partially helps on CLUTRR and SpartQA but worsens ProntoQA by overshooting the ground-truth label distribution.The steering intervention changes the model’s label bias rather than consistently improving downstream reasoning.
Loading 2608.29431v1…