acceptodds
Under review as a conference paper at ICLR 2027

Robust Steerability: Closed-Loop Control of Language Model Representations Under Distribution Shift

23 views

Abstract

Activation steering can change the behavior of large language models, but an intervention calibrated on one prompt distribution may fail when deployment inputs alter the network's internal dynamics. Neurostimulation faces an analogous problem and has increasingly moved from fixed, open-loop stimulation toward closed-loop, state-dependent control. Following this principle, we formulate activation steering as finite-horizon feedback control over transformer depth. We identify local layer-to-layer dynamics from calibration activations and represent their mismatch with the true network through a disturbance channel fitted to the empirical covariance of one-step prediction residuals. We then synthesize an controller that minimizes the worst-case amplification of this mismatch, and define robust steerability, , as the inverse of the smallest achievable amplification. We show that the associated certificate localizes fragility to specific layers and disturbance directions, and that nominal LQR steering is the large- H_\infty steering reduces mean jailbreak attack success from 9.3\% to 1.7\% relative to the strongest baseline, improves truthfulness\(\times\)informativeness under Spanish-prompt transfer by 5.4 points on average, and retains 77--95\% of unsteered math accuracy under language steering, compared with 0--85\% for A-LQR. Gains under context-length shift are model-dependent. Finally, S_rob $ varies systematically with the steering task and increases with model scale within a task. These results suggest that modeling uncertainty matters most when steering is applied outside its calibration regime, and that biological and artificial neural networks pose related control problems.

Primary area:
foundation or frontier models, including LLMs
Keywords:
activation steering, large language models, robust control, distribution shift, representation engineering, interpretability, closed-loop neurostimulation, NeuroAI

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_xQ4UOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper presents a technique to evaluate whether an activation steering method will be reliable when applied out of the regime where it has been developed. It further proposes a more robust steering technique that can resist deployment shifts.

The evaluation technique is inspired by closed loop methods in neurostimulation, and based on modeling activations through the layers of a transformers as a discrete time dynamical system. Specifically, how much the dynamics deviates from a calibration dynamics, more precisely how this deviation is amplified through the layers, will dictate if the steering approach is reliable.

The authors then define robust steerability, and show that it can additionally identify vulnerable directions in activation space.

The experiments validate the methods and theory across four deployments shift for language tasks.

Critical strengths

This paper provides a refinement on a notion of steerability, moving away from a binary assessment of whether a network can be steerable, but rather to which extent (in terms of distributional deployment shift) it can be steered.

The experiments in four different settings are convincing.

Critical weaknesses

I was excited by the analogy with closed loop stimulation experiments in human brains at first, but ultimately didn’t find the analogy to hold since physical time and layer are conceptually different. The closed loop part is interesting, but I’m not sure the parallel between brains and machines brings something. To be precise, I know that depth-as-time is a classic way of viewing propagation through a transformer. However, depth already has an analogy in biological brains and time already has an analogy in neural networks (tokens). Hence, I feel that the narrative mixes the two.

Figure 3 is too dense and the caption not detailed enough.

Questions for authors

I'm confused about Figure 7, panel B: could the authors explain in more details what it shows, and in particular, why is time now tokens?

3 comments

We thank the reviewer for recognizing the refined notion of steerability and the convincing experiments. We did not expect this level of neuro depth in a review: we optimized the manuscript for an ML-venue audience and the LLM subsections, and our legends' brevity reflects the neuro detail we judged suitable there. Encouraged by this solid NeuroAI review, we are expanding the brain-analogy sections.


W1: Layer vs. physical time

We agree with the reviewer: depth has an analogue in the brain (space) and time has an analogue in neural networks (tokens). This is exactly the mapping we intend, and the narrative should not have mixed the two. Part of the confusion came from our notation: the index in the dynamical equation (line 133) and Fig. 3C was written as temporal but should be spatial. This is now corrected. We take layers as the variable of the dynamics and write a dynamical equation for the progression of activity across layers within a fixed time window (here in token). We apply the exact same layer-to-space analogy to the brain: we define a hierarchy over cortical regions (Fig3.D), index them by k, and study the propagation of activity across them for a fixed window of time (Δ\Delta). In both systems the dynamics propagate across space, and time is held fixed (physical time in the brain, tokens in LLMs).

