acceptodds
Under review as a conference paper at ICLR 2027

Not All Errors Are Equally Steerable: Predicting Error Types to Guide Activation Steering

Abstract

Activation steering aims to correct language-model errors by modifying internal activations at inference time, but different errors may not be equally amenable to the same intervention. We ask whether layer-wise token trajectories can reveal which errors are most likely to benefit from steering. Tracing the correct-answer token across layers reveals two recurring patterns. In a no-entry error, the correct token never becomes a plausible output candidate; in a competitive-decay (CD) error, it becomes competitive at an intermediate layer but loses its advantage before the final output. Under the same fixed steering method, CD errors are recovered approximately four times as often as no-entry errors. Motivated by this gap, we introduce Information Dynamics, a framework for characterizing these trajectories and predicting error type. Although the taxonomy is defined using the reference answer, error type can be predicted from the model’s own early candidate dynamics without access to the reference answer at inference time. Across six open-weight models and eight tasks, these dynamics consistently distinguish CD from no-entry errors. We use these predictions to allocate a fixed steering budget toward more recoverable errors. Across seven tasks on Qwen3-8B and Llama-3.1-8B, steering only 20% of errors selected by predicted CD probability corrects more than twice as many errors as random selection under the same intervention budget. These results show that layer-wise token dynamics provide a practical signal of error steerability and can substantially improve activation steering.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, error analysis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ediVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether layer-wise token trajectories can help identify which errors are steerable. The authors identify two types of errors: no-entry errors and competitive-decay errors, and show that the latter are easier to steer. Motivated by this finding, they further propose a framework that dynamically allocates more steering budget to competitive-decay errors during inference-time steering.

Critical strengths

The paper is in general well-written and easy to follow. The problem studied is interesting and well-motivated. Characterizing error types based on their steerability or steering difficulty from the perspective of a model’s internal layer-wise dynamics, and using this to guide inference-time interventions, is intriguing and of practical significance.

Critical weaknesses

One of my major concerns is the practical benefit of the proposed method. First, according to Table 3, among the successfully recovered errors, no-entry errors account for 64% (16/25) on Qwen and 66.7% (14/21) on Llama. I agree that intervention efficiency is higher for CD errors. However, CD errors are far less common than no-entry errors (roughly a 1:8-1:10 ratio in both models). The authors should therefore better justify the benefit of preferentially allocating steering budget to CD errors, especially when the intervention budget increases and more no-entry errors are considered. Second, Table 4 shows only marginal improvements over the endpoint RF baseline. Specifically, the proposed framework recovers only 1 additional error on Llama and 2 on Qwen compared with using final-layer features alone. Table 4 reports results under a 20% intervention budget. Although the authors provide additional results for budgets ranging from 10% to 50%, these ablations do not include comparisons with endpoint RF. The authors should report endpoint RF performance across these budgets and on larger evaluation sets to better justify the benefit of modeling layer-wise dynamics rather than relying solely on final-layer features.

Another concern relates to the experimental setting. As I understand it, the main predictor-guided steering experiments are conducted exclusively on the error set, which means that the authors must first identify the incorrectly predicted samples in the evaluation set and exclude the correctly predicted ones. Although the CD predictor does not require a reference answer at inference time, constructing the error set on which the CD predictor is applied still requires correctness information for the entire evaluation set in advance. However, in real-world scenarios, the gold labels for test samples may not be available, so the select and steering operation would need to be applied to a mixture of correct and incorrect samples. The authors do not report the steering results for initially correctly predicted samples, even though the steering operation could potentially turn initially correct predictions into incorrect ones.

Questions for authors
  1. Could the authors report a performance comparison between the proposed method and endpoint RF across all intervention budgets and, if possible, on a larger evaluation set?

  2. Could the authors report the results of applying the proposed method to the full evaluation set rather than only to the error set?

  3. Does CD status remain predictive of recovery after controlling for the reference token’s final-layer rank or margin?

4.How does the router behave on initially correct samples? According to Appendix B.1, CD error samples and correctly predicted samples can look similar in the middle layers. If the CD predictor assigns high preference scores to correctly predicted samples, it is possible that many correct samples that should not be steered would be selected from a mixed pool.

8 comments
AuthorsOpenReview ↗

Official Response to Reviewer ediV 1/3

We sincerely thank Reviewer ediV for the positive review, and for recognizing that our paper is well-written, our problem is well-motivated, and using layer-wise dynamics to guide steering is intriguing and of practical significance.

In this rebuttal, we have:

  • Clarified why prioritizing CD errors maximizes efficiency under tight budgets, while naturally recovering No-entry errors as budgets grow, as explained in W1 below.

  • Evaluated our method on the full evaluation set of 646 questions across seven tasks without oracle error filtering, achieving a +2.17 percentage point net accuracy gain with zero correct answers harmed, as reported in W3 and Q2 below.

  • Provided a complete comparison against Endpoint RF across all budgets (10% to 50%) and on an expanded MMLU evaluation set, as detailed in W2 and Q1 below.

  • Reported matched control experiments confirming CD errors recover more often even when final-layer rank and margin are matched, as presented in Q3 below.

  • Audited initially correct answers and confirmed zero negative impact: our router rarely selects them, and steering introduces zero new errors, as detailed in Q4 below.


W1. Benefit of prioritizing CD errors over more numerous No-entry errors.

Summary: We completely agree that if you steer every single error (100% budget), No-entry errors make up most recoveries simply because they are eight times more. But when intervention budgets are limited (e.g., 10%–20%), targeting CD errors first gives the highest return on investment because their per-sample success rate is over 4× higher (17.3% vs. 3.9%). Crucially, our router does not discard No-entry errors; as the budget expands, it naturally selects and fixes them. At a 50% budget, at least 8 of our 17 recoveries were No-entry errors.

  • Why prioritize CD under a limited budget? CD errors offer greater potential for correction because the correct answer has already become competitive in intermediate layers. Prioritizing these errors directs limited interventions toward cases where this existing competition can be strengthened. On Qwen, CD errors recover at 17.3% (9/52) compared to just 3.9% (16/409) for No-entry errors, concentrating limited interventions where they work best.

  • Our method naturally captures and recovers No-entry errors as the budget expands: Our router ranks errors continuously rather than making a hard binary cut.

    • At a 50% budget on Qwen, our method recovered 17 errors. Because the entire pool only contains 9 recoverable CD errors, at least 8 of these 17 recoveries must be No-entry errors.
    • At a 20% budget, at least 4 of 13 recoveries on Qwen and 3 of 10 on Llama were No-entry errors.
  • Key takeaway: We prioritize CD errors because the correct answer is already competitive in intermediate layers, giving greater correction potential through steering. As the budget expands, our continuous ranking also selects and recovers No-entry errors.


W2. Marginal improvements over Endpoint RF and the value of layer-wise dynamics.

