acceptodds
Under review as a conference paper at ICLR 2027

MetaSteer: Context-Conditioned, Non-linear Steering via Attention-Projection Adaptation

Abstract

Steering large language models typically relies on linear, context-independent interventions in activation space, an assumption that recent work has challenged and that can induce an information bottleneck when a fixed representation must encode many behavioral distinctions. We introduce _MetaSteer_, a method that learns nonlinear interventions with context-dependent effects and applies them to attention projection matrices, producing activation effects that vary with the input context by construction and requiring no linear concept-geometry assumption. Framed as preference-based optimization, MetaSteer is trained once on a pooled preference corpus and transferred zero-shot to unseen concepts and out-of-distribution contexts. We find that, despite using low-rank adapters, MetaSteer induces structured, context-dependent changes in hidden-state trajectories while partially preserving aspects of their local trajectory dynamics, including velocity and curvature. We evaluate MetaSteer on three controlled text-generation benchmarks and three agentic settings across multiple model families and scales. MetaSteer matches or outperforms strong task-specific steering baselines on most aggregate comparisons in the zero-shot regime. Across the evaluated settings, stronger text-generation steering is associated with stronger agentic steering performance. We further discuss geometric trajectory effects, capability retention, and safety considerations raised by transferable steering.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
activation steering, representation engineering, interpretability, representation geometry, controllable text generation, low-rank adaptation, preference learning, zero-shot generalization, attention mechanisms

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_FhqeOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Paper Summary

MetaSteer attaches rank-16 LoRA adapters to the q/k/v/o attention projections of every layer and trains them with the standard DPO objective on a pooled corpus built from AxBench Concept16K / Concept16K-v2.

Training data.

  • Each example is rendered as ### Instruction\n{x}\n\n### Steering Direction\n{c}, where cc is a natural-language concept description.
  • y+y^+ is a concept-containing response and y−y^- is a concept-free response.

Validation.

  • After training once per backbone, the same adapter is evaluated "zero-shot" on:
    • AxBench, CLaS-Bench, and PersonalityBench (Table 1);
    • Dictator/Ultimatum games and SocialEval (Tables 2–3);
    • JailbreakBench (Fig. 2c);
    • MMLU and TruthfulQA (Table 5).
  • The paper argues that the induced effect is nonlinear and context-dependent, unlike fixed activation vectors.
  • It supports this with a hidden-state trajectory-geometry analysis (Sec. 6, App. I) and a CAA-alignment analysis (App. J).
Critical strengths
  1. Broad evaluation footprint. The paper evaluates six backbones from three families, three text-generation benchmarks, three agentic settings, a capability-retention check, and a safety probe. The agentic evaluation in particular goes beyond what most steering papers attempt.
  2. Transparent data and training protocol. App. M.5 documents pair construction, the concept-disjoint split, the fixed hyperparameters, and the absence of checkpoint selection. It is also candid that instructions are shared across splits (Table 10).
  3. Candid safety finding. The authors report that the steering-direction slot can be used to elicit harmful content (Fig. 2c, App. L) rather than omitting it.
  4. Judge-reliability effort. Multi-judge agreement (App. D) and a small human cross-check (App. E) are more than many steering papers provide.
Critical weaknesses
  1. The method is concept-conditioned LoRA-DPO, and the decisive control is missing. MetaSteer trains LoRA on q/k/v/o with standard DPO on prompts that contain the concept as text, and supplies the concept as text again at inference. The claimed distinguishing properties (nonlinearity, context dependence) hold for any LoRA fine-tune of attention. The paper never reports the same backbone with the same template and no adapter, which is needed to separate the effect of training from simply putting the concept in the prompt. Directly comparable concept-trained steering methods (HyperSteer, RePS) are also not run.

  2. The evidence for the distinctive claims does not measure what is claimed. The trajectory-geometry analysis, which supports a central abstract claim, appears to use Concept16K dataset responses rather than MetaSteer generations ("steps supplied by the dataset", App. I), with no steered-vs-unsteered comparison. The "zero-shot" claim lacks an overlap audit, even though training and AxBench evaluation concepts come from the same GemmaScope inventory and instructions are not held out (App. M.5).

Questions for authors

All questions are clarifications that could change my assessment.

  1. What are the results for each backbone with the same ### Steering Direction template and no adapter?
  2. In App. I, are the analyzed responses MetaSteer generations or dataset responses?
  3. What is the overlap of concepts and instructions between the training pool and the evaluation benchmarks?
6 comments

First, we thank Reviewer Fhqe for taking the time to read our research and for providing valuable feedback. Below are our responses to the critical weaknesses and our answers to the questions.

Critical Weakness 1: The method is concept-conditioned LoRA-DPO, and the decisive control is missing.

Part 1: Nonlinearity and context dependence hold for any LoRA fine-tune of attention. We acknowledge this. Our reason for including Equations 4–8 in the paper was to show that MetaSteer is mathematically sound, not to show uniqueness.

Part 2: Performance comparison with the adapter off. The following table shows the change in performance when the adapter is on versus off, using the same prompt template in both cases. The concept score consistently goes up while the instruction score goes down. The fluency score also goes up in most cases, which finally leads to a performance gain in the harmonic mean. This is consistent with the main claim of the paper, which is learning a multi-objective optimization that transfers to new settings.

