acceptodds
Under review as a conference paper at ICLR 2027

Geometric Alignment: Training Against Concentrated Behavioral Geometry

Abstract

Many alignment interventions assume undesirable behavior lives in a compact representation direction. Before removing it, we ask whether a coherent direction exists and whether ordinary training can avoid it without an inference-time edit. We answer with three falsifiable gates—concentration (A), geometry-held-out exclusion after geometric training (B), and an opener-CE diagnostic (C)—and the training bridge L = L_task + λ G_bad, exact for linear readouts as w^T Σ_bad w. On a planted residual-stream world the chain passes under geometric training with no post-hoc scrub (held-out response drop 0.989). On a Gemma-4-12B educational proxy, KL-regularized LoRA with a projected-energy penalty cuts bad-subspace energy by 32.5% on a six-prompt report slice while the specified opener-CE diagnostic stays within its Gate C band, again with no activation scrub; bridge CE stays near-flat, so Gate B is energy, not CE. The same A→B→C design on Qwen2.5-7B-Instruct (same prompts and splits) yields a 31.8% report-slice energy drop and opener-CE +0.1%: a protocol replication, not evidence of universal generalization. A second family (sycophancy, Gemma) is NO-GO at layer 16 and GO (−44.3%) at layer 48: concentration licenses the question, but layer still matters. On these prompt families, concentrated displacement geometry can be suppressed in projected energy without moving that opener-CE diagnostic out of band; if concentration fails, the gates say NO-GO.

Primary area:
interpretability and explainable AI
Keywords:
geometric alignment, representation geometry, subspace exclusion, LoRA, falsifiable gates

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_zfRrOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper asks whether a undesirable behavioral family (compliance and sycophancy prefix pairs) occupies a compact subspace of the residual stream, and whether ordinary training can suppress it without an inference-time edit. It proposes a three-gate protocol:

  • Gate A (concentration): top-4 energy of the centered covariance Σb=Cov(Δz)\Sigma_b=\mathrm{Cov}(\Delta z) of paired activation differences, compared against a matched-energy isotropic control.
  • Gate B (exclusion by training): relative drop in held-out projected energy R=E∥P^⊤h∥2R=\mathbb{E}\|\hat P^\top h\|^2 (or held-out response E∣W⋅Δz∣\mathbb{E}|W\cdot\Delta z| in the synthetic case) after training with L=Ltask+λ GbadL=L_{\text{task}}+\lambda\,\mathcal{G}_{\text{bad}}. GO requires a drop of at least 30%.
  • Gate C (capability): a benign "opener" cross-entropy must stay within 15% relative change.

Proposition 1 shows that for a linear readout the penalty equals w⊤Σbadww^\top\Sigma_{\text{bad}}w. Evidence comes from three settings: (i) a synthetic contaminated-label residual stream (held-out response drop 0.989, clean accuracy non-decreasing); (ii) Gemma-4-12B with KL-regularized LoRA and a normalized projected-energy penalty (32.5% drop on a 6-prompt report slice, opener-CE +0.8%); (iii) Qwen2.5-7B-Instruct with the same prompts and splits (31.8%, opener-CE +0.1%). A sycophancy family is NO-GO at layer 16 and GO (−44.3%-44.3\%) at layer 48. Activation scrubbing is relegated to an appendix and is only marginal.

Critical strengths
  1. Candid, falsifiable framing. The paper keeps its negative results: sycophancy fails at layer 16, null prefix families are rejected by the mid-layer rule, scrubbing is only marginal, and the CE ablation shows compliance CE is driven by the μ\mu term and not by geometry. The observation that concentration (Gate A) licenses the question but does not imply trainable suppression (Gate B) is useful, and the limitations are stated plainly.

  2. Clean synthetic sanity check. The planted-subspace setup with contaminated labels (Sec. 7.1) is well constructed, because it makes exclusion non-vacuous. The λ\lambda sweep (Appendix A) shows a smooth, monotone drop in held-out response (0.042 to 0.989) with clean accuracy essentially unchanged. This confirms that the penalty behaves as Proposition 1 predicts in the linear case.