To demonstrate the value of layer-wise dynamics across intervention budgets and on a larger evaluation set, we compare Candidate Dynamics with Endpoint RF across four budget levels and on additional MMLU questions.

At the 20% budget, recovery gains were modest (+2 on Qwen and +1 on Llama). On Qwen, four errors were corrected only by Candidate Dynamics, while two were corrected only by Endpoint RF. Across five paired seeds, however, Candidate Dynamics achieved higher CD AUROC in every run on both models. Mean AUROC was 0.843 ± 0.035 vs. 0.764 ± 0.006 on Qwen and 0.850 ± 0.027 vs. 0.743 ± 0.005 on Llama. The corresponding ensemble AUROCs were 0.948 vs. 0.802 and 0.937 vs. 0.781.

Candidate Dynamics holds this advantage because final-layer probabilities cannot separate errors caused by complete factual ignorance from errors caused by competitive decay, where the model internally recalled the correct answer but lost it late. Endpoint RF treats both as generic low-confidence errors, whereas Candidate Dynamics tracks early candidate competition to selectively prioritize recoverable cases. For example, on an ARC-Challenge question incorrectly answered as (A) Aphelion, Endpoint RF discarded it due to low final-layer confidence (0.38). In contrast, Candidate Dynamics detected that the correct answer (B) Perihelion surged to Rank 0 (+1.85 nats) at Layers 16–22, assigned it a high CD probability of 0.84, and successfully recovered it under steering, securing a genuine recovery that Endpoint RF missed.

Under identical five-seed setups on the 461-error Qwen pool, Candidate Dynamics corrected two to three more errors than Endpoint RF at all four tested budgets from 10% to 50%, doubling the recovery count at 10%.Furthermore, Candidate Dynamics corrected more errors at every tested budget on the additional MMLU evaluation.This confirms that modeling layer-wise dynamics provides a stable, practical advantage across diverse budgets and larger datasets.

AuthorsOpenReview ↗

Official Response to Reviewer ediV 2/3

W3. Full-dataset evaluation without oracle error filtering.

To test real-world deployment where error labels are unknown in advance, we evaluated our method on the official question-disjoint evaluation split of all 646 questions across seven benchmark datasets (containing 185 correct answers and 461 errors) without any pre-filtering. Our router ranks candidates purely by early internal trajectories without needing ground-truth answers. At the 20% budget, Candidate Dynamics achieved a net accuracy gain exceeding Endpoint RF while preserving all selected correct answers. Crucially, this gain is consistent across all domains without negative side effects: accuracy improved from 34.1% to 36.5% on Knowledge & Multi-hop QA (TriviaQA, HotpotQA; +6 net answers), from 14.8% to 16.9% on Multiple-Choice Reasoning (ARC-C, MMLU, CSQA; +5 net answers), and from 40.6% to 42.6% on Comprehension & Inference (QuALITY, MNLI; +3 net answers). Furthermore, steering introduced zero new errors across all seven tasks: every selected correct answer remained correct after intervention at both tested budgets.


Q1. Comparison with Endpoint RF across all intervention budgets (10% to 50%) and on a larger evaluation set.

Summary: We evaluated Endpoint RF across all four budget tiers on the 461-error Qwen pool under identical 5-seed training setups, as well as on an expanded evaluation set of 1,024 MMLU errors. Candidate Dynamics consistently beat Endpoint RF across all budget tiers on both pools, recovering 2 to 3 additional answers on the original pool and up to 7 additional answers on the expanded set.

1. Four-budget sweep on original Qwen error pool (461 errors)

Qwen Intervention BudgetEndpoint RF RecoveriesCandidate Dynamics RecoveriesNet Gain over Endpoint RF
10% (46 errors)3 (6.5%)6 (13.0%)+3 questions (2.0× baseline)
20% (92 errors)11 (12.0%)13 (14.1%)+2 questions
30% (138 errors)12 (8.7%)15 (10.9%)+3 questions
50% (230 errors)15 (6.5%)17 (7.4%)+2 questions
  • Across all four budgets, watching the trajectory consistently recovered 2 to 3 more answers than looking only at the final layer. This pattern is consistent with the motivation for using trajectory features. Endpoint RF observes the final-layer state, while Candidate Dynamics also tracks whether a competing candidate emerges and is subsequently suppressed across layers. This additional information helps distinguish errors with similar endpoint scores but different responses to steering, as supported by the matched comparisons in Q3. Ranking errors using this information yielded more corrections at each tested budget.

  • At a 10% budget, Candidate Dynamics doubled the yield of Endpoint RF (13.0% vs. 6.5%).

  • On Llama at a 20% budget (472 evaluation errors), Candidate Dynamics recovered 10/94 (10.6%) errors vs. 9/94 (9.6%) for Endpoint RF.

2. Independent evaluation on additional MMLU questions (1,024 errors)

  • Questions and setup. To test whether the gains over Endpoint RF extend to a larger set of previously unseen questions across more subjects, we sampled 1,200 test questions from the remaining 53 MMLU subjects. None matched the original 1,900 questions by dataset ID or prompt text. We kept both five-seed selectors, the steering direction, and α=1.8\alpha = 1.8 fixed.

  • 20% budget. Qwen3-8B answered 1,024 questions incorrectly. Each selector chose 205 errors to steer. Candidate Dynamics corrected 31/205 errors (15.1%), whereas Endpoint RF corrected only 24/205 (11.7%), recovering 7 fewer answers than our method. Both corrected 20 of the same errors. Dynamics alone corrected 11, Endpoint alone corrected 4. Of the 15 differences, 11 favored Dynamics (73.3%, paired odds ratio 2.75, two-sided exact McNemar p=0.118p = 0.118).

  • Other budgets. Recovery counts for Candidate Dynamics versus Endpoint RF were 14 versus 7 at 10%, 38 versus 31 at 30%, and 47 versus 42 at 50%.

Together, these results show that Candidate Dynamics corrected more errors than Endpoint RF at every tested budget on both evaluation pools. The gains extended to previously unseen MMLU subjects with both selectors and the steering setup held fixed.

AuthorsOpenReview ↗

Official Response to Reviewer ediV 3/3

Q2. Results of applying the proposed method to the full evaluation set.

Summary: We applied the frozen router directly to all 646 held-out evaluation questions (185 correct, 461 errors) without oracle filtering. At both 20% and 30% budgets, Candidate Dynamics outperformed Endpoint RF and broke zero correct answers.

Evaluation Setting (All 646 Evaluation Questions)Intervened QuestionsErrors SelectedCorrect SelectedErrors CorrectedCorrect DegradedNet Gain (Δ\Delta Correct)Overall Accuracy
Baseline (No Steering)00000—28.6% (185/646)
Random Selection (20% Budget)*12992.136.95.036.9×d36.9 \times d≤+5.0\le +5.0≤29.4%\le 29.4\%
Endpoint RF (20% Budget)1291209111+1030.2% (+1.55 pp)
Candidate Dynamics (20% Budget)1291263140+1430.8% (+2.17 pp)
Endpoint RF (30% Budget)19417816132+1130.3% (+1.70 pp)
Candidate Dynamics (30% Budget)1941886160+1631.1% (+2.48 pp)
  • Random selection would fix 25×129/646≈5.025 \times 129/646 \approx 5.0 errors on average. Here dd is the fraction of the 185 correct answers that steering would turn wrong. The upper bounds assume d=0d=0.