We understand the reviewer drew a common analogy from the neurostimulation literature, which mostly models dynamics through time. However, state-dependent closed-loop stimulation is not limited to temporal dynamics. Neural states are inherently spatiotemporal, and their spatial organization can be functionally consequential for stimulation. Manipulating relative phase across cortical regions can enhance or impair cognition [1]. Direction-specific traveling-wave stimulation induces spatially organized responses and distinct cognitive outcomes [2], and TMS-EEG evidence shows that the response to stimulation depends on the ongoing global spatial configuration of activity [3,4]. Effective stimulation thus depends not only on when an intervention is delivered, but also on the spatial organization of the state it acts on. This is exactly analogous to state-based stimulation that uses a "spatial dynamical" equation to estimate the state; older procedures mostly estimate it from "temporal dynamics" at a single recording site.

We borrow only one concept from this literature: state-based closed-loop stimulation is more effective than open-loop. The parallel adds an empirical result: the same spatial-dynamics formalism shows local dynamics in both brains and LLMs (Fig. 7B, Q1). We made the space/time distinction explicit in the introduction and Fig. 3 and Fig.7 caption.

Background for brain–transformer analogy: Large-scale cortical recordings have established traveling waves as a prominent organizing principle of neural spatiotemporal dynamics. Building on this, recent works by Muller, Churchland, Sejnowski et al. has identified potential computational parallels between cortical traveling waves and Transformers, particularly in the spatial encoding of temporal context and attention [5,6]. Beyond superficial brain–LLM analogies, we argue spatial organization of neural dynamics provides a more mechanistically grounded framework than equating transformer depth with biological time. Our work contributes to this framework; a fuller treatment is underway in a separate manuscript.


W2: Figure 3 is too dense and the caption too brief

We expanded the Figure 3 caption to clarify each panel, using the additional camera-ready page.


Q1: Fig. 7B and "time as tokens"

Once the dynamics are written across layers/electrodes (space), the remaining question is how time is frozen. We take time to be the context length: tokens in LLMs, and the temporal window fed to the spatial dynamical equation in the brain. Panel B shows that in both LLMs and brain regions, the residual in predicting the future spatial state increases with this horizon: the longer the context, the harder it is for the linearized model to explain the layer-wise (or region-wise) dynamics. This means the spatial dynamics is local in both biological brains and LLMs. The time/token horizon is the variable on which we test the spatial dynamics, not the variable along which they propagate. We extended the legend to convey this.

We hope these explanations and improvements address the reviewer’s concerns and are reflected in the final evaluation.


References

1. https://elifesciences.org/articles/22001
2. https://www.pnas.org/doi/10.1073/pnas.2527296123
3. https://www.nature.com/articles/s41467-019-10581-7
4. https://www.jneurosci.org/content/44/39/e0782242024
5. https://www.cell.com/trends/neurosciences/fulltext/S0166-2236(24)00149-8
6. https://arxiv.org/abs/2511.11668
Reviewer_xQ4UOpenReview ↗

Thank you for clarifying my concerns

I thank the authors for clarifying my misunderstanding/concerns on the link between neuroscience and AI, and on the figures. I will keep my score.

Thank you for your support and constructive feedback, which helped clarify our NeuroAI contribution. We are pleased that your concerns have been resolved and would be grateful if you would consider updating your rating to reflect your current assessment of the paper.

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

The paper formulates activation steering as finite-horizon H∞ feedback control, using calibration activations to estimate local dynamics and structured model uncertainty. It introduces robust steerability as a diagnostic and evaluates frozen controllers under four distribution shifts, with the strongest results on TruthfulQA and HarmBench.

Critical strengths
  1. The uncertainty-aware formulation addresses a relevant limitation of nominal steering: distribution shifts can alter how interventions propagate through representations.
  2. Consistent improvements on TruthfulQA and HarmBench across several models provide promising evidence that robust feedback can improve steering beyond the calibration distribution.
Critical weaknesses
  1. The comparison does not isolate robust synthesis from hyperparameter optimization: H∞ receives benchmark-specific tuning, whereas A-LQR uses inherited fixed configurations.
  2. The predictive value of robust steerability remains insufficiently established. Its dependence on separately selected performance weights and limited behavioral validation complicates its interpretation as a comparable reliability measure.
Questions for authors
  1. Does H∞ retain its advantage when A-LQR receives the same calibration split, cost-weight search space, and selection budget?
  2. How well does calibration residual covariance capture OOD error geometry, and do gains persist against scale-matched isotropic, diagonal, and randomly rotated alternatives?
  3. Under fixed performance weights and consistent normalization, does SrobS_{rob} predict held-out degradation better than residual variance, nominal closed-loop disturbance gain, and perturbation sensitivity?
3 comments

We thank the reviewer for their thoughtful and encouraging assessment, particularly for recognizing our contributions to uncertainty-aware activation steering and the consistent performance improvements demonstrated across multiple benchmarks under distribution shift. We particularly appreciate the reviewer’s recognition of the work’s informativeness and technical rigor. Moreover, the reviewer’s insightful questions motivated additional control experiments that substantially strengthened (1) the behavioral validation of our robust steerability measure and (2) the evidence supporting a fair comparison between our proposed controller and nominal LQR.