Change from enabling the adapter (on minus off):

ModelΔConceptΔInstructionΔFluencyΔHarmonic
Gemma-2-2B+0.62−0.44+0.32+0.32
Gemma-2-9B+0.64−0.34+0.06+0.30
Llama-3.1-8B+0.66−0.09+0.04+0.26
Llama-3.2-3B+0.89−0.17+0.17+0.46
Qwen3-1.7B+0.47−0.11+0.27+0.29
Qwen3-4B+0.28−0.13+0.25+0.18

Part 3: Direct comparison with HyperSteer and RePS. The following table compares the performance of the technically adjacent methods that the reviewer asked about against MetaSteer. The experiment is currently limited to two Gemma backbones. The results suggest that MetaSteer significantly outperforms them.

MethodGemma-2-2BGemma-2-9B
Prompt, unadapted (same template, adapter off)1.191.22
RePS0.7560.892
HyperSteer0.7421.09
MetaSteer*1.511.52

Scores are the AxBench harmonic mean of concept, instruction and fluency scores.

*MetaSteer is not strictly held-out: 13/499 AxBench concepts (2.61%) and 348/2,954 instructions (11.78%) exactly match the training pool, and roughly 10% of concepts have a near-paraphrase in training (see overlap audit).

Critical Weakness 2: The evidence for the distinctive claims does not measure what is claimed.

Part 1: The trajectory-geometry analysis. The initially reported structure preservation was based only on dataset samples. During the rebuttal, we conducted the same analysis on text generated by MetaSteer and reached the same structure-preservation conclusion. In a complementary experiment, we changed the embedder model to MetaSteer, and the structure preservation remained.

Part 2: The "zero-shot" claim. Our overlap audit shows ~3% concept overlap, ~11% instruction overlap, and ~10% semantic duplication in the AxBench benchmark, while the other benchmarks remained reasonably non-overlapping. We acknowledge this, and in the next version we will limit this claim according to the results. In another complementary experiment, we investigated the association between dataset overlap and performance and found an association of |ρ| ≤ 0.17 for Axbench (see overlap audit).

Question 1

The answer to this question is already presented in our response to Critical Weakness 1, Part 2.

Question 2

The initial version only had ready-made dataset samples embedded using unadapted models (and the claim was that steering, as a general phenomenon, appears to preserve general structure while changing position). In the rebuttal, we extended the analysis to all four combinations: (dataset samples or generated samples) × (embedding with the unadapted model or the adapted model). We see the same pattern.

EmbedderDataAccumulationPositionVelocityAccelerationMenger
UnadaptedAxBenchOff+0.0536+0.2877+0.2668+0.2712
UnadaptedAxBenchOn+0.0114+0.2474+0.2449+0.0363
UnadaptedGeneratedOff+0.0420+0.1877+0.1636+0.1923
UnadaptedGeneratedOn+0.0059+0.1758+0.1599+0.1233
AdaptedAxBenchOff+0.0538+0.2876+0.2667+0.2715
AdaptedAxBenchOn+0.0111+0.2472+0.2449+0.0329
AdaptedGeneratedOff+0.0425+0.1878+0.1635+0.1956
AdaptedGeneratedOn+0.0072+0.1769+0.1613+0.1235

The table reports the similarity gain between steered texts relative to the shuffled control, aggregated over six models.

Question 3: Overlap audit

Setup. The reference set is the deduplicated union of the Concept16K v1 and v2 training splits (352 instructions, 27,422 concepts, 226,696 concept–instruction pairs). The evaluation sets are AxBench (499 concepts in the audited file, 2,954 instructions), CLaS-Bench (32 concepts, 2,240 instructions), PersonalityBench (793 concepts, 459 instructions), SocialEval/WorldTree (7 concepts, 1,015 instructions), and Dictator and Ultimatum (6 concepts and 1 instruction each). For each evaluation item, we compute (i) exact string match and (ii) its maximum similarity to any training item of the same type, using word-level Jaccard (lexical) and cosine similarity under sentence-transformers/all-mpnet-base-v2 (semantic; identical strings are excluded).

Exact match.

BenchmarkConcepts (exact / n)Instructions (exact / n)
AxBench13 / 499 (2.61%)348 / 2,954 (11.78%)
CLaS-Bench0 / 320 / 2,240
PersonalityBench0 / 7930 / 459
SocialEval (WorldTree)0 / 70 / 1,015
Dictator0 / 60 / 1
Ultimatum0 / 60 / 1

Maximum similarity to training (median / P95).

BenchmarkConcept JaccardConcept cosineInstruction JaccardInstruction cosine
AxBench0.556 / 0.8590.825 / 0.9660.286 / 1.0000.549 / 0.787
CLaS-Bench0.200 / 0.2100.452 / 0.5610.000 / 0.1070.260 / 0.367
PersonalityBench0.121 / 0.1490.568 / 0.6580.110 / 0.1350.317 / 0.376
SocialEval (WorldTree)0.143 / 0.2500.680 / 0.7170.061 / 0.0830.327 / 0.385
Dictator0.129 / 0.1430.709 / 0.7250.157 / 0.1570.469 / 0.469
Ultimatum0.129 / 0.1430.709 / 0.7250.161 / 0.1610.491 / 0.491

Findings.

