acceptodds
Under review as a conference paper at ICLR 2027

JAM: Controllable and Responsible Text Generation via Causal Reasoning and Latent Vectors Manipulation

Abstract

Inference-time representation steering typically relies on a globally tuned intervention strength, even though different inputs may require different amounts of control. We introduce JAM (Just A Move), which formulates latent steering as constraint satisfaction rather than score maximization. A learned linear probe defines a desired half-space: if an input-specific first-pass representation already satisfies the learned constraint, JAM abstains; otherwise, it computes in closed form the minimum- displacement of that representation required to reach the decision boundary. The resulting prompt- and decode-side components are then used as steering signals during a controlled second generation pass. The learned control surfaces generalize to strictly prompt-disjoint test data, with AUROC between 0.858 and 0.934. Under a matched HHH evaluation across Meta-Llama-3-8B and Mistral-7B-Instruct-v0.2, JAM does not reduce Base ROUGE-2 in any of the six settings while keeping perplexity close to the unmodified model; fixed-strength CAA substantially reduces reference overlap in several honest/helpful settings. Controlled distance sweeps show no consistent reference-overlap benefit from moving 20%–50% beyond the analytically defined boundary distance, supporting the boundary as a tuning-free conservative calibration point. JAM also reduces measured toxicity in the reported RealToxicityPrompts evaluation, with a grammaticality trade-off. The boundary computation itself adds only 1.1 ms, and examples that already satisfy the learned constraint avoid a second generation pass. Together, these results support boundary distance as a simple input-specific rule for deciding whether and how strongly to steer.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Activation Steering, Controllable Text Generation, Representation Engineering, Adaptive Steering, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_4sDCOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors present JAM, an activation steering method that formulates activation steering as an optimization problem with constraints. Given a binary logistic classifier on the activations that project them to the score space ∈[0,1]\in [0,1], JAM introduces the minimum-norm intervention that would result in a "non-toxic" classification. This intervention is the one that transports the current activation orthogonally to the classification hyperplane if classified toxic, and abstains if already classified non-toxic. The authors find that JAM works well in practice and does not significantly impact text fluency.

Critical strengths

N/A; highly redundant with already published work, see Weaknesses.

Critical weaknesses

The presented approach is identical to LiSeCo (https://openreview.net/forum?id=a3o2pzZuvE), an already published controlled generation method in the literature (which has been on Arxiv since 2024, and published June 2026); see Theorem 1 in their paper. Indeed, LiSeCo is even more general than the presented method, permitting controlling activations to level sets that correspond to a given classifier score p(attribute)∈[0,1]p(\text{attribute}) \in [0,1] (in contrast to solely the boundary margin presented in this work). Unfortunately, it is difficult to see how this paper advances beyond LiSeCo and its follow-ups (e.g., MERA, https://openreview.net/forum?id=fUCPq5RvmH).

Questions for authors

None

Reviewer_w3aLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper casts inference-time representation steering as a constraint-satisfaction problem rather than score maximization. A probe is fitted on the concatenation of the prefill-token and last generated-token hidden states and decides whether the representation already lies in the desired half-space; if it does, JAM returns the first-pass answer unchanged and applies no intervention. Otherwise, it computes the displacement to the learned hyperplane in closed form, splits this displacement into a prompt-side and a decode-side component, and injects them during a second, controlled generation pass.

Critical strengths
  1. A reusable, tuning-free boundary-constrained intervention paradigm. The work uses the signed margin of a linear probe to settle two things at once: whether an intervention is needed, and how large the intervention should be for the current example, computed analytically. Unlike most adaptive steering approaches, which rely on an additional trained heuristic controller or a reinforcement-learning policy, JAM obtains the perturbation from a closed-form minimum-distance projection, so its definition is mathematically transparent.

