acceptodds
Desk rejected at ICLR 2027

What Does Steering Change? Decoding with Base-Relative Steering Evidence

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Activation steering controls a frozen language model by adding a direction to its hidden states, but the outcome hinges on the steering strength. Weak steering may leave the output largely unchanged, while stronger steering can degrade content. We study this sensitivity by examining how steering changes next-token predictions relative to the unsteered model. At a fixed strength, steering-induced divergence varies widely within generations, and steering can promote alternative tokens before changing the most probable continuation. Crucially, base-relative gain predicts which alternatives become dominant at stronger interventions beyond what steered probability alone predicts. These observations motivate decoding with base-relative steering evidence. We instantiate this approach in Multi-Alpha Consensus Decoding (MAC), which evaluates several strengths, filters unreliable predictions, and scores tokens by base-relative gain anchored to base likelihood. In controlled comparisons, base-relative gain scoring can produce stronger style transfer than mixing steered probabilities. Across two authorship-transfer benchmarks and four model backbones, MAC achieves the highest target-style similarity among the compared methods, with a measurable cost in meaning preservation. Our results suggest that steering-strength sensitivity is partly a decoding problem that base-relative steering evidence helps address.

Primary area:
foundation or frontier models, including LLMs
Keywords:
activation steering, inference-time intervention, controllable generation, adaptive decoding, adaptive steering

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_MKLJOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

setup:

  • inference-time activation steering setting
  • heuristic technique based pre-computed contrastive vector is already available for a target task (the paper focuses on instruction following constraints and writing style guide)
  • vector gets added to the hidden layers of an LM at inference time

problem:

naive addition behaves unevenly from token to token and its not trivial to calibrate it -- two extremes cases are it's hard to get rid of high confident base model tokens and on the other hand you risk running into repetitive loops and other such degenerate patterns.

proposed solution:

  • the authors notice that when steering does not change the top token, it often boosts some other tokens by 10x+ relative to the base model, and they try to use this gain in their proposed decoding technique
  • as no single "strength" for adding the steering vector works at every position, their method (MAC) runs the base model along with 5 different steering strengths in parallel at each step. it then scores tokens by combining the base log-probability with a weighted average of their gains across the strengths.
  • to keep this log-gain score from picking noisy tokens or running into degeneracy, they use several heuristics on top.
  • empirically, MAC gets noticeably stronger style transfer than prior baselines (while trading off some meaning preservation), whereas on instruction following its gains are only significant in the setting where the instruction is removed from the prompt.
Critical strengths
  1. practical set of tricks that make log-gain usable without collapsing into degeneracy. using raw log-prob gain relative to the base model is brittle (because of tail noise and flattened distributions easily blowing up) and the authors propose a nice set of guards that actually make it work in practice. the tricks intuitively make sense and the authors explain / justify with simple token-level observations:
  • drop strengths whose distribution becomes too flat or low confidence
  • if all remaining strengths still agree on the same top-1 token (from the base model's top 3), output it directly without running the gain score
  • restrict scoring to only the top-5 tokens from each branch
  • clip negative log-gains to zero and subtract a repetition penalty over the last few tokens
  1. solid empirical evaluation and ablations within their chosen setup: across 4 models and 2 style-transfer datasets MAC consistently gets stronger style transfer than previous relevant baselines, and the ablations are detailed enough to see what each of the bigger pieces contribute.
Critical weaknesses
  1. weak motivation: the authors have a really nice observation (steering can boost runner-up tokens by 10x+ before having an effect on top tokens) but adding decoding heuristics on top of it in a contrived setting feels a bit shallow. the results with the IF setup somewhat hint at this where when the instruction is in the prompt, MAC has no gains. there could be better uses for this observation; maybe as an eval or scaling-law signal to see how models across sizes already surface the right tokens internally before they reach rank 1 or something like that. the paper makes no effort to justify otherwise.

  2. i would have liked to see the failure modes (for which several guards were added) folded into the scoring formulation more elegantly. from the ablations it looks like the guards do quite a lot of heavy lifting.

Questions for authors
  1. can you evaluate the main baselines with similar guards added as MAC (the usability filter, repetition penalty etc.)?

  2. is there a more realistic task beyond style transfer where MAC shows a clear gain when the full prompt is actually given? the IF eval is not convincing.

Reviewer_WNEKOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper asks what activation steering actually changes at the next-token level and uses the answer to design a decoder. Part 1 is an analysis across four backbones (Llama-3.2-3B, Llama-3.1-8B, Qwen2.5-3B, Phi-4-mini) on 85,500 positions from MUD style-transfer prefixes, comparing steered pt,α against unsteered pt,0 over α ∈ {0.1,…,1.0}. Findings: (RQ1) at fixed α, within-generation KL divergence varies by ≥10× (p90/p10); (RQ2) the "flip strength" that changes the argmax varies across positions (72–84% of adjacent positions differ); (RQ3) steering frequently promotes an alternative ≥10× while leaving the argmax unchanged (16–44% of positions at α = 0.3); (RQ4) base-relative gain g = pt,α/pt,0 predicts which alternatives later reach rank-1 beyond current steered probability — adding log-gain raises AUC from 0.736–0.783 to 0.790–0.838 across all 36 backbone×strength cells, with coarse probability-stratum controls.