Q1 / W1 (Does H∞ retain its advantage under equal calibration?)

Yes. First, we clarify that A-LQR in our original experiments followed the calibration pipeline of Skifstad et al.; where the model was identical, we inherited their reported settings for a faithful published-baseline comparison.

We additionally performed the requested matched-budget study on Qwen2.5-0.5B-Instruct / HarmBench. LQR (8D) and H∞ use the same 8D basis, calibration split, dynamics, control channels, reference, intervention, 12-setting cost grid, 50 validation prompts per setting, and AXBench Overall selection rule.

Table R1a. Complete calibration grid (AXBench Overall ↑ ; bold = maximum).

Q/RQf/RA-LQR (full)LQR (8D)H∞
.01.01.364.494.772
.01.1.734.420.744
.011/√10.694.024.044
.1.01.734.001.001

Both 8D controllers independently select the same costs (Q/R, Qf/R) = (.01, .01), yet H∞ achieves .772 vs. .494 on calibration. We froze each controller at its best calibration and tested 240 held-out behaviors, each with five jailbreak variants:

Table R1b. Held-out HarmBench ASR (%) ↓ ; mean ± jackknife SE.

ControllerDirectJailbreak (OOD)
Original32.50 ± 2.3124.42 ± 1.59
A-LQR (full)29.58 ± 2.8116.17 ± 1.61
LQR (8D)17.08 ± 2.1911.92 ± 1.17
H∞12.92 ± 2.365.33 ± 0.83

H∞ reduces OOD ASR by 55.3% relative to equally calibrated A-LQR, with identical selected weights. Thus, the improvement persists without an extra search budget or more favorable representation.


Q2 (Does calibration covariance capture OOD geometry? What about isotropic, diagonal, or rotated controls?)

We took the reviewer’s question seriously and added a new appendix section (I) comparing our method against isotropic and randomly rotated variants of the steering policy. On HarmBench, the rotated H∞ variant exhibited a 1.25-fold degradation in the joint steering performance metric, while the isotropic variant achieved near-zero performance. Both variants were scale-matched to our proposed method. Together, these results demonstrate the sensitivity of robust steering performance to disturbance geometry.

Moreover, we interpret the empirical residual covariance as defining the disturbance channel, without requiring it to match the OOD covariance. The matched-budget experiment holds the fitted system and evaluation protocol fixed while varying only the synthesis method (nominal LQR versus H∞), establishing an advantage for robust synthesis under distribution shift. We thank the reviewer for motivating this clarification and believe that disentangling robust optimization from the choice of disturbance geometry opens a promising direction for mechanistic investigation.


Q3 / W2 (Is SrobS_{rob}) useful despite calibration dependence?)

In light of this question and other validity related question about SrobS_{rob} we now incorporated a new section in our main results section, fully exploring the robust steerability and its behavioral gain and calibration control analysis. Please see page 10.

Above these empirical new diagnostics and control analysis we particularly argue that SrobS_{rob} provides pre-deployment capabilities beyond residual variance. Figure R1e is a nominal example of applicability of SrobS_{rob} to track steering gains cross task which is fundamentally unreachable from residual variance analysis.


We believe the new sections strengthen the diagnostic's empirical grounding and we hope consequently it resolves any remaining attribution concern. Given the reviewer's positive assessment of soundness and significance, we would welcome favorable reassessment of the recommendation in light of explanations and additional experiments provided for weakness 1 and 2.

Thank you for the response and additional experiments. The connection between neurostimulation and LLM steering is particularly interesting, and I appreciate the authors’ effort to bring control-theoretic ideas into this area. The clarifications and new results have addressed my main concerns, and I am happy to recommend acceptance of this paper.

Thank you for recommending acceptance; your questions directly improved the paper. We would be grateful if you could update the rating to reflect your recommendation.