  • Zero degradation on correct predictions: Candidate Dynamics preserved all selected correct answers at both budgets (0 degraded). In contrast, Endpoint RF introduced errors at both budgets.

  • Reliable net accuracy gains: Overall accuracy increased at both budgets without any oracle intervention.


Q3. Does CD status remain predictive of recovery after controlling for final-layer rank or margin?

Summary: Yes. We evaluated recovery under the main steering intervention (α=1.8\alpha=1.8) across two matched settings: (1) matching on final rank (≤1\le 1) and margin (≤0.5\le 0.5 nats) across the pool (44 pairs), and (2) additionally matching within the same task (41 pairs). Both settings show higher CD recovery (18.2% vs. 4.5%, and 17.1% vs. 2.4%).

Matching SettingPairsCD RecoveriesNo-entry RecoveriesDiscordant Pairs (CD / NE)Two-sided Exact McNemar ppPaired Odds Ratio
Cross-task Matched (∣ΔrL∣≤1,∣ΔmL∣≤0.5\vert\Delta r_L\vert \le 1, \vert\Delta m_L\vert \le 0.5)448 / 44 (18.2%)2 / 44 (4.5%)7 / 1p=0.0703p=0.0703OR = 7.0
Within-task Matched (Same task +∣ΔrL∣≤1,∣ΔmL∣≤0.5+ \vert\Delta r_L\vert \le 1, \vert\Delta m_L\vert \le 0.5)417 / 41 (17.1%)1 / 41 (2.4%)6 / 0p=0.03125p=0.03125∞\infty
  • Cross-task matched details (44 pairs): Mean final rank 8.0 (CD) vs. 8.2 (No-entry); mean final margin −2.08 vs. −2.14 nats. 1 both corrected, 7 only CD corrected, 1 only No-entry corrected, 35 neither corrected (p=0.0703p=0.0703, paired OR = 7.0).

  • Within-task matched details (41 pairs): Drawn from ARC-Challenge, MMLU-4subj, and TriviaQA. Mean final rank 7.9 (CD) vs. 8.1 (No-entry); mean final margin −2.05 vs. −2.11 nats. 1 both corrected, 6 only CD corrected, 0 only No-entry corrected, 34 neither corrected (p=0.03125p=0.03125).

  • Conclusion: Both matched analyses demonstrate that intermediate trajectories capture recoverability differences that remain even after closely matching on endpoint rank and margin.


Q4. Router behavior on initially correct samples: Do correct answers consume budget?

Our method spends little intervention budget on initially correct answers and introduced no new errors among those selected in our evaluation. Although correct answers and CD errors can share patterns in some intermediate layers, the router combines early candidate-dynamics features into a CD preference score, so local visual similarity need not imply similar routing scores. We examined this on 646 held-out Qwen3-8B questions from TriviaQA, HotpotQA, ARC-Challenge, MMLU-4subj, CSQA, QuALITY, and MNLI, including 185 initially correct answers and 461 errors, without correctness-based filtering. Correct answers received a mean score of 0.031 ± 0.024, compared with 0.742 ± 0.165 for CD errors. Our method consequently allocated only 2.3% of intervention slots to correct answers at the 20% budget and 3.1% at 30%. Most correct answers were therefore left untouched. For the selected correct answers, we applied the same fixed steering operation (α=1.8\alpha=1.8) and checked the regenerated outputs. All three selected correct answers at 20% remained correct, as did all six at 30%, including the original three.


Summary

We thank Reviewer ediV for the constructive feedback. Prioritizing CD errors maximizes efficiency when the budget is tight, while our ranking naturally expands to treat and recover No-entry errors as the budget grows (at a 50% budget, at least 8 of the 17 recoveries were No-entry errors). Across all tested budgets, Candidate Dynamics recovered 2 to 3 more answers than Endpoint RF on the original pool, and +7 answers on the expanded 1,024-error set. Most importantly, when tested on all 646 held-out evaluation questions without oracle filtering, accuracy rose from 28.6% to 30.8% with zero correct answers harmed. We hope these results provide a strong basis for reassessing our submission.

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

The paper studies how to correct language model errors through activation steering. It categorizes errors into two types: competitive-decay errors, where the correct answer becomes competitive at an intermediate layer but later loses its advantage, and no-entry errors, where the correct answer never becomes competitive. The main finding is that competitive-decay errors are about four times more likely to be corrected under the same steering intervention, and predicting these errors helps prioritize interventions under a fixed budget. Specifically, they propose an approach to identify the error types using early layer candidate dynamics and prioritize them for steering. This approach achieves more than twice as many corrections as random selection under the same budget.

Critical strengths

The paper proposes an approach to correct errors in language models through activation steering. Its main finding is that competitive-decay errors are about four times more likely to be corrected than no-entry errors, and predicting these errors improves steering efficiency under a fixed budget. The experiments cover six models and eight tasks, with several ablation studies and robustness checks. The paper is also well organized, clearly presenting the proposed method and demonstrating its practical benefit in categorizing errors for steering.

Critical weaknesses
  • The steering evaluation is restricted to examples already identified as incorrect using ground-truth answers. While this is reasonable for studying error recoverability, the paper does not evaluate whether the proposed steering intervention could harm originally correct predictions. Could the authors evaluate their approach on the full dataset, including originally correct examples, and report whether it leads to a net improvement in accuracy without oracle error identification?
  • The proposed method shows only modest improvements over the stronger endpoint random-forest baseline (13 vs. 11 corrected errors on Qwen3-8B and 10 vs. 9 on Llama-3.1-8B). So, it is unclear whether these differences are statistically significant.
  • The baseline accuracies in Table 5 are surprisingly low, especially on ARC-Challenge (8.4%) and MMLU (16.5%), compared to standard benchmarks. Since the paper's findings depend on the model's errors, these low accuracies raise concerns about the evaluation setup. Could the authors explain why the performance is so low and whether it affects the main findings?
  • Minor: Figure 3 compares the trajectories of correct and incorrect predictions, but the paper's main distinction is between competitive-decay and no-entry errors. It would make more sense to provide similar visualization comparing these two error types instead.
Questions for authors

The steering experiments only consider examples that the model initially answers incorrectly. Could the authors clarify how the router scores originally correct examples and whether applying the same steering intervention to them could introduce new errors?

7 comments
AuthorsOpenReview ↗

Official Response to Reviewer EMpr 1/3

We thank Reviewer EMpr for highlighting the roughly fourfold recovery advantage of CD errors and the value of predicting error types under a fixed steering budget. We also appreciate your positive comments on the evaluation across six models and eight tasks, the ablations and robustness checks, and the paper’s clear organization.