  • AxBench is not disjoint from training. 13 concepts (2.61%) and 348 instructions (11.78%) match exactly. The instruction overlap reflects a shared instruction pool (App. M.5): 348 of the 352 training instructions appear in AxBench, while 2,606 of 2,954 AxBench instructions (88.2%) do not appear in training. The P90 concept cosine is 0.939, so roughly 10% of AxBench concepts have a near-paraphrase in training.
  • The other five benchmarks show no meaningful overlap. They have no exact matches, no item exceeds a Jaccard of 0.40 (concepts) or 0.33 (instructions), and the maximum cosine is at most 0.73 (concepts) and 0.54 (instructions). Values of roughly 0.5–0.7 between short concept descriptions reflect topical and stylistic relatedness rather than overlap.
  • Association with performance (AxBench). For ~10% of the generated evaluation instances per model (n = 3,000), we computed Spearman's ρ between each instance's maximum cosine similarity to training and its scores.
    • Instruction similarity vs. instruction score: ρ = 0.23–0.25 for all six backbones (p < 0.001).
    • Instruction similarity vs. harmonic mean: ρ = −0.005 to 0.18, significant for four of six backbones.
    • Concept similarity vs. concept score and harmonic mean: |ρ| ≤ 0.17, significant in 3 of 12 cells (p < 0.05, uncorrected).
  • Conclusion. We cannot rule out a small contribution from overlap, particularly through shared instructions. In the next version, we will restrict the zero-shot claim for AxBench to held-out concepts with shared instructions. The zero-shot claim for the other five benchmarks is unaffected.

Thank you for the responses and new experiments. The adapter-off control and the overlap audit address my Q1 and Q3. Remaining points:

  1. Adapter-off control (Q1). The adapter adds +0.18 to +0.46 HM over the same template without it, so the gain is not explained by the prompt alone. The gain comes from Concept (+0.28 to +0.89) at a cost in Instruction (−0.09 to −0.44), so please report the component scores alongside HM in the main table. Given your agreement that Eqs. 4–8 hold for any attention LoRA, I suggest framing the method as concept-conditioned preference fine-tuning, with the contribution being the transfer setting rather than a new steering mechanism.

  2. HyperSteer / RePS. Both score below the unadapted prompt (e.g., Gemma-2-9B: 1.09 / 0.892 vs. 1.22), so prompting remains the relevant reference. Please also clarify what "held-out" means for RePS, which is trained per concept.

  3. Overlap (Q3). Thank you for the audit. Since restricting the AxBench claim is planned, could you report AxBench HM on the subset excluding exact and near-paraphrase concepts (e.g., cosine > 0.9) and the 348 shared instructions? This would quantify the effect directly.

  4. Trajectory geometry (Q2). Using MetaSteer generations resolves my question. However, all rows still compare ordered vs. shuffled segments, and the adapted and unadapted embedders give nearly identical values (e.g., 0.0536 vs. 0.0538). This suggests the metric does not detect the adapter's effect. A matched comparison of steered vs. unsteered generations for the same prompts would be needed to support "preservation." Was the Steering Direction slot included when embedding with the adapted model?

If item 3 is reported and the framing is revised as in item 1, I would be glad to raise my score.

Reviewer_XrW2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MetaSteer, a steering method for LLMs. It trains LoRA adapters on all four attention projections (Wq,Wk,Wv,WoW_q, W_k, W_v, W_o) with standard DPO. The training data is a pooled preference corpus built from AxBench Concept16K and Concept16K-v2. The steering concept is given as text inside the prompt, under a fixed “Steering Direction” template. One adapter is trained per backbone and kept fixed. The authors argue that attention depends on the input, so the same fixed adapter produces different activation shifts in different contexts. This differs from adding a fixed steering vector. They give a short derivation showing that the induced change in the attention logit is a bilinear form in the query and key tokens. The method is evaluated zero-shot on AxBench, CLaS-Bench and PersonalityBench. It is also tested on agentic settings, namely the Dictator Game, the Ultimatum Game and SocialEval.

Critical strengths

The framing is useful. The paper treats steering as a design space with three axes: geometric assumption, intervention site and learning procedure. This helps place the work in a crowded area. Training a single concept-conditioned intervention that transfers to unseen concepts is a sensible goal. It is closer to what practitioners need than training one vector per concept. The empirical scope is broad. There are six backbones from three families and two scales each, three text benchmarks, and three agentic settings. The evaluation follows each benchmark’s own scoring pipeline. The setup is also quite careful in places.

Critical weaknesses

My main concern is about what the method actually is. MetaSteer is LoRA on q/k/v/o, trained with DPO, where the target behavior is written into the prompt. Functionally, this is fine-tuning the model to follow a new kind of instruction (“Steering Direction: X”). The arguments about context dependence and nonlinearity (Eqs. 4–8) hold for any LoRA fine-tuning of attention, and indeed for the base model itself. So they do not separate MetaSteer from ordinary parameter-efficient fine-tuning. The paper presents this as activation steering, but it does not have the usual properties of steering.