Reviewer_UfC4OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the reliability of activation steering under distribution shift. Fixed or closed-loop activation interventions typically rely on local dynamical models estimated on calibration data; such models become mismatched when the language, prompt form, or context length changes, and the error is amplified across subsequent layers. The main contributions are threefold. First, robust control is formally introduced into activation steering, explicitly incorporating "possible model mismatch" into controller synthesis rather than correcting it post hoc. Second, interpretable diagnostic quantities (Srob and the worst-case perturbation direction) are provided for comparing the local fragility of different model–task instances. Third, validation is conducted on four benchmarks, with a different shift variable designed for each benchmark.

Critical strengths
  1. It writes "possible model mismatch" into the controller design itself. Rather than constructing yet another fixed intervention vector, the authors formulate activation steering as finite-horizon feedback control: layer depth serves as the time horizon, and the covariance of calibration residuals serves as the disturbance channel. This incidentally yields diagnostic quantities that can localize fragile layers and directions. Its value lies in the framework rather than in any particular controller, and it is worth the community continuing to study.

  2. There are consistent positive gains across four independently constructed shift scenarios. On TruthfulQA, all four models improve in T×I; on HarmBench, the attack success rate decreases across three models and six conditions; on MGSM, Qwen outperforms A-LQR in both accuracy and Spanish correlation under Chinese/French/Japanese.

Critical weaknesses
  1. Srob is interpreted more broadly than the evidence can support. It is an error-amplification metric for an eight-dimensional surrogate system under a specific cost and perturbation measure—not accuracy, not safety rate, and not a behavioral failure threshold. Each model–task instance also selects its own cost, and uniformly rescaling the cost can change the score without changing controller behavior. The appendix explores the association between the score and refusal using only three points, two of which are already near the ceiling.

After all, the paper positions Srob as a "pre-deployment reliability diagnostic," and this is precisely the core new contribution that distinguishes it from existing steering methods. If this layer does not hold up, the contribution degenerates into "a controller with more thorough tuning." 2. The existing comparisons cannot attribute the gains to robust synthesis, and some "capability preservation" cases did not actually steer successfully. H∞ performs a separate tuning grid search for each benchmark, whereas A-LQR uses a fixed configuration, and the appendix's descriptions of the two implementations' dimensions are inconsistent—so it is impossible to disentangle how much of "H∞ is better" comes from robustness versus tuning. Meanwhile, on MGSM, Granite retains about 95% mathematical accuracy but has a Spanish correlation of 0; on HarmBench 1B, H∞'s overall steering quality and fluency are actually lower than A-LQR's. The former undermines causal attribution, and the latter indicates that some "preserved capability" may simply reflect that the intervention did not take effect. Together, these two points currently prevent me from raising my score to accept. If an equal-budget LQR comparison and a capability comparison after matching steering achievement rates are added, I will switch to weak accept.

Questions for authors

(1)Under the same eight-dimensional basis, the same calibration split, the same cost grid, and the same selection budget, how much advantage does H∞ still have over a recalibrated LQR? If an additional isotropic-perturbation H∞ control is added that matches the overall scale, can the respective contributions of the "worst-case objective" and the "empirical covariance geometry" be separated? (2)Can you report a per-sample joint metric of "target language/style achieved AND task success," or compare accuracy and benign-request utility after matching steering achievement rates? Currently Granite retains about 95% accuracy on MGSM but has a Spanish correlation of 0, so it is impossible to judge from this whether the steering actually took effect. (3)Although I think most of the figures in this paper are quite sophisticated and well done, there are still some flaws I care about. For example, some icons in Figure 1 overlap, and the shading of colors is also quite intertwined. You should know that many reviewers will print it in black and white, which would make it even harder to see clearly. Also, in the Figure 3D plot, for example, the text "LAO" overlaps with the dots, and the key issue is that black and dark blue affect each other.

3 comments

We thank the reviewer for the thoughtful assessment and for recognizing the value of a robust-control framework for activation steering. We appreciate the direct request for two experiments as conditions for weak accept; we have now conducted both, and the results strengthen our claims.


W2/Q1 (Do H∞ outperform LQR with identical calibration budgets?)

First, we clarify that A-LQR was calibrated throughout our original experiments using the pipeline of Skifstad et al. Where the model was identical to theirs, we reused their published calibrated settings, not an arbitrarily fixed baseline.

We additionally ran the requested fully matched Qwen2.5-0.5B-Instruct/HarmBench comparison. LQR (8D) and H∞ share the 8D basis, calibration split, dynamics, control channels, reference, intervention, 12-setting cost grid, 50 validation prompts per setting, and AXBench Overall selection criterion.