Critical weaknesses
  1. The headline metric is the quantity being optimized, and it is not tied to behavior. Eq. (3) directly minimizes projected energy ∥P^⊤h32comp∥2\|\hat P^\top h_{32}^{\text{comp}}\|^2 on the compliance forward at layer 32. Observing that this energy is lower on six other prompts is largely expected and does not show that the targeted behavior or its causal pathway was reduced. The paper's behavior-adjacent measures do not move: compliance CE changes by −0.8%-0.8\% (Gemma) and −3.3%-3.3\% (Qwen), and the preference margin is slightly adverse (−1.83→−1.67-1.83\to-1.67). The CE ablation shows CE responds only to the μ\mu term, and inference-time scrubbing raises compliance CE by 44.7% whereas training does not. LoRA is applied at layers {28,…,36}\{28,\dots,36\}, including layers downstream of the penalized layer 32, so the model may simply recompute the same information elsewhere. The paper concedes that an energy drop is not a claim that behavior fell (Sec. 8), but that leaves the central claim ("training can suppress concentrated behavioral geometry") supported only by a proxy that the method itself optimizes.

  2. Statistical and protocol weaknesses.

    • The reported number rests on a 6-prompt report slice that is a post-hoc partition of a 13-prompt pool already used in an earlier pilot (−30.3%-30.3\%). There are no seeds or confidence intervals, and the 30% GO threshold is cleared by about 2.5 points in a single run. The selected checkpoint is early-stop@80, which is also the maximum step count, so no early stopping actually occurred.
    • The mid-layer Gate A requirement was added after seeing results (Sec. 6 addendum), and layer choice acts as a tuned hyperparameter (sycophancy fails at layer 16 yet passes at layer 48 with −44.3%-44.3\%).
    • Gate A is weak evidence of concentration: top-4 energy is 0.527 at Gemma layer 32 against 0.284 for the isotropic control, the rank-2 penalty covers only 0.363 of train energy and 0.335 of held-out centered energy, and the null families also reach GO at layer 48.
    • Gate C is a single opener-CE metric, and the paper itself states that general capability preservation is not shown. The Qwen "protocol replication" uses the same prompts and splits, so it says little about generality.
  3. No baselines and limited novelty. There is no comparison with a random rank-2 subspace penalty, a mean-difference (μ\mu) direction penalty, training-time LEACE/INLP-style variants, representation noising or circuit breakers, or a plain KL/preference baseline. The theory is a one-line quadratic identity (Proposition 1), and Remarks 1 and 2 are imported from a self-cited preprint. The related training-time representation methods are acknowledged in Sec. 2 but not compared against, so the contribution over them is unclear.

Questions for authors
  1. (Material.) Does the energy-penalized model show any change in an outcome that is not the training target, such as the rate of compliant or sycophantic generations judged independently, or the response to a steering vector along P^\hat P? Without this, what does the energy drop show beyond successful optimization?

  2. (Material.) Does the effect survive measurement away from the penalized site, for example energy at layers 36 or 40, a different token window, or a probe retrained after training? Is the bad subspace removed, or merely rotated or recomputed downstream?

  3. (Material.) Please add controls: a random rank-2 subspace penalty and a mean-difference direction penalty at matched λ\lambda, with multiple seeds, confidence intervals, and a larger prompt set (n≫6n\gg 6). How stable is the GO/NO-GO call across alternative splits of the 13-prompt pool?

  4. (Clarification.) The selected checkpoint is λ=20\lambda=20 with early-stop@80, equal to the maximum number of steps. Was early stopping ever active? Would the selected λ\lambda or the report-slice drop change if the calib and report slices were swapped?

  5. (Clarification.) How were the layer choices (32 for Gemma, 16 for Qwen) made, given that the mid-layer rule was added after the null-family results? For sycophancy, would layer 48 have been selected without the failure at layer 16?