Questions for authors

Did you check for overlap, exact or near-duplicate, between the Concept16K training concepts and the AxBench evaluation concepts? How many evaluation concepts appear in training?

5 comments

First, we thank Reviewer XrW2 for taking the time to review our paper and for the valuable feedback. Below are our responses to the main concern and the question.

Main concern: Where MetaSteer sits and where its gains come from

Part 1: MetaSteer is mathematically sound, not unique. We agree that the context dependence and nonlinearity argued in Equations 4–8 hold for any LoRA fine-tune of attention, and indeed for the base model itself. We included these equations to show that MetaSteer is mathematically well-founded, not to claim that they distinguish it from parameter-efficient fine-tuning.

In the next version, we will state this explicitly and remove any wording that suggests otherwise. We will also describe MetaSteer as concept-conditioned preference fine-tuning used as a steering method. Its contribution is the transfer setting (a single adapter, trained once on pooled concept-conditioned preferences, that takes the concept as text and transfers to unseen concepts/contexts without per-concept training), not a new steering mechanism.

Part 2: Where the performance gain comes from. The most relevant reference is the same model with the same template and the adapter off (coefficient = 0), since supplying the concept in the prompt is the simplest alternative. The table shows the change from enabling the adapter (on minus off). Enabling the adapter raises the harmonic mean for all six backbones (+0.18 to +0.46) through the concept score (+0.28 to +0.89), at a cost in instruction score (−0.09 to −0.44).

ModelΔConceptΔInstructionΔFluencyΔHarmonic
Gemma-2-2B+0.62−0.44+0.32+0.32
Gemma-2-9B+0.64−0.34+0.06+0.30
Llama-3.1-8B+0.66−0.09+0.04+0.26
Llama-3.2-3B+0.89−0.17+0.17+0.46
Qwen3-1.7B+0.47−0.11+0.27+0.29
Qwen3-4B+0.28−0.13+0.25+0.18

For context, we also compare MetaSteer with technically adjacent trained methods:

  • HyperSteer [3]: the closest transferable method. A hypernetwork trained on pooled concepts generates the steering signal, using an architecturally identical model. We report the best variance among concept-only, concept + prompt and cross-attention.
  • RePS: a preference-trained steering vector added to the activations, trained per concept. It differs from MetaSteer in intervention site (residual stream vs. attention projections) and in per-concept training.
  • LoRA, LoReFT, SFT: standard fine-tuning methods trained on each single concept and evaluated on unseen instructions. LoRA is the closest to ordinary parameter-efficient fine-tuning.
MethodGemma-2-2BGemma-2-9B
Prompt, unadapted (same template, adapter off)1.191.22
RePS0.7560.892
HyperSteer0.7421.09
LoReFT0.7220.777
SFT0.714–
LoRA0.6410.602
MetaSteer*1.511.52

Scores are the AxBench harmonic mean of concept, instruction and fluency scores. MetaSteer partially overlaps with its training pool on AxBench (see Question 1).

All trained baselines score below the unadapted prompt on both backbones, consistent with the corresponding paper [1], where prompting is the strongest reference method on AxBench. We therefore treat the prompt as the primary reference and the trained baselines as secondary context.

Part 3: Steering properties. Following the definition in [1], steering applies an intervention to a model's activations to control its generations, most commonly by adding a fixed vector. A fixed vector is the usual instance, not a requirement of the definition. MetaSteer controls generation through its effect on activations: the adapter displaces hidden states by an amount that depends on context (Eq. 9). It has the following properties:

  • Concept supplied at inference: A new concept is specified as text (similar to HyperSteer [3]). No per-concept training or vector is needed.
  • Controllability and reversibility: As stated in (Eq. 3), scaling the coefficient λ_f changes the intensity of the intervention, and λ_f = 0 recovers the unadapted model. We have not characterized how steering strength varies with intermediate values of λ_f, so we make no claim about that.
  • Specificity (defined in [2]): Specificity is the precision of the modification: a highly specific method changes the targeted behavior while minimizing unintended changes to unrelated behaviors. MetaSteer holds Specificity to a good extent (see Capability retention below, comparison with SKOP line 308).

Question 1: Overlap audit

Short answer. 13 of 499 AxBench evaluation concepts (2.61%) appear exactly in the training pool, and roughly 10% have a near-paraphrase there. The other five benchmarks have no exact matches.

Setup. The reference set is the deduplicated union of the Concept16K v1 and v2 training splits (27,422 concepts). The evaluation concepts are AxBench (499), CLaS-Bench (32), PersonalityBench (793), SocialEval/WorldTree (7), and Dictator and Ultimatum (6 each). For each evaluation concept, we compute exact string match and its maximum similarity to any training concept, using word-level Jaccard (lexical) and cosine similarity (semantic) under sentence-transformers/all-mpnet-base-v2 embedding (identical strings are excluded).

Exact match.

BenchmarkConcepts (exact / n)
AxBench13 / 499 (2.61%)
CLaS-Bench0 / 32
PersonalityBench0 / 793
SocialEval (WorldTree)0 / 7
Dictator0 / 6
Ultimatum0 / 6

Maximum similarity to training (median / P95).