Table R1a. Complete calibration grid (AXBench Overall ↑ ; bold = maximum).

Q/RQf/RA-LQR (full)LQR (8D)H∞
.01.01.364.494.772
.01.1.734.420.744
.011/√10.694.024.044
.1.01.734.001.001

Both 8D controllers independently select the same costs (Q/R, Qf/R) = (.01, .01), yet H∞ achieves .772 vs. .494 on calibration. We froze each controller at its best calibration and tested 240 held-out behaviors, each with five jailbreak variants:

Table R1b. Held-out HarmBench ASR (%) ↓ ; mean ± jackknife SE.

ControllerDirectJailbreak (OOD)
Original32.50 ± 2.3124.42 ± 1.59
A-LQR (full)29.58 ± 2.8116.17 ± 1.61
A-LQR (8D)17.08 ± 2.1911.92 ± 1.17
H∞12.92 ± 2.365.33 ± 0.83

Thus H∞ cuts OOD ASR by 55.3% relative to recalibrated A-LQR, with identical budget, representation, and selected costs, so unequal tuning does not explain the improvement. Our isotropic H∞ ablation in Appendix I completes this section, replicating the separation of "worst case objective" and "empirical covariance geometry".


W2/Q2 (Does capability persist when steering actually succeeds?)

We computed the requested joint metric: Spanish output AND correct mathematics, using paired Qwen3-4B MGSM results (100 prompts per language). Chinese and French inputs give virtually identical steering rates across controllers:

Table R2. Matched-achievement MGSM

InputSpanish A-LQRH∞Spanish & correct A-LQRH∞Fluency A-LQRH∞
Chinese9910020511.151.52
French979828461.111.42
Mean98992448.51.131.47

On the 196 prompts where both methods steered to Spanish, H∞ answered 95 correctly versus 48 for A-LQR, ruling out weaker steering as the explanation for its capability advantage. Granite's resistance to Spanish steering is not specific to H∞ (other methods also fail), indicating a model-dependent limitation rather than a weakness of robust control. Nie et al. (EMNLP 2025) identify model-dependent language-transition mechanisms; Huq et al. (COLM 2026) document Granite's distinctive architecture and irregular layerwise representations. We report these negative results to highlight architectural differences that may limit steering.


W1 (Additional evidences in support of SrobS_{rob})

We thank the reviewer for this important question. We conducted additional analyses and added a new results section to the main text (Fig. R1) examining behavioral gain correlated with SrobS_{rob} and its calibration agnosticity.

SrobS_{rob} correlated with steering gain: We now substantiate our interpretation of SrobS_{rob} with additional direct empirical evidence. Fig. R1(a–c) Analysis of held-out ID/OOD performance across TruthfulQA, HarmBench, and MGSM reveals a significant positive correlation between (SrobS_{rob}) and relative behavioral gain, supporting pre-deployment diagnostic nature of our measure. Fig. R1(e) further shows encouraging cross-task trends from TruthfulQA steerability to HarmBench ASR reduction.

Calibration robustness: While absolute SrobS_{rob} values depend on the cost matrices, Fig. R1(d) show consistent cross-model patterns across 32 shared cost configurations. Llama-3-8B consistently has the highest SrobS_{rob}, and Qwen-2.5-14B the lowest, demonstrating the relative steerability trend aren't artifacts of a particular calibration choice. Fig. R1(f) also shows how directly H∞ marginal limits connect to LQR, supporting our proposition 2 and resolving potential calibration concerns.


Q3 (Figure readability)

Visual suggestions are implemented.


With the two directly requested experiments now being provided, we hope the reviewer will weigh them in their final assessment. Should any questions remain, please do not hesitate to ask.

I thank the authors for the response. I maintain my score.

Thank you for your continued engagement. Your original review indicated that an equal-budget LQR comparison and a capability comparison at matched steering would support a weak-accept recommendation. We conducted both experiments, at substantial computational cost, as requested (Tables R1b and R2): 5.33% vs. 11.92% OOD ASR at equal budget, and 95 vs. 48 correct responses at matched steering. Reviewer AtpL independently raised the equal-budget concern and considered it resolved by these results.

We hope these additional experiments have also helped address your concerns and would greatly appreciate your perspective on whether any questions remain. We would be happy to provide further clarification while the discussion is open and grateful for your consideration of these results in reassessing the score as stated.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.