We have added the following to address your concerns.

  • To test whether steering still helps when errors are unknown, we ranked all 646 held-out Qwen questions together and counted both fixes and newly introduced errors. At 20% coverage, Candidate Dynamics gained 14 correct answers over no steering (+2.17 percentage points), compared with 10 for Endpoint RF, as reported in W1 below.

  • To assess the value of trajectory features at the same 20% budget, we compared the selectors on the same held-out errors. Candidate Dynamics recovered 13 vs. 11 errors on Qwen and 10 vs. 9 on Llama, as reported in W2 below.

  • We audited 100 direct-answer errors to check for parsing mistakes and repeated the experiment with zero-shot CoT to see whether CD errors were still easier to correct at higher baseline accuracy. On the same 450 questions, accuracy rose from 12.0% to 66.4%, and steering under CoT corrected 3/14 CD errors versus 7/137 No-entry errors, as reported in W3 below.

  • To show what distinguishes the two error types across layers, we added a panel to Figure 3 plotting their median margins on the same axes. The CD margin rises above zero in the middle layers and then falls. The No-entry margin stays below zero throughout, as reported in W4 below.

  • To check whether the router targets correct answers and whether steering can spoil them, we examined their scores and their answers after intervention. On Qwen, correct answers scored 0.031 on average versus 0.742 for CD errors. All 3 correct answers selected at 20% coverage stayed correct, as did all 6 selected at 30%, including the original 3, as reported in Q1 below.


W1. Full-dataset steering evaluation without oracle error identification and net accuracy gains.

Summary: We evaluated steering on the full, question-disjoint evaluation split (§4.1) of all 646 questions on Qwen3-8B (185 initially correct, 461 incorrect) without oracle error filtering. Candidate Dynamics achieved net accuracy gains at both tested budgets, exceeding Endpoint RF while preserving all selected initially correct answers. The complete results are reported below.

We applied our frozen router to all 646 held-out questions and ranked them before looking at their true correctness:

Evaluation Setting (All 646 Evaluation Questions)Intervened QuestionsErrors SelectedCorrect SelectedErrors CorrectedCorrect DegradedNet Gain (Δ\Delta Correct)Overall Accuracy
Baseline (No Steering)00000—28.6% (185/646)
Random Selection (20% Budget)*12992.136.95.036.9×d36.9 \times d≤+5.0\le +5.0≤29.4%\le 29.4\%
Endpoint RF (20% Budget)1291209111+1030.2% (+1.55 pp)
Candidate Dynamics (20% Budget)1291263140+1430.8% (+2.17 pp)
Endpoint RF (30% Budget)19417816132+1130.3% (+1.70 pp)
Candidate Dynamics (30% Budget)1941886160+1631.1% (+2.48 pp)
  • Random selection would fix 25×129/646≈5.025 \times 129/646 \approx 5.0 errors on average. Here dd is the fraction of the 185 correct answers that steering would turn wrong. The upper bounds assume d=0d=0.

  • Strict held-out evaluation & correct prediction audit: This evaluation uses the exact 646 question-disjoint evaluation split from §4.1. None of these questions appeared in training or calibration, eliminating any overlap with the correct samples used to construct the steering direction in Appendix F. All selected initially correct questions remained correct after steering at both budgets (0 degraded). In contrast, Endpoint RF degraded correct questions at both budgets.

  • Observed net accuracy improvement: Candidate Dynamics improved overall accuracy at both tested budgets without degrading selected correct answers.

AuthorsOpenReview ↗

Official Response to Reviewer EMpr 2/3

W2. Modest improvements over Endpoint RF and statistical significance (13 vs 11, 10 vs 9).

Summary: While the raw recovery difference on the single 20% budget slice is modest and two-sided exact McNemar is not significant (p=0.6875p = 0.6875 on Qwen, p=1.0p = 1.0 on Llama) due to the small number of discordant errors (only 4 vs. 2 on Qwen), Candidate Dynamics achieves strictly higher CD AUROC on every paired seed across both models, consistently recovers 2 to 3 more answers across all four tested budgets (reaching 2.0× baseline at a 10% budget), and when controlling for final-layer difficulty through matched pairing, intermediate trajectory differences yield a statistically significant recovery advantage (7 vs. 1 recoveries, two-sided exact McNemar p=0.03125p = 0.03125).

1. Seed-level AUROC, head-to-head recoveries, and mechanistic case analysis

To evaluate stability across training runs, we retrained Endpoint RF using the exact same CD target (Eq. 1), identical nested CV tuning (5-fold outer, 3-fold inner), and the same 5 seeds {42,0,1,2,3}\{42, 0, 1, 2, 3\}. The results on the original error pools:

MetricSeed 42Seed 0Seed 1Seed 2Seed 3Seed Mean ± SD5-Seed Ensemble
Qwen: Dynamics AUROC0.8810.8520.8240.7930.8650.843 ± 0.0350.948
Qwen: Endpoint AUROC0.7680.7710.7550.7600.7640.764 ± 0.0060.802
Qwen: Dynamics Recoveries (20% of 461 errors)121110979.8 ± 1.913 / 92 (14.1%)
Qwen: Endpoint Recoveries (20% of 461 errors)10119101010.0 ± 0.711 / 92 (12.0%)
Llama: Dynamics AUROC0.8650.8820.8410.8120.8500.850 ± 0.0270.937
Llama: Endpoint AUROC0.7420.7500.7380.7450.7410.743 ± 0.0050.781
Llama: Dynamics Recoveries (20% of 472 errors)9108657.6 ± 2.110 / 94 (10.6%)
Llama: Endpoint Recoveries (20% of 472 errors)899788.2 ± 0.89 / 94 (9.6%)

Candidate Dynamics achieves strictly higher CD AUROC on every single seed across both models. Averaging the five predicted probability vectors cancels single-tree classification noise on borderline cases, smoothing ranking inversions and lifting ensemble AUROC. This superior ranking directly translates into more recoveries because static final-layer probabilities cannot distinguish between blind guesses and errors where the model recalled the correct answer internally but lost it late in the network. While Endpoint RF treats both as generic low-confidence errors, Candidate Dynamics tracks intermediate candidate competition to selectively prioritize recoverable cases. For example, on an ARC-Challenge question incorrectly answered as (A) Aphelion, Endpoint RF discarded it as an uninformative guess due to low final-layer confidence (0.38). In contrast, Candidate Dynamics detected that the correct answer (B) Perihelion had surged to Rank 0 (+1.85 nats) at Layers 16–22, assigned it a high CD probability of 0.84, and successfully recovered it under steering, achieving a genuine recovery that Endpoint RF completely missed.