BenchmarkJaccardCosine
AxBench0.556 / 0.8590.825 / 0.966
CLaS-Bench0.200 / 0.2100.452 / 0.561
PersonalityBench0.121 / 0.1490.568 / 0.658
SocialEval (WorldTree)0.143 / 0.2500.680 / 0.717
Dictator0.129 / 0.1430.709 / 0.725
Ultimatum0.129 / 0.1430.709 / 0.725

Dictator and Ultimatum have only 6 concepts, so their rows are descriptive only.

Findings.

  • AxBench is the closest benchmark to the training concepts. Its P90 cosine is 0.939, so roughly 10% of AxBench concepts have a near-paraphrase in training.
  • The other five benchmarks show no meaningful concept overlap. No concept exceeds a Jaccard of 0.40, and the maximum cosine is at most 0.73.
  • Association with performance (AxBench). We computed Spearman's ρ between each concept's maximum cosine similarity to training (SC) and the concept score (C) and harmonic mean (HM). The associations are weak (|ρ| ≤ 0.17) and significant in only 3 of 12 cells (p < 0.05, uncorrected). Most are positive.
ModelSC & CSC & HM
Gemma-2-2B-IT0.0479 (0.2851)0.1002 (0.0252)
Gemma-2-9B-IT−0.0119 (0.7912)−0.0149 (0.7398)
Llama-3.1-8B-Instruct0.0514 (0.2515)0.0502 (0.2627)
Llama-3.2-3B-Instruct0.1073 (0.0165)0.1103 (0.0137)
Qwen3-1.7B0.1547 (0.2150)0.1678 (0.1781)
Qwen3-4B0.0906 (0.5682)0.1258 (0.4271)

Cells show ρ (p-value). Bold marks p < 0.05.

  • Performance on the excluded subset: To quantify the effect directly, we re-scored MetaSteer on the AxBench subset that excludes exact-match and near-paraphrase concepts and the 348 instructions shared with training (12,233 rows removed). Component scores for each backbone are given in our response to Reviewer Fhqe.
ModelFull HMExcluded HMΔConceptΔInstructionΔFluencyΔHarmonic
Gemma-2-2B1.511.50−0.04+0.03−0.03−0.01
Gemma-2-9B1.521.52+0.01+0.01−0.040.00
Llama-3.1-8B1.551.48−0.07−0.01−0.10−0.07
Llama-3.2-3B1.621.49−0.07+0.01−0.25−0.13
Qwen3-1.7B1.571.51−0.050.00−0.12−0.06
Qwen3-4B1.591.48−0.08+0.02−0.19−0.11

The harmonic mean changes by −0.13 to 0.00 (mean −0.06) and stays between 1.48 and 1.52. The concept and instruction scores change little (−0.08 to +0.03). The decrease comes mainly from fluency (−0.03 to −0.25), and the reason is not yet clear to us and needs more analysis. The Gemma backbones, trained on Concept16K-v2 (derived from LlamaScope, whereas the AxBench concepts come from GemmaScope, Table 9), are essentially unchanged. The other backbones, trained on Concept16K-v1 (GemmaScope-derived).

Conclusion: We cannot rule out a small effect of near-duplicate concepts on AxBench. In the next version, we will restrict the zero-shot claim for AxBench accordingly and report the excluded-subset results alongside the full results. The zero-shot claim for the other five benchmarks is unaffected.

Capability retention

The submitted version compared the unadapted model only with the adapted model given the steering template but no steering signal (adapted without concept). In a complementary experiment (asked by Reviewer z1nu), we measured how much a generic steering signal degrades general capabilities. The signal is "Be precise and accurate" (adapted with concept). The table shows results for the three configurations, with 0/2/4-shot results and the six backbones aggregated.

DatasetUnadaptedAdapted, no conceptAdapted, with conceptUnadapted vs. adapted, no conceptUnadapted vs. adapted, with concept
MMLU0.6480.6470.571−0.1 pts (p = 0.081)−7.8 pts (p = 0.003)
TruthfulQA MC10.3760.3790.355+0.3 pts (p = 0.531)−2.1 pts (p = 0.352)
TruthfulQA MC20.5510.5600.546+0.9 pts (p = 0.107)−0.5 pts (p = 0.898)

Without a concept, no change is detectable on any dataset. With the active concept, only MMLU shows a significant drop, and the TruthfulQA changes are not detectable. We will qualify the capability-retention claims in the paper accordingly.

References

[1] Zhengxuan Wu, Aryaman Arora, Atticus Geiger, Zheng Wang, Jing Huang, Dan Jurafsky, Christopher D. Manning, and Christopher Potts. AxBench: Steering LLMs? Even simple baselines outperform sparse autoencoders. arXiv preprint arXiv:2501.17148, 2025.

[2] Simon Ostermann, Daniil Gurgurov, Tanja Baeumel, Michael A. Hedderich, Sebastian Lapuschkin, Wojciech Samek, and Vera Schmitt. From weights to activations: Is steering the next frontier of adaptation? In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 29854–29879, 2026.