Reviewer_i5CtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates whether behavioral representations in large language models are concentrated in low-dimensional subspaces and whether such representations can be suppressed through training without relying on inference-time activation editing. The authors propose a three-gate validation protocol (A→B→C) to examine the geometric concentration of behavioral representations, the effectiveness of geometry-based suppression after training, and its impact on a specified capability diagnostic. Experiments in a synthetic setting and on two language models, Gemma and Qwen, show that the proposed method substantially reduces held-out response in the synthetic setting and projected energy in the estimated subspaces in the LLM experiments, while keeping the corresponding capability diagnostics within predefined bounds.

Critical strengths
  1. The research question is meaningful. This paper explores suppressing specific behavioral representations through training rather than activation editing at inference time, offering a valuable direction for investigating interventions in the internal representations of large language models.
  2. The limitations are reported with reasonable transparency.The authors evaluate the proposed geometry-based training approach on synthetic data and two model families, Gemma and Qwen, and report some unsuccessful results. The paper also explicitly acknowledges the retrospective data split and the limitations of the evaluation metrics, which helps readers interpret the experimental findings appropriately.
Critical weaknesses
  1. Lack novelty and insufficient justification for key design choices. The core method largely combines existing techniques, including PCA-based subspace estimation, quadratic regularization, KL constraints, and LoRA fine-tuning. The theoretical analysis also relies heavily on established results from linear-quadratic optimization. As such, the main contribution appears to lie more in the design of the A→B→C empirical validation protocol than in the development of a new training algorithm. However, the current ablation studies do not sufficiently isolate the contribution of the target behavior's geometric structure. A systematic analysis of key components and design choices, such as intervention rank, layer selection, and regularization scheme, is also lacking. Consequently, the distinctiveness and necessity of the proposed method relative to existing techniques remain insufficiently demonstrated.
  2. Insufficient experimental validation, with limited evidence supporting the method's practical effectiveness. The optimization objective and the primary evaluation metrics are closely coupled. The LLM experiments mainly demonstrate a reduction in projection energy within the target subspace, which is itself directly optimized during training. However, independent evaluations of actual model behavior are lacking, making it difficult to determine whether the reduction in geometric energy translates into meaningful behavioral improvements. Furthermore, the experimental scale is relatively small. The core Gemma results are based on only six retrospective report prompts, with no uncertainty estimates or sufficient validation of generalization across tasks. This makes it difficult to assess the robustness of the reported results and the resulting GO decisions.
  3. In terms of presentation, some figures and tables suffer from small sizes, crowded labels, and suboptimal layouts. The paper also introduces a considerable amount of specialized terminology, while key results are scattered across the main text and appendix. These issues affect the overall readability and coherence of the paper.
Questions for authors

Q1. Beyond the reduction in target subspace projection energy, could the authors provide an independent evaluation of actual model behavior to demonstrate that lower geometric energy indeed translates into improvements in the model's generated outputs? Q2. Could the authors include carefully matched training-time baselines (e.g., random subspace regularization) to establish whether the observed improvements primarily stem from the estimated behavioral geometry rather than general representation regularization? Q3. Given that the main results are based on only six retrospective report prompts, could the authors validate their findings on a new, independent test set and provide uncertainty estimates across multiple random seeds or different data splits to demonstrate the robustness of the Gate B GO decision?

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

The paper appears to investigate directional alignment in learned neural network representations. It proposes a diagnostic for determining whether representations exhibit coherent alignment directions, before applying methods to suppress this alignment during training or remove it at inference. Beyond this broad outline, serious problems with the writing prevented me from understanding the methods and contributions.

Critical strengths

I could not identify any substantive strengths in this paper, which reads as largely AI-generated with limited to no human oversight.

Critical weaknesses

The paper and its central ideas are extremely difficult to follow. Short sentences and phrases are strung together with semicolons, with little context, explanation, or discussion. The introduction opens with a two-sentence paragraph that provides no background, clear problem statement, or supporting citations. The conclusion begins: “On the studied families, if the displacement cloud concentrates, training can cut that projected energy while the specified opener-CE diagnostic stays in band.” This sentence exemplifies the paper’s opaque language and unexplained terminology. These pervasive writing problems make it difficult to understand the claims or assess the contribution.

Questions for authors

None, I couldn't get past the serious writing issues.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.