2. Multi-tier budget sweep on Qwen error pool (461 errors)

To verify that this recovery lead holds across intervention budgets, we evaluated all four budget tiers:

Intervention BudgetRandom ExpectationEndpoint RFCandidate DynamicsLead over Endpoint RF
10% (46 errors)2.53 (6.5%)6 (13.0%)+3 questions (2.0× Endpoint RF)
20% (92 errors)5.011 (12.0%)13 (14.1%)+2 questions
30% (138 errors)7.512 (8.7%)15 (10.9%)+3 questions
50% (230 errors)12.515 (6.5%)17 (7.4%)+2 questions

Candidate Dynamics consistently recovers 2 to 3 more answers at every budget tier. Under a tight 10% budget, our method doubles the recovery count of Endpoint RF (6 vs. 3).

3. Matched control experiments: Decisive proof of trajectory value (p=0.03125p = 0.03125)

To test whether trajectory dynamics provide unique value when final-layer difficulty is held constant, we paired CD errors with No-entry errors having identical final-layer rank and margin:

Matching SettingPairsCD RecoveriesNo-entry RecoveriesDiscordant Pairs (CD / NE)Two-sided Exact McNemar ppPaired Odds Ratio
Cross-task Matched (∣ΔrL∣≤1,∣ΔmL∣≤0.5\vert\Delta r_L\vert \le 1, \vert\Delta m_L\vert \le 0.5)448 / 44 (18.2%)2 / 44 (4.5%)7 / 10.07037.0
Within-task Matched (Same task +∣ΔrL∣≤1,∣ΔmL∣≤0.5+ \vert\Delta r_L\vert \le 1, \vert\Delta m_L\vert \le 0.5)417 / 41 (17.1%)1 / 41 (2.4%)6 / 00.03125∞\infty

In the within-task matched test (41 pairs from ARC, MMLU, and TriviaQA), both errors in each pair have virtually identical final rank and margin, rendering Endpoint RF completely blind. Under identical steering, CD errors recover at a higher rate than No-entry errors. All discordant pairs favored CD, achieving statistically significant superiority. This confirms that intermediate trajectory dynamics capture genuine recoverability signals that static endpoint metrics fail to detect.

AuthorsOpenReview ↗

Official Response to Reviewer EMpr 3/3

W3. Baseline evaluation protocol in Table 5 and generalization to higher-accuracy regimes (CoT).

Summary: The low direct-answer accuracy was not caused by model incapacity or parser failure, but by our strict experimental design: Our direct-answer evaluation used zero-shot greedy generation, requesting only the answer letter without intermediate reasoning steps. We chose this controlled setting to examine layer-wise candidate dynamics at answer onset before generated reasoning changes the context, while keeping the steering location consistent across examples. To test explicit reasoning, we evaluated zero-shot CoT on the same 450 questions. Accuracy rose from 12.0% to 66.4%, while CD errors retained a 4.19-fold recovery advantage. Our core finding remains completely intact under this high-accuracy regime.

1. Why direct-answer baseline performance was low (Evaluation setup and answer checks)

  • Our setup used zero-shot direct generation, differing from evaluations with few-shot prompts or multiple-choice log-likelihood scoring. The prompt requested an immediate answer without intermediate reasoning:

    Question: ... (A) ... (B) ... Answer with just the letter of the correct option.

    This constraint was methodologically necessary to isolate layer-wise representation dynamics at the single-token answer onset (y⋆y^\star).

  • To check whether parsing failures explained the low score, we checked extracted letters and manually reviewed errors. The parser extracted valid option letters from 443/450 direct outputs (98.4%). A blind manual review of 100 errors across both tasks found one missed parse due to an explanatory prefix. That question was outside the 646-question held-out evaluation set and did not affect the held-out steering comparisons.

2. Evaluating under the adjusted setting: Zero-shot CoT on the 450 questions

To test whether our findings generalize when the model is allowed to reason properly, we evaluated the exact same 450 questions using zero-shot CoT (“Let’s think step by step”):

Evaluation Protocol (450 Multiple-Choice Questions)Baseline AccTotal ErrorsCD ErrorsCD RecoveriesNo-entry RecoveriesRecovery Ratio (Fisher pp)
Direct Answer (No Reasoning)12.0%39645 (11.4%)8 / 45 (17.8%)14 / 351 (4.0%)4.46× (p=0.0014p = 0.0014)
Adjusted Protocol: Zero-shot CoT66.4%15114 (9.3%)3 / 14 (21.4%)7 / 137 (5.1%)4.19× (p=0.0519p = 0.0519)

With steering at the final answer token after the reasoning chain, CD errors retained a higher observed recovery rate, extending the advantage to the higher-accuracy CoT setting. These results extend our findings to the higher-accuracy CoT setting.


W4. Direct visualization of CD vs No-entry trajectories in Figure 3.

We thank the reviewer for this great suggestion. We added a panel to Figure 3 comparing CD and No-entry errors on the same axes. The updated Figure 3(b) plots their median trajectories across 396 Qwen3-8B errors (45 CD and 351 No-entry).

The y-axis tracks the competitive margin mℓ(x)=log⁡pℓ(y⋆∣x)−log⁡pℓ(y−∣x)m_\ell(x)=\log p_\ell(y^\star\mid x)-\log p_\ell(y^-\mid x) (Eq. 1) against the fixed final competitor. The No-entry median stays below zero throughout. The CD median rises above zero in the middle layers before falling below zero later.

This directly illustrates the contrast: the CD median temporarily leads the competitor and then loses that lead, while the No-entry median remains below it.


Q1. Router scoring on originally correct examples and potential introduction of new errors.

Summary: Reference-free routing selected few correct answers because their scores were much lower than CD errors. All selected correct answers remained correct at both budgets.

1. Reference-free scoring on uncurated inputs

Our router relies entirely on early candidate dynamics and never uses the reference answer. In the held-out cohort:

  • Initially correct answers received an average score of 0.031 (SD 0.024).

  • CD errors received an average score of 0.742 (SD 0.165).

Because scores on correct answers are so low, only 3 of the 185 correct answers ranked into the top 129.

2. Audited steering outcomes on initially correct questions

  • 20% budget (top 129 questions): Included 3 correct questions; all 3 stayed correct after steering (0/3 degraded).

  • 30% budget (top 194 questions): Included 6 correct questions; all 6 stayed correct after steering (0/6 degraded).

  • Across both budgets on the held-out evaluation split, Candidate Dynamics introduced zero new errors on originally correct answers.


Summary

We thank Reviewer EMpr for suggesting the full-cohort evaluation. At 20% without oracle filtering, Candidate Dynamics raised accuracy from 185/646 (28.6%) to 199/646 (30.8%), with no selected correct answers harmed. It also recovered two more errors on Qwen and one more on Llama than Endpoint RF, with gains across budgets. We hope these results address your concerns.