[3] Jiuding Sun, Sidharth Baskaran, Zhengxuan Wu, Michael Sklar, Christopher Potts, and Atticus Geiger. HyperSteer: Activation steering at scale with hypernetworks. arXiv preprint arXiv:2506.03292, 2025.

Reviewer_z1nuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

MetaSteer trains LoRA adapters on the query, key, value, and output attention projections using DPO over pooled concept-conditioned preference data. A single adapter per backbone remains fixed at inference, while the desired concept is supplied in the prompt, producing context-dependent activation changes. Experiments across six models evaluate transfer without further adaptation to three text-generation benchmarks and three agentic settings, reporting improvements in most aggregate comparisons. The paper additionally analyzes hidden-state trajectory geometry, general capability retention, evaluator reliability, and safety implications.

Critical strengths

Its strongest contribution is the breadth of this transfer evaluation: six backbones, three text-generation benchmarks, and three agentic settings. Aggregate results improve across all six models on AxBench and CLaS-Bench, and across five of six models on the agentic evaluations.

Critical weaknesses
  1. What produces the improvement remains unclear. Context-dependent, nonlinear activation effects already arise from applying LoRA within attention. The experiments do not isolate the benefits of adapting all four projections, using DPO, or pooling concept supervision. Matched-data ablations would clarify the contribution.

  2. The geometric analysis does not directly establish preservation. Appendix I compares ordered response trajectories with shuffled-order controls. It does not clearly compare matched trajectories before and after MetaSteer. Furthermore, low alignment with a CAA direction cannot exclude another fixed direction.

  3. The scope of unseen-concept transfer needs clarification. The internal split separates concept IDs but shares all 142 or 216 instructions across splits, which is explicitly disclosed. However, I could not find an explicit overlap audit between the Concept16K training corpus and the main Concept500 evaluation inventory.

  4. Capability retention during active steering is less established. The retention tests use ordinary multiple-choice evaluation without an active steering concept. Meanwhile, AxBench instruction-following scores decrease in five of six models, despite improved aggregate scores. This indicates a remaining tradeoff between concept compliance and task utility.

Questions for authors
  1. Beyond standard DPO training of attention LoRA, what is the distinctive contribution of MetaSteer? What evidence separates the benefit of adapting all four attention projections from that of pooled concept supervision?

  2. Were Appendix I’s responses generated by MetaSteer, and were their representations extracted with the adapters enabled? Which matched comparison with the unadapted model supports preservation of trajectory dynamics, given that Tables 7–8 compare original and shuffled segment order?

  3. How was overlap between Concept500 evaluation concepts and Concept16K training concepts checked and excluded? Appendix M.5 establishes disjointness within the training corpus’s splits, but the separation from the main AxBench evaluation inventory remains unclear.

6 comments

First, we thank Reviewer z1nu for taking the time to review our paper and for the valuable feedback. Below are our responses to each comment and question.

Comment 1 and Question 1: What produces the improvement

We agree that context dependence and nonlinearity (Eqs. 4–8) hold for any LoRA fine-tune of attention. We included these equations to show that MetaSteer is mathematically well-founded, not to claim uniqueness.

Matched-data ablations: The most relevant control is the same model with the same template and the adapter off (or coefficient = 0). The table shows the change from enabling the adapter (on minus off). Enabling the adapter raises the harmonic mean for all six backbones (+0.18 to +0.46) through the concept score (+0.28 to +0.89), at a cost in instruction score (−0.09 to −0.44). The gain is therefore not explained by the template alone.

ModelΔConceptΔInstructionΔFluencyΔHarmonic
Gemma-2-2B+0.62−0.44+0.32+0.32
Gemma-2-9B+0.64−0.34+0.06+0.30
Llama-3.1-8B+0.66−0.09+0.04+0.26
Llama-3.2-3B+0.89−0.17+0.17+0.46
Qwen3-1.7B+0.47−0.11+0.27+0.29
Qwen3-4B+0.28−0.13+0.25+0.18

For context, we also compare MetaSteer with technically adjacent methods as Fhqe suggested.

  • HyperSteer [1]: the closest transferable method. It extracts an attention pattern from a natural-language description of the concept (using an identical model) and applies it to attention heads for steering. It has three variants (concept only, concept + prompt, cross-attention). We report the best-performing variant here.
  • RePS: a preference-trained steering vector added to the activations, trained per concept.
  • LoRA, LoReFT, SFT: standard fine-tuning methods, trained on each single concept and evaluated on unseen instructions. (concepts remains same)
MethodGemma-2-2BGemma-2-9B
Prompt, unadapted (same template, adapter off)1.191.22
RePS0.7560.892
HyperSteer0.7421.09
LoReFT0.7220.777
SFT0.714–
LoRA0.6410.602
MetaSteer*1.511.52

Scores are the AxBench harmonic mean. MetaSteer partially overlaps with its training pool on AxBench (see the overlap audit section).

All trained baselines score below the unadapted prompt, consistent with Table 1 in the paper and with the AxBench paper [3], where prompting is the strongest reference method on AxBench. We therefore treat the prompt as the primary reference and the trained baselines as secondary context.