  2. It effectively mitigates a classic pain point of controllable generation: information degradation of the original text caused by intervention. Across all six (model × attribute) settings of the matched HHH evaluation, JAM's ROUGE-2 is never below the base model and PPL stays close to the base. In contrast, the representative fixed-strength baseline CAA shows large ROUGE-2 drops on the honesty and helpfulness tasks — that is, it tampers with or loses the source facts and reference information in order to achieve attribute control. By perturbing only the examples that violate the constraint, and by the minimum distance, JAM protects the fidelity of the original generation while pursuing attribute improvement. The accompanying mechanism ablations (intervention-distance sweep, 2×2 factorial, intervention-layer sweep) empirically show that moving to the decision boundary is already sufficient and that increasing the perturbation beyond it yields no stable metric gain, which mechanically supports the rationale of this conservative, tuning-free policy.

Critical weaknesses
  1. Limited cross-model robustness, with clearly divergent attribute gains. Experiments are conducted only on two 7–8B models (Llama-3-8B, Mistral-7B-Instruct-v0.2). On Llama-3-8B, the GPT-4 pairwise preference shows a sizeable gain for JAM over attribute prompting; on Mistral-7B-Instruct-v0.2, however, the overwhelming majority of pairs are draws and the attribute improvement is not obvious. This indicates that the practical gain depends heavily on the intrinsic representation properties of the base model, and that consistent attribute improvement across models cannot be guaranteed. The toxicity task also shows a clear trade-off: toxicity falls while grammaticality degrades.

  2. Narrow generalization scope; no evidence on larger or newer base models. The experiments are confined to the 7–8B range and are not validated on larger models (e.g. 13B/34B) or on more recently released open base models, so it cannot be confirmed whether this boundary-minimum-distance steering paradigm still holds.

  3. The novelty argument and the related-work comparison are insufficient, and the background on adaptive magnitude is not adequately reviewed. Section 5 concentrates on fixed-strength activation-steering methods such as CAA and ITI, but it is unclear whether other work on input- or example-adaptive intervention strength in inference-time steering has appeared in the past year or two. Does any existing approach already modulate the intervention magnitude per example? The reader cannot judge whether the idea of per-example adaptive magnitude already has precedent, nor whether JAM's novelty lies in "proposing adaptive magnitude" or in "realizing adaptive magnitude and intervention gating without extra training, through decision-boundary geometry plus a closed-form minimum-L2 projection".

Questions for authors

See above.

Reviewer_R5ZtOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes JAM (Just A Move), a LLM steering method based on constraint satisfaction. The main idea is to use a linear probe over latent vectors to detect and classify desired or undesired behavior, e.g., harmful versus safe output, and only if that condition is satisfied apply generation steering.

Critical strengths

The paper shows that is possible to use simple linear probes to understand and modify the model behavior, which confirms the Linear Representation Hypothesis nicely and showcases its generality and usefulness.

Critical weaknesses

The largest issue of this paper is the lack of novelty since very similar approaches have been proposed in recent years. For example, Xu et al., “Uncovering Safety Risks of Large Language Models through Concept Activation Vector,” NeurIPS 2024, introduced SCAV, Its §2.3.1, Eqs. 3-4 already provides a very similar min distance intervention condition using linear probes and its Appendix C provides a proof of optimality which this paper does not. This is the same gated half-space projection operation used by JAM, and even though it differs from SCAV in some details and application, my understanding is that contributions 1 and 2 are not novel.

Furthermore, there is substantial prior work that needs to be referenced and discussed as the paper greatly lacks comparison with the existing alternative methods [1, 2, 3] and has little fundamental motivation [4,5] for why this should work in the first place.

  • [1] Adaptive Activation Steering, Wang et al, WWW 2025
  • [2] Programming Refusal with Conditional Activation Steering, Lee et al., ICLR 2025
  • [3] AdaSteer, Zhao et al., EMNLP 2025
  • [4] The Linear Representation Hypothesis and the Geometry of Large Language Models, Park et al., ICML 2024
  • [5] Frame Representation Hypothesis, Valois et al., TACL 2025

Also, the criticism and comparison with ITI and CAA need further details, experiments, and reference their peer-reviewed versions instead of pre-prints.