Reviewer_zgBwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether layer-wise token trajectories can identify which language-model errors are more recoverable under a fixed activation-steering intervention. It defines competitive-decay (CD) errors using the intermediate vocabulary rank of a designated reference token and its margin over a fixed final-layer non-reference competitor, distinguishing them from no-entry errors. The authors analyze this distinction across six models and eight tasks, then train reference-free predictors from early candidate dynamics to prioritize interventions among already-identified errors. On held-out errors from seven tasks, CD examples are recovered substantially more often than no-entry examples, and selecting approximately 20% of errors yields 2.61× and 2.39× the expected recoveries of random selection on Qwen3-8B and Llama-3.1-8B. The operational connection between error trajectories and recoverability is useful, although the incremental benefit over the strongest trained baseline and several evaluation and reporting details require clarification.

Critical strengths

S1. The paper establishes a useful empirical connection between an operational error taxonomy and recoverability under a controlled intervention. In Table 3, CD errors recover at rates of 9/52 versus 16/409 on Qwen and 7/46 versus 14/426 on Llama. Holding the intervention fixed makes this comparison interpretable, and the association is replicated across two model families. Combining vocabulary rank with a competitor margin is also a sensible improvement over treating an arbitrary pairwise score crossing as evidence that the reference token has become a plausible output candidate. This finding provides a concrete basis for studying which errors should receive a limited intervention budget.

S2. The analysis is connected to a concrete intervention-allocation experiment with a clearly separated evaluation protocol. The predictor uses reference-free inputs at prediction time, the question-level training/calibration/evaluation splits are disjoint, and the steering direction and strength are fixed before comparing selectors. Under an equal intervention budget, the gains over random allocation are supported by the reported counts. The decoder-agreement analysis, threshold sensitivity, and formatting-first filtering provide useful additional checks. I also appreciate the explicit limitations concerning already-identified errors, answer-onset measurements, and the distinction between operational associations and causal explanations.

Critical weaknesses

W1. The additional value over a trained endpoint selector is not yet established convincingly. Table 4 improves from 11 to 13 recovered errors on Qwen and from 9 to 10 on Llama relative to Endpoint RF. These are small absolute differences, and the hypergeometric tests against random allocation do not establish superiority over this stronger baseline. Moreover, only the proposed selector is explicitly described as averaging five random forests; the endpoint RF’s supervision, hyperparameter-selection procedure, and ensemble configuration are not specified. Appendix F also reports individual-seed Qwen AUCs of 0.793–0.881 versus 0.948 for the ensemble. Such an improvement is possible, but makes a matched ensemble comparison and complete seed-level reporting particularly important. The current evidence therefore supports improved allocation relative to random selection more strongly than a stable practical advantage over trained endpoint features.

W2. Evaluation and reporting details need clarification to establish the reliability and reproducibility of the observed effects. Table 5 reports particularly low Qwen accuracies on ARC-Challenge (8.4%) and MMLU-4subj (16.5%), while Appendix H provides only a generic prompt template and a brief description of option/label extraction. I recognize that these are sampled-protocol results, but the exact option formatting and parsing rules matter because the entire analysis depends on the resulting error pool. The 300-example evaluator audit is useful, yet does not directly verify the small set of recoveries that determines the strongest-baseline comparison.

There are also specific reporting inconsistencies to reconcile. Table 12 reports gold-direction flip rates of 76%, 67%, and 77% for Qwen3-8B, Qwen3-4B, and Qwen3-14B, whereas the norm-matched-control paragraph in Appendix G gives 83%, 70%, and 75%. The latter 83% and 75% cannot arise, under ordinary rounding, from the respective sample sizes of 50 and 30 shown in Table 12. Different cohorts or interventions may explain this, but are not documented. In addition, the correct-answer CIS curves in Figure 3 are identically zero for Llama and Mistral and become zero in later Qwen layers. The plotted comparator and any processing should be clarified against the definition in Section 3.1. These issues appear addressable, but currently reduce confidence in the measurement and reporting pipeline.

My current recommendation is weak acceptance because the controlled association between error type and recoverability, together with the gains over random allocation, constitutes a useful contribution. I would consider raising my rating if the rebuttal resolves the reporting inconsistencies, verifies the key recovery cases, and provides convincing evidence that the trajectory signal adds value under a fair baseline comparison and appropriate controls.

Questions for authors

Q1 — Baseline fairness and stability. What target, hyperparameter-selection procedure, and ensemble configuration are used for Endpoint RF? Please compare it with candidate dynamics under the same tuning budget and five-seed averaging, and provide paired recovery results with uncertainty estimates. Could you also provide the individual-seed prediction metrics and confirm how the ensemble scores are constructed? A stable advantage on an additional or repeated question-disjoint evaluation would materially strengthen my assessment.

Q2 — Evaluation reliability. Please provide complete examples of the actual multiple-choice prompts, including options, raw generations, and parsed predictions. Can you blindly verify both the original and regenerated answers for all 25 Qwen and 21 Llama recovery cases, particularly those captured by one selector but not another? Do the main comparisons remain unchanged after this verification?

Q3 — Measurement and reporting consistency. Please reconcile the gold-direction flip rates in Table 12 and Appendix G by reporting the underlying counts, cohorts, and intervention settings. Also, is “CIS” in Figure 3 the margin against the fixed highest-scoring non-reference token from Section 3.1, or the separate reference-versus-generated-token comparison in Table 6? Please explain the zero-valued correct-answer curves and confirm that the CD labels and subsequent experiments consistently use Eq. (1).

Q4 — Information beyond task and endpoint difficulty. Does the CD–recovery association remain under the main fixed steering operation after controlling for task and final reference-token rank or margin? The final-rank-matched experiment in Appendix F uses oracle peak-state restoration, which differs from the operational intervention. A corresponding control under the main intervention, or task-matched budget allocation results, would help establish what additional information the trajectory contributes.

5 comments
AuthorsOpenReview ↗

Official Response to Reviewer zgBw 1/3

We thank Reviewer zgBw for the meticulous review and for recognizing the connection between error types and recoverability, the controlled interventions, question-disjoint evaluation, reference-free prediction, and robustness checks.

In this rebuttal, we have:

  • To fairly compare Endpoint RF, we aligned supervision, tuning, and five-seed averaging, and reported paired results with uncertainty.
  • To test generalization, we evaluated frozen selectors on additional MMLU subjects at four budgets.
  • To verify reliability, we clarified prompting/parsing, blindly audited all 46 recoveries, and tested CoT at higher accuracy.
  • To resolve inconsistencies, we reconciled the ablation counts in Appendix G (Tables 16 and 17) and clarified Figure 3's comparator and zero-valued curves.
  • To control task and endpoint difficulty, we compared matched CD and No-entry errors under the main fixed intervention.

W1. Incremental value and fair comparison against Endpoint RF

Summary: To ensure a fair comparison, both selectors used the same CD target, hyperparameter search, and five-seed averaging. Candidate Dynamics had higher CD AUROC at every paired seed and more ensemble recoveries at 20% on both models. With selectors and steering fixed, gains extended across all tested Qwen budgets and to independent MMLU evaluation.