Training all four attention projections, not just one: We ran this ablation on Gemma-2-2B-IT by adapting each of q, k, v, and o alone under the same setup as the full adapter. The results are in the Head Isolation section below. In the next revision, we will extend it to other backbones and to combinations of projections.

Comment 2 and Question 2: Geometric analysis

In the submitted Appendix I, the analyzed responses were dataset responses embedded by the unchanged model. During the rebuttal, we repeated the analysis for all four combinations: (dataset or MetaSteer-generated responses) × (unchanged model or MetaSteer embedder). The finding remained persistent in all four settings (see the table below).

EmbedderDataAccumulationPositionVelocityAccelerationMenger
UnadaptedAxBenchOff+0.0536+0.2877+0.2668+0.2712
UnadaptedAxBenchOn+0.0114+0.2474+0.2449+0.0363
UnadaptedGeneratedOff+0.0420+0.1877+0.1636+0.1923
UnadaptedGeneratedOn+0.0059+0.1758+0.1599+0.1233
AdaptedAxBenchOff+0.0538+0.2876+0.2667+0.2715
AdaptedAxBenchOn+0.0111+0.2472+0.2449+0.0329
AdaptedGeneratedOff+0.0425+0.1878+0.1635+0.1956
AdaptedGeneratedOn+0.0072+0.1769+0.1613+0.1235

Similarity gain of ordered segments over the shuffled control, aggregated over six backbones.

The results show that steered responses, which are generated to answer different instructions while expressing various concepts, have meaningful structural similarity (including velocity, acceleration, and curvature), although position varies much like a random order (no meaningful gain in similarity over the shuffled control). This reinforces the notion that steering, as a phenomenon, changes position while preserving structure. It is worth noting that the embedder model is completely blind to instruction and concept.

CAA alignment: We agree that low alignment with a CAA direction cannot exclude another fixed direction. As stated in Section 3.3, the score is a diagnostic against one linear baseline.

Comment 3 and Question 3: Scope of unseen-concept transfer

The submitted version did not audit overlap between the Concept16K training pool and the evaluation benchmarks. We have now done so for all six evaluation sets and re-scored MetaSteer on an AxBench subset that excludes overlapping items. 13 of 499 AxBench concepts (2.61%) and 348 of 2,954 AxBench instructions (11.78%) match training exactly, and roughly 10% of AxBench concepts have a near-paraphrase in training. The other five benchmarks show no exact matches.

Setup. The reference is the deduplicated union of the Concept16K v1 and v2 training splits (352 instructions, 27,422 concepts, 226,696 pairs). The evaluation sets are AxBench (499 concepts in the audited file, 2,954 instructions), CLaS-Bench (32 concepts, 2,240 instructions), PersonalityBench (793 concepts, 459 instructions), SocialEval/WorldTree (7 concepts, 1,015 instructions), and Dictator and Ultimatum (6 concepts and 1 instruction each). For each evaluation item, we compute exact string match and its maximum similarity to any training item of the same type, using word-level Jaccard (lexical) and cosine similarity under sentence-transformers/all-mpnet-base-v2 (semantic; identical strings are excluded).

Exact match.

BenchmarkConcepts (exact / n)Instructions (exact / n)
AxBench13 / 499 (2.61%)348 / 2,954 (11.78%)
CLaS-Bench0 / 320 / 2,240
PersonalityBench0 / 7930 / 459
SocialEval (WorldTree)0 / 70 / 1,015
Dictator0 / 60 / 1
Ultimatum0 / 60 / 1

Maximum similarity to training (median / P95).

BenchmarkConcept JaccardConcept cosineInstruction JaccardInstruction cosine
AxBench0.556 / 0.8590.825 / 0.9660.286 / 1.0000.549 / 0.787
CLaS-Bench0.200 / 0.2100.452 / 0.5610.000 / 0.1070.260 / 0.367
PersonalityBench0.121 / 0.1490.568 / 0.6580.110 / 0.1350.317 / 0.376
SocialEval (WorldTree)0.143 / 0.2500.680 / 0.7170.061 / 0.0830.327 / 0.385
Dictator0.129 / 0.1430.709 / 0.7250.157 / 0.1570.469 / 0.469
Ultimatum0.129 / 0.1430.709 / 0.7250.161 / 0.1610.491 / 0.491

Dictator and Ultimatum have only one instruction and six concepts, so their rows are descriptive only.

Findings.

