acceptodds
Under review as a conference paper at ICLR 2027

Reasoning Geometry of Language Models: How Prompting, Training, and Scale Shape Judgment Trajectories

Abstract

Language-model reasoning exhibits behaviorally anchored functional geometry, but prompting and post-training reshape it in fundamentally different ways. Controlled contrasts identify reasoning-associated () and judgment-associated () factors: reasoning is -weighted (), judgment is -weighted (), yielding . A factorial reveals asymmetric interaction: reasoning remains -weighted across policy states, while judgment geometry depends on reasoning context. Prompting and post-training act as distinct geometric operators: prompting reweights engagement of this functional organization, whereas post-training reorients the representation frame in which it is expressed. Held-out Procrustes alignment removes 95% of the final-layer – rotation gap, showing that most angular change reflects shared frame reorientation rather than loss of functional structure. Despite coordinate variation, relative functional organization persists across probes, training states, and tasks. Interventions along either factor are no more consequential than norm-matched random directions, separating behavioral readability from selective steerability. Endpoint shift predicts reflection success (AUROC ). Together, these results identify relational trajectory geometry as a substrate for model observability, distinguishing functional change from representational reorientation across prompting and post-training.

Primary area:
interpretability and explainable AI
Keywords:
representation geometry, reasoning trajectories, language model diagnostics, functional geometry, post-training

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_dzx1OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies how prompting and post-training change hidden-state geometry using divergence, curvature, and two learned readouts associated with reasoning and judgment. It reports relative behavioral separation, shared frame reorientation across checkpoints, and a separate reflection-success prediction result.

Critical strengths
  • Alignment-aware analysis. Held-out Procrustes alignment removes most of the apparent difference between the two factors’ rotation angles. This provides a useful caution against interpreting raw representational rotation as function-specific change.

  • Behavioral validation. The authors evaluate the readouts using independent reasoning-demand and judgment-policy contrasts, alternative probes, and additional tasks. The intervention controls also appropriately distinguish behavioral readability from selective steerability.

Critical weaknesses
  • Unclear functional interpretation. The learned readouts show different sensitivities to reasoning-demand and judgment-policy changes, but their functional labels are unstable across probes and tasks. The evidence supports a relative response difference more clearly than a reproducible characterization of reasoning and judgment.
  • Limited explanatory contribution. Suffix invariance follows directly from causal masking, while curvature effects depend on the measured span and metric. The interpretation that prompting changes functional engagement therefore relies heavily on readout comparisons that are themselves sensitive to probe choice and depth.
  • Disconnected monitoring result. Reflection-success prediction uses raw hidden-state distances rather than the learned readouts, leaving the practical value of the proposed functional organization unestablished. The different instructions and contexts across reflection passes also warrant matched controls to clarify what the predictor captures.
Questions for authors

What is the benefit of learning these two readouts? The reflection predictor does not use them, and steering along them performs no better than random directions. What do they help us understand or accomplish beyond analyzing the original hidden states?

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

This paper investigates the difference of reasoning geometry of large language models across prompt methods, post training processes, and parameter amounts. Specifically, authors train MLP projection heads v (separates entity identity, which is considered to be related to judgment in this work) and u (tracks within-entity temporal order, which is considered to be related to reasoning in this work) on a separate -identity tracking pretext. Varied reasoning and judgement demand change v and u. Experimental results on Qwen models and GSM8k datasets show that prompting reweights engagement of reasoning and judgement, whereas post-training reorients the representation frame. Besides, endpoint shift can predict reflection success.

Critical strengths
  • It is interesting to investigate the reasoning geometry of large language models under different settings.
Critical weaknesses
  • This paper is very hard to follow. Some symbols are introduced without any accompanying explanation, like line 135 "ϕ_u(ℓ) and ϕ_v(ℓ) denote fixed projection heads introduced in Section 5.3".

  • The findings mainly constrained to Qwen models and GSM8k datasets. Authors do not show the generalization to other tasks like MATH or SWE‑bench. The investigation on post-training is coarse for there may be too many stages from Qwen-base model to Qwen-Instruct model. It is better to investigate a fully open-sourced models like Olmo-3.

  • Some claims even may not generalize to all heads (Line 315-317: full signed pattern SR>0, SJ<0 holds only in a subset).

Questions for authors

Refer to the critical weakness. Can your findings be generalized to other large language models like LLaMA-3 or Olmo-3?

Reviewer_Ps7cOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how model activations change with different prompts, additional training, and model size. The authors train an MLP probe that maps model activations to two 64-dimensional feature vectors, u and v. They then compare prompts that ask the model to solve a problem or decide whether to accept a supplied answer, and use these comparisons to label u and v as related to reasoning and judgment. They argue that prompting changes these two measurements differently. They also find that rotating the original model’s activations makes the two models’ measurements more similar, which they interpret as evidence that some relationships survive training. Separately, they find that activation differences between an initial answering prompt and a prompt asking the model to review its answer help predict successful correction.

Critical strengths

None

Critical weaknesses
  1. The interpretation of u and v as reasoning- and judgment-associated measurements is insufficiently established.

The projections are trained to track entity identity and temporal progression. Their subsequent interpretation rests on differences between prompts requesting versus supplying a derivation, and between permissive versus conservative answer policies. These comparisons also change wording, supplied information, and expected responses. Distinguishing these prompts does not establish that the projections track reasoning or judgment themselves, particularly when activations are measured before the answer is generated.

  1. The claimed fundamental distinction between prompting and post-training is not convincingly demonstrated.

Prompting is analyzed through changes under a fixed model and projection network. Post-training is additionally analyzed by fitting a transformation between the models’ activations. These different analyses do not establish that the underlying changes are fundamentally different.

  1. The writing makes the paper very difficult to understand.

It uses extensive jargon and vague language, introducing terms before explaining them, starting in the abstract. Phrases such as “behaviorally anchored functional geometry” and “relative functional organization” appear before the reader has a concrete account of what is measured. Many numerical results follow, but their meaning and connection to the main claims are often unclear. The introduction’s second paragraph largely repeats the abstract’s claims and numerical results instead of developing the motivation and research gap. Also, the introduction repeatedly presents the paper’s findings and interpretations as general facts. As a result, I must read the paper before I can understand the abstract.

Questions for authors
  1. What evidence establishes that u and v track reasoning and judgment beyond sensitivity to the corresponding instructions ?

  2. What comparable analysis supports the distinction between prompting changing engagement and post-training changing the representation frame ?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.