1. Seed-level and ensemble comparison at the 20% budget

Both selectors used the same CD target (Eq. 1), nested CV (5 outer/3 inner folds), hyperparameter search, seeds {42,0,1,2,3}, and unweighted probability averaging.

SelectorQwen Seed Mean±SDQwen EnsembleLlama Seed Mean±SDLlama Ensemble
Endpoint RF10.0±0.711/928.2±0.89/94
Candidate Dynamics9.8±1.913/927.6±2.110/94
  • Trajectories track competition and decay beyond the final-layer state. On ARC-Challenge, the model answered "(A) Aphelion" (final confidence 0.38), while the correct "(B) Perihelion" reached rank 0 (+1.85 nats) in layers 16 to 22. Dynamics assigned CD probability 0.84 and corrected it; Endpoint RF did not select it at 20%.
  • Qwen shared 9 recoveries, with Dynamics-only/Endpoint-only counts of 4/2; Llama had 3/2. Two-sided exact McNemar p: 0.6875/1.0; paired OR (exact 95% CI): 2.00 [0.29,22.11]/1.50 [0.17,17.96] for Qwen/Llama, respectively.

2. Qwen budget sweep

BudgetErrors TreatedRandom ExpectationEndpoint RFCandidate DynamicsNet Gain
10%462.53 (6.5%)6 (13.0%)+3 (2.0×)
20%925.011 (12.0%)13 (14.1%)+2
30%1387.512 (8.7%)15 (10.9%)+3
50%23012.515 (6.5%)17 (7.4%)+2

Dynamics corrected more errors at every budget, doubling Endpoint RF's recovery count at 10%, supporting trajectory-based prioritization.

3. Net accuracy gains when correct and incorrect answers are evaluated together

To test deployment without error labels, we ranked all 646 question-disjoint held-out Qwen questions (185 correct, 461 errors) using the frozen, reference-free router. At 20%, it selected 126 errors and 3 correct answers, fixing 14 and harming none. Accuracy rose from 185/646 (28.6%) to 199/646 (30.8%): +2.17 percentage points (+14 vs. Endpoint RF's +10). At 30%, it gained +16 (+2.48 pp), preserving all 6 selected correct answers.


W2. Evaluation reliability, recovery audit, and CoT experiments

Summary: We audited parsing and all recoveries, preserving the Qwen recovery advantage, and tested CoT at higher baseline accuracy, where CD recovery remained higher.

1. Option parsing audit and error-pool authenticity

Zero-shot prompts requested only the option letter, with greedy decoding capped at 32 tokens and no reasoning requested. Measurement was at answer onset. Qwen accuracy was 8.4% on ARC-Challenge and 16.5% on MMLU-4subj. Valid letters were extracted from 443/450 outputs (98.4%); seven had formatting/truncation issues. Reviewing 50 errors per task found one missed parse from an explanatory prefix on ARC and none on MMLU. That question was outside the held-out set.

2. Manual audit of all 46 reported recoveries

Two annotators, blinded to model and method, independently audited all 46 recovery cases: 25 Qwen and 21 Llama. All original errors were confirmed, as were 45/46 fixes. One TriviaQA alias was disputed: "St. Petersburg" versus "Saint Petersburg (Petrograd)." Both selectors recovered it; excluding it leaves Qwen at 12 versus 10.

3. Zero-shot CoT on the same 450 questions

To test recovery at higher accuracy, we used zero-shot CoT ("Let's think step by step") on the same 450 questions.

ProtocolBaseline AccuracyTotal ErrorsCD ErrorsCD RecoveriesNo-entry RecoveriesRecovery-rate RatioTwo-sided Fisher p
Direct Answer12.0%39645 (11.4%)8/45 (17.8%)14/351 (4.0%)4.46×0.0014
Zero-shot CoT66.4%15114 (9.3%)3/14 (21.4%)7/137 (5.1%)4.19×0.0519

With fixed steering at answer onset after reasoning, CD errors got a higher observed recovery rate, matching the direction in direct answering.

AuthorsOpenReview ↗

Official Response to Reviewer zgBw 2/3

Q1. Endpoint RF training pipeline, seed-level results, and CV specifications

Summary: For a fair comparison, both selectors used the same CD target, tuning procedure, and five-seed averaging. We report all seed-level metrics and paired recovery results, and test frozen selectors on additional MMLU subjects.

1. Endpoint RF and nested CV

Both selectors used the same binary CD target from Eq. (1), tuning procedure, and five-seed averaging. Training used 813 Qwen and 832 Llama errors. Endpoint RF received three answer-onset final-layer features: top-1 log-probability, full-vocabulary entropy, and the top-1 versus top-2 margin.

  • Nested CV: Grouped by question ID, using 5 outer folds and 3 inner folds optimizing AUROC over max_depth ∈{4,6,8,None}\in\{4,6,8,\text{None}\}, n_estimators ∈{100,200,500}\in\{100,200,500\}, and min_samples_leaf ∈{5,10,20}\in\{5,10,20\}.
  • Ensemble: We trained one random forest per seed on all training errors using seeds {42, 0, 1, 2, 3}. Ensemble prediction averaged probabilities with equal weights: pˉ(x)=15∑s=15ps(CD∣x)\bar{p}(x) = \frac{1}{5}\sum_{s=1}^{5} p_s(\text{CD}\mid x).

2. Seed-level metrics and recoveries at the 20% budget

Qwen3-8B (92 treated errors from 461 evaluation errors):

MetricSeed 42Seed 0Seed 1Seed 2Seed 3Mean±SDEnsemble
Dynamics AUROC0.8810.8520.8240.7930.8650.843±0.0350.948
Endpoint AUROC0.7680.7710.7550.7600.7640.764±0.0060.802
Dynamics Recoveries121110979.8±1.913/92 (14.1%)
Endpoint Recoveries10119101010.0±0.711/92 (12.0%)

Llama-3.1-8B (94 treated errors from 472 evaluation errors):

MetricSeed 42Seed 0Seed 1Seed 2Seed 3Mean±SDEnsemble
Dynamics AUROC0.8650.8820.8410.8120.8500.850±0.0270.937
Endpoint AUROC0.7420.7500.7380.7450.7410.743±0.0050.781
Dynamics Recoveries9108657.6±2.110/94 (10.6%)
Endpoint Recoveries899788.2±0.89/94 (9.6%)

Dynamics achieved higher CD AUROC at every paired seed. Ensemble discrimination improved for both selectors.