In that sense, the experimental results are not sufficient evidence with the Table 2 comparing ROUGE-2 and perplexity and not demonstrating that the outputs become more honest, helpful, or harmless, focusing on showing just preservation of quality. The paper needs to evaluate control and preservation together and better ablations to support all the claims.

Questions for authors
  1. Can you position JAM against SCAV carefully and show how they differ? Isolate what is unique about your design.
  2. Can you explain the method from first principles or motivating it more carefully?
Reviewer_VxRMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

JAM reframes inference-time representation steering as constraint satisfaction instead of score maximization. A linear probe is trained on a hidden-state representation that concatenates the last prefill token and the last generated token (layer 20) to define a desired half-space for an attribute (harmless, honest, helpful, plus toxicity). At inference a first-pass generation produces a representation. If the signed margin is already non-negative JAM abstains and returns the first-pass answer. Otherwise it computes the minimum l2 displacement to the learned decision boundary in closed form (a standard projection onto a hyperplane), splits it into prompt-side and decode-side components, and steers a second generation pass. Probes generalize to strictly prompt-disjoint test data (AUROC 0.858 to 0.934). Under a matched HHH evaluation on Llama-3-8B and Mistral-7B, JAM does not reduce Base ROUGE-2 in any of six settings, while CAA collapses ROUGE-2 in several honest and helpful settings. A distance sweep shows no consistent benefit from steering 20 to 50 percent beyond the boundary. Toxicity is reduced with a grammaticality trade-off, and boundary computation adds about 1.1ms.

Critical strengths
  1. The core idea is simple and stated precisely. The signed margin of a learned linear boundary answers two questions at once, whether to steer and by how much, and it replaces a global steering strength with an input-specific, tuning-free calibration point. Abstaining on examples that already satisfy the constraint is sensible.

  2. The methodology is careful. Strictly prompt-disjoint probe splits, frozen validation-selected baselines, and a clear separation between the matched comparison and the mechanism studies all make the paper trustworthy.

Critical weaknesses
  1. The main result does not show that JAM actually improves the target attribute. The headline is a preservation claim (JAM does not reduce Base ROUGE-2), but ROUGE-2 values are tiny (0.02 to 0.08) and ROUGE-2 measures reference overlap, not harmlessness or honesty. The GPT-4 pairwise evaluation is mixed (Llama3 harmless 40.3% win vs 31.0% loss, Mistral mostly draws), so there is no convincing evidence that steering changes the attribute as intended.

  2. The comparison to CAA is confounded and the paper's own analysis undercuts its argument. CAA differs from JAM in both direction construction and intervention policy, so the preservation result cannot be attributed to adaptive magnitude. The 2x2 factorial (Table 8) reports that the two proposed factors, input-dependent magnitude and abstention, have small, mixed-sign, non-significant effects on ROUGE-2 after correction. The benefit appears to come from a gentler probe-normal direction, not from the calibration rule. The toxicity gain also comes with a large grammaticality drop (0.763 to 0.737, and 0.642 to 0.566).

Questions for authors
  1. The 2x2 factorial reports that neither adaptive magnitude nor abstention significantly changes ROUGE-2. What then is the demonstrated benefit of the boundary-conditioned calibration rule over a fixed strength along the same probe-normal direction?

  2. ROUGE-2 against positive references is a weak proxy for attribute satisfaction. Can you provide an independent measure, ideally human or judge-based, showing that JAM outputs are actually more harmless, honest, or helpful rather than just less different from the base model?

  3. For toxicity, the grammaticality regression is large. Is there a setting that reduces toxicity while holding grammaticality near baseline, or is the trade-off inherent to the steering direction?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.