  • AxBench is the closest benchmark to the training data. The instruction overlap reflects a shared instruction pool (App. M.5): 348 of the 352 training instructions appear in AxBench, while 2,606 of 2,954 AxBench instructions (88.2%) do not appear in training. The concept P90 cosine is 0.939, so roughly 10% of AxBench concepts have a near-paraphrase in training.
  • The other five benchmarks show no meaningful overlap. No item exceeds a Jaccard of 0.40 (concepts) or 0.33 (instructions), and the maximum cosine is at most 0.73 (concepts) and 0.54 (instructions). Values of roughly 0.5–0.7 between short concept descriptions reflect topical and stylistic similarity.
  • Association with performance (AxBench). For ~10% of the generated evaluation instances per model (n = 3,000), we computed Spearman's ρ between the maximum cosine similarity to training and the instance scores. Instruction similarity is modestly associated with the instruction score for all six backbones (ρ = 0.23–0.25) and weakly with the harmonic mean (ρ = −0.005 to 0.18, significant for four of six). Concept similarity has weak associations throughout (|ρ| ≤ 0.17, significant in 3 of 12 cells).
ModelSI & ISC & CSI & HMSC & HM
Gemma-2-2B-IT0.2274 (<0.001)0.0479 (0.2851)0.1846 (<0.001)0.1002 (0.0252)
Gemma-2-9B-IT0.2301 (<0.001)−0.0119 (0.7912)0.1324 (<0.001)−0.0149 (0.7398)
Llama-3.1-8B-Instruct0.2279 (<0.001)0.0514 (0.2515)−0.0049 (0.8429)0.0502 (0.2627)
Llama-3.2-3B-Instruct0.2519 (<0.001)0.1073 (0.0165)0.0400 (0.1061)0.1103 (0.0137)
Qwen3-1.7B0.2493 (<0.001)0.1547 (0.2150)0.0858 (0.0005)0.1678 (0.1781)
Qwen3-4B0.2367 (<0.001)0.0906 (0.5682)0.1019 (<0.001)0.1258 (0.4271)

SC and SI: maximum cosine similarity of the concept and the instruction to training. C, I and HM: concept score, instruction score and harmonic mean. Cells show ρ (p-value). Bold marks p < 0.05.

continuation of findings in next comment

  • Performance on the excluded subset. To quantify the effect directly, we re-scored MetaSteer on the AxBench subset that excludes exact-match and near-paraphrase concepts and the 348 shared instructions (12,233 rows removed).
ModelFull HMExcluded HMΔConceptΔInstructionΔFluencyΔHarmonic
Gemma-2-2B1.511.50−0.04+0.03−0.03−0.01
Gemma-2-9B1.521.52+0.01+0.01−0.040.00
Llama-3.1-8B1.551.48−0.07−0.01−0.10−0.07
Llama-3.2-3B1.621.49−0.07+0.01−0.25−0.13
Qwen3-1.7B1.571.51−0.050.00−0.12−0.06
Qwen3-4B1.591.48−0.08+0.02−0.19−0.11

The harmonic mean changes by −0.13 to 0.00 and stays between 1.48 and 1.52. The concept and instruction scores change little (−0.08 to +0.03), and the decrease comes mainly from fluency (−0.03 to −0.25). The Gemma backbones, trained on Concept16K-v2 (derived from LlamaScope, whereas the AxBench concepts come from GemmaScope; see Table 9), are essentially unchanged. The other backbones, trained on Concept16K-v1 (GemmaScope-derived), show larger decreases. The reason for the fluency decrease is not yet clear to us and needs more analysis.

Conclusion: We cannot rule out a small contribution from overlap on AxBench, particularly through shared instructions. In the next version, we will restrict the zero-shot claim for AxBench and report the excluded-subset results alongside the full results. The zero-shot claim for the other five benchmarks is unaffected.

Comment 4: Capability retention during active steering

We agree that the initial version lacks a capability-retention test during active model steering. We ran this experiment during the rebuttal window and see a clear drop in performance. In essence, this is a particular form of the tradeoff between instruction-following capability and concept score, which is established in Comment 1 and in the paper's multi-objective optimization framing.

The following table contains capability retention on three datasets for both an empty steering signal and a general steering signal ("Be precise and accurate"). The table shows results for the three configurations, with 0/2/4-shot results and the six backbones aggregated.

DatasetUnadaptedAdapted, no conceptAdapted, with conceptUnadapted vs. adapted, no conceptUnadapted vs. adapted, with concept
MMLU0.6480.6470.571−0.1 pts (p = 0.081)−7.8 pts (p = 0.003)
TruthfulQA MC10.3760.3790.355+0.3 pts (p = 0.531)−2.1 pts (p = 0.352)
TruthfulQA MC20.5510.5600.546+0.9 pts (p = 0.107)−0.5 pts (p = 0.898)

Without a concept, no change is detectable on any dataset. With the active concept, only MMLU (14K samples) shows a significant drop, and the TruthfulQA changes are not detectable at our sample size (817 samples each). We will revise the statements that the model's general capabilities are largely preserved (Introduction and Section 7) accordingly.

References

[1] Jiuding Sun, Sidharth Baskaran, Zhengxuan Wu, Michael Sklar, Christopher Potts, and Atticus Geiger. HyperSteer: Activation steering at scale with hypernetworks. arXiv preprint arXiv:2506.03292, 2025.

[2] Haoyan Luo, Mateo Espinosa Zarlenga, and Mateja Jamnil. Don't lose focus: Activation steering via key-orthogonal projections. arXiv preprint arXiv:2605.06342, 2026.

[3] Zhengxuan Wu, Aryaman Arora, Atticus Geiger, Zheng Wang, Jing Huang, Dan Jurafsky, Christopher D. Manning, and Christopher Potts. AxBench: Steering LLMs? Even simple baselines outperform sparse autoencoders. arXiv preprint arXiv:2501.17148, 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.