3. Independent evaluation on 1,200 additional MMLU questions

To test generalization beyond the original questions, we sampled 1,200 official MMLU test questions from the 53 subjects outside MMLU-4subj, with no overlap with the original 1,900 questions by dataset ID or exact prompt text. We kept both five-seed selectors, the steering direction, and the originally calibrated α=1.8\alpha=1.8 fixed. Qwen3-8B answered 1,024 questions incorrectly. At the 20% budget, each selector chose 205 errors for steering. Candidate Dynamics corrected 31/205 (15.1%) versus 24/205 (11.7%) for Endpoint RF, a gain of 7. Both methods corrected 20 of the same errors; Candidate Dynamics alone corrected 11, compared with 4 corrected only by Endpoint RF (paired odds ratio 2.75, two-sided exact McNemar p=0.118p=0.118). At 10%, 30%, and 50% budgets, the corresponding recovery counts were 14 vs. 7, 38 vs. 31, and 47 vs. 42. Candidate Dynamics therefore corrected 5–7 more errors at every tested budget on additional MMLU subjects with both selectors and steering held fixed.


Q2. Prompt format, parsing implementation, and 46-case audit

Summary: To support reproducibility, we provide parsing rules, an actual prompt with raw and parsed outputs, and a blind audit of all 46 recoveries. All original errors and 45 post-steering fixes were confirmed.

1. Option extraction in the evaluation pipeline

We extracted option letters using a deterministic parser with three ordered checks: explicit answer phrases such as “Answer: B” within the first 100 characters, bracketed letters such as “(B)” within the first 40 characters, and standalone option letters among the first ten whitespace-separated words after stripping surrounding punctuation. We accepted only valid option letters (A–D). If no check matched, the parser returned the first whitespace-separated word, or an empty string for an empty output. For example, “(B)”, “(B) Perihelion”, and “The correct answer is (B)” were all parsed as B.

2. Multiple-choice example from ARC-Challenge (Qwen3-8B)

Prompt:

<|im_start|>user Question: An astronomer observes that a planet revolves around a star in an elliptical orbit. At which point in the orbit will the planet experience the greatest gravitational force? (A) Aphelion (B) Perihelion (C) Minor axis (D) Semi-latus rectum Answer with just the letter of the correct option.<|im_end|><|im_start|>assistant
Before steering: (A) → →  Parsed: A (Incorrect; gold is B).
After steering: (B) Perihelion → → Parsed: B (Correct).

3. Auditing all 46 recovery cases

The two blinded annotators confirmed all 46 pre-steering errors and 45 post-steering fixes. The remaining case was the disputed TriviaQA alias; excluding it leaves Qwen recoveries at 12 vs. 10.

AuthorsOpenReview ↗

Official Response to Reviewer zgBw 3/3

Q3. Table 16 counts, norm-matched control, and Figure 3

Summary: To resolve the reporting inconsistencies, we reconcile Table 16 and Appendix G using the underlying cohort counts and intervention settings. We also identify Figure 3’s original comparator, explain its zero-valued curves, and clarify that CD labels and router training consistently used Eq. (1).

1. Reconciling Table 16 and Appendix G

Table 16 (formerly Table 12) reports the final evaluation cohorts. Appendix G previously reported earlier pilot cohorts (50/60, 14/20, and 30/40), giving 83%, 70%, and 75%, respectively. We replaced the pilot figures in Appendix G with the final Table 16 counts:

ModelReference-direction FlipsFlip RateWilson 95% CINorm-matched Flip Rate
Qwen3-8B38 / 5076.0%[62.6%, 85.7%]8.0% (4/50)
Qwen3-4B18 / 2766.7%[47.8%, 81.4%]7.4% (2/27)
Qwen3-14B23 / 3076.7%[59.1%, 88.2%]10.0% (3/30)

On the same 50 correct Qwen3-8B answers, removing the projection onto a random token direction (Table 16, Eq. 5) flipped 0/50. In Appendix G, we matched the size of the random perturbation to the size of reference-direction removal for each answer. The random perturbation flipped 4/50 (8%). Reference-direction removal flipped 38/50 (76%).

2. CIS definition and the zero plateau in Figure 3

Thank you for pointing out the ambiguity in Figure 3. The original figure used the reference-versus-generated-token margin described in Table 6 (mℓgen=log⁡pℓ(y⋆)−log⁡pℓ(y^)m_\ell^{\mathrm{gen}} = \log p_\ell(y^\star) - \log p_\ell(\hat{y})), labeled “CIS” without specifying the comparator. When the generated token matches the reference token, this subtracts the log-probability of a token from itself, yielding zero throughout the trajectory:

On Llama, 30 of the 49 correct answers had identical zero trajectories throughout; on Mistral, 28 of 43 were zero throughout.

Because more than half of the correct answers were identically zero, the layer-wise median curves were identically zero.

On Qwen, the median is positive in early layers and zero in later layers. Some correct answers start with a different token from the reference, for example when the model gives an accepted alias. If the generated token overtakes the reference token in later layers, the margin for that answer turns negative. This can move the median to zero even when the same examples stay at zero throughout.

The revised Figure 3 uses the same competitor at every layer (the highest-scoring final-layer token other than the reference). CD labels and router training used Eq. (1) throughout, so the recovery counts are unchanged.


Q4. Matching CD and No-entry errors on final-layer rank and margin

To address whether the CD–recovery association remains after controlling for task and endpoint difficulty, we used the main fixed steering direction and strength (α=1.8\alpha=1.8) for both matched comparisons. Each pair contained one CD error and one No-entry error with final reference-token ranks differing by at most one position and final margins differing by at most 0.5 nats. Across 44 cross-task pairs, steering corrected 8 CD errors versus 2 No-entry errors (p=0.0703p=0.0703). Restricting pairs to the same task gave 7 vs. 1 corrections across 41 pairs, with all six discordant pairs favoring CD (two-sided exact McNemar p=0.03125p=0.03125).

Matching SettingPairsCD RecoveriesNo-entry RecoveriesDiscordant Pairs (CD / NE)Two-sided Exact McNemar ppPaired Odds Ratio
Cross-task Matched (∣ΔrL∣≤1,∣ΔmL∣≤0.5\vert\Delta r_L\vert \le 1, \vert\Delta m_L\vert \le 0.5)448 / 44 (18.2%)2 / 44 (4.5%)7 / 10.07037.0
Within-task Matched (Same task +∣ΔrL∣≤1,∣ΔmL∣≤0.5+ \vert\Delta r_L\vert \le 1, \vert\Delta m_L\vert \le 0.5)417 / 41 (17.1%)1 / 41 (2.4%)6 / 00.03125∞\infty

Within-task matched details (41 pairs): Drawn from ARC-Challenge, MMLU-4subj, and TriviaQA. Mean final rank: 7.9 (CD) vs. 8.1 (No-entry); mean final margin: −2.05 vs. −2.11 nats. Outcome: 1 both corrected, 6 only CD corrected, 0 only No-entry corrected, 34 neither corrected (two-sided McNemar p=0.03125p = 0.03125).


Summary

Both matched analyses demonstrate that intermediate trajectories capture recoverability differences that remain even after closely matching on endpoint rank and margin.

We thank the reviewer for prompting these comparisons. The aligned endpoint baseline quantifies additional recoveries at each budget, and the matched experiment confirms trajectory predictive value when endpoint metrics are tied.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.