Part 2 introduces Multi-Alpha Consensus Decoding (MAC), a training-free decoder that evaluates the unsteered branch plus a 5-value strength grid in parallel, filters unusable strengths (entropy/peak/rank conditions), weights strengths by smoothed base-relative evidence under a Gaussian prior centered at α=0.2, and scores tokens by St(x) = log pt,0(x) + λ Σ qt(α)ut,α(x) − γ ct(x) with a direct-consensus fast path. Evaluation: authorship transfer (MUD, LaMP Tweet) and IFEval, vs unsteered/fixed-strength/Hybrid Search/DAC/MoLaCE/DIRECTER. MAC attains the highest StyleCAV in all settings (e.g. MUD 0.370/0.414 on 8B/3B vs MoLaCE 0.226/0.346), GPT-4.1 pairwise win rates 67–89% (Holm p<0.001), at a measurable SBERT (meaning) cost. On IFEval it gives the highest compliance (92.0% with instructions, +3.6pp but not significant; 46.1% without, +14.1pp significant) with a quality cost when instructions are absent.

Critical strengths
  1. A predictive, actionable analysis. The finding that base-relative gain predicts later argmax transitions beyond current steered probability (RQ4, AUC +0.054–0.073 at α=0.3, improvement in every cell, with probability-stratified controls showing 50–56% vs 12–15% takeover at matched pt,α) is the strongest part of the paper. It is not merely descriptive: it directly justifies the decoder's scoring signal, so the analysis and method are tightly coupled rather than bolted together.
  2. Honest, well-controlled evaluation of the resulting trade-off. MAC is shown to sit on the style–meaning Pareto frontier with statistically significant style gains and transparent reporting of the meaning/quality cost (SBERT down; IFEval quality 3.88 vs 4.20 without instructions). The ablations disentangle weighting from token objective (Figure 6 interaction, +0.238 StyleCAV) and show base-relative decoding helps even with a single strength (Figure 7), which substantiates that the mechanism — not just ensembling — drives the result.
Critical weaknesses
  1. The "best" claim is partly an operating-point choice, on metrics that reward the intervention. MAC's headline advantage is "highest target style," but MoLaCE consistently retains higher SBERT (meaning), i.e. it occupies a different point on the same trade-off; MAC's win is largely a move toward style at a measured cost in meaning. Because the primary style metric (StyleCAV) and the GPT-4.1 "which shares the author's style" prompt both reward stronger stylization while the judge is not shown the source (Appendix D.2), the evaluation is structurally biased toward the more aggressive method. A stronger case would fix a meaning floor and compare style, or report a single scalar that penalizes meaning loss, across all baselines — not only in the Appendix C controls.
  2. Modest/uncertain gains on the verifiable task, and cost. On IFEval-with-instructions, the compliance gain (+3.6pp) is not significant after Holm correction and quality is essentially tied; the clear win (without instructions, +14.1pp) comes with a significant text-quality drop (3.88 vs 4.20). So on the one benchmark with objectively verifiable constraints, MAC either does not significantly help or helps at a real quality cost. Combined with the 1.42× decoding overhead and the need to still select a strength grid, the practical value beyond style transfer is not firmly established.
Questions for authors
  1. For style transfer, can you report results at a matched meaning floor (e.g. select each method's operating point to SBERT ≥ a fixed threshold, then compare StyleCAV)? Does MAC still dominate MoLaCE when meaning is held comparable?
  2. The GPT-4.1 judge does not see the source text, so it cannot penalize meaning drift. Can you add a source-aware pairwise protocol (style and fidelity) to check that MAC's preference is not an artifact of rewarding over-stylization?
  3. StyleCAV/StyleDistance are embedding metrics that may be gamed by surface features. Do human judgments (even a small study) agree with the StyleCAV ranking?
  4. The strength grid is the only tuned setting but is selected on a development set per backbone. How sensitive are the results to grid choice (Appendix J), and would a single fixed grid across all backbones materially change conclusions?
  5. On IFEval without instructions, can the quality drop (3.88) be mitigated by a stronger base-likelihood weight λ or a meaning/quality term, and if so does compliance survive?
  6. Please position MAC against concurrent authorship-steering work, particularly Can Activation Steering Capture Multidimensional Authorship Style? (arXiv:2609.04792, whose A3S construction you already use) and Predicting Steering Vectors and Adapter Weights for Few-Shot Author-Style Transfer (arXiv:2610.03163), and against recent adaptive-scale methods (e.g. Dynamically Scaled Activation Steering, arXiv:2512.03661). Is the base-relative decoding axis orthogonal to their strength-scaling, and do they stack?
Reviewer_Ak5YOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The limitations are discussed on page 9. Please assess whether the paper should be desk-rejected directly.

The following scores are for reference only.

Critical strengths

The limitations are discussed on page 9. Please assess whether the paper should be desk-rejected directly.

The following scores are for reference only.

Critical weaknesses

The limitations are discussed on page 9. Please assess whether the paper should be desk-rejected directly.

The following scores are for reference only.

Questions for authors

The limitations are discussed on page 9. Please assess whether the paper should be desk-rejected directly.

The following scores are for reference only.

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

Related papers

Discussion (0)

Sign in to comment.