acceptodds
Under review as a conference paper at ICLR 2027

Latent Reasoning Does Not Lack Signal, It Lacks Invariance: Subspace-Constrained Refinement for Robust Reasoning

Abstract

Latent reasoning methods in LLMs perform multi-step inference in embedding space, avoiding the token overhead of Chain-of-Thought while preserving reasoning capacity. Recent training-free refinements steer these latent trajectories using contrastive signals from reference checkpoints, yielding 2-5% accuracy gains at zero parameter cost. We show that these gains are fragile along three aspects: (i) hyperparameter sensitivity, where the step size and memory rate have task-specific optima with no per-instance adaptation; (ii) reference-pair dependence, where the contrastive direction from a single checkpoint pair can reflect pair-specific artifacts rather than stable reasoning structure; and (iii) input sensitivity, where updates in the full embedding space let surface features push equivalent inputs onto different trajectories. We argue that reliable latent refinement requires trajectory-level stability, not only answer-level accuracy, and that the refinement must degrade gracefully when its own signal is unreliable. We propose Trajectory-Invariant Latent Refinement (TILR), a training-free framework with two components. First, we identify a low-rank subspace in which reference models consistently differ across a calibration set, and project contrastive updates onto this subspace to filter unstable variation. Second, we apply an adaptive gating mechanism that scales update magnitude according to the alignment between the current signal and the learned subspace, suppressing unreliable corrections and recovering the unmodified backbone when alignment is low. Across six reasoning benchmarks, TILR consistently improves over prior refinement methods while substantially reducing sensitivity to input perturbations and reference selections. In particular, TILR improves answer agreement under semantically equivalent paraphrasing by 11.4% and reduces trajectory variance by  30% relative to the Coconut backbone, while maintaining comparable accuracy. These results demonstrate that constraining refinement to invariant subspaces yields more reliable latent reasoning without requiring retraining or backpropagation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Latent Reasoning, Chain-Of-Thought Reasoning, Representation Subspaces, Inference-time Robustness, Invariance

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_niJGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a training-free optimization framework for latent space reasoning, Trajectory-Invariant Latent Refinement (TILR), aiming to address the sensitivity of existing latent space reasoning correction methods to hyperparameters, reference model selection, and input perturbations. Specifically, TILR first extracts a stable correction subspace from the representational differences between strong and weak reference models using Singular Value Decomposition (SVD), and then dynamically adjusts the correction strength of the latent states using an Adaptive Stability Gate, thereby improving the consistency and robustness of the reasoning trajectory. The authors conducted experiments on six mathematical and symbolic reasoning benchmarks and further validated the applicability of the method on different model architectures. Experimental results show that TILR achieves a certain accuracy improvement compared to existing methods while maintaining low additional computational overhead, and significantly improves answer consistency and trajectory stability under input perturbations.

Critical strengths
  1. Training-free optimization has practical value. TILR can improve latent space inference without updating model parameters or performing backpropagation, offering certain deployment convenience.
  2. The adaptive correction mechanism is value. The Adaptive Stability Gate adjusts the update intensity based on the alignment between the correction direction and the learning subspace, helping to suppress unreliable corrections under low alignment conditions.
  3. The paper has a complete and clear structure.
Critical weaknesses
  1. The theoretical basis for invariant subspaces is insufficient. Current cannot directly guarantee their invariance to the input or reference model.
  2. The reliability of adaptive gating is insufficient. Subspace alignment does not necessarily represent the correctness or effectiveness of the correction direction, and gating offers only a small accuracy gain compared to individual projection, lacking a more direct mechanism for verification.
  3. On some datasets, TILR shows only a small improvement over Refinement.
  4. The experimental model scale and task coverage are limited. The main experiments are primarily based on GPT-2, with the appendix only extending to 1B–1.5B models. Furthermore, the tasks focus on mathematical and symbolic reasoning, which is insufficient to verify the method's generalization ability on larger-scale models and complex reasoning tasks.
Questions for authors
  1. It is recommended to analyze the correlation between subspace alignment and actual correction gains.
  2. Further validate the effectiveness of TILR on 7B or larger models, and on more complex tasks such as multi-hop inference and open-domain inference?
  3. It is recommended that the authors report the complete inference cost, including forward computation of the reference model and subspace calibration, and provide an end-to-end efficiency comparison with the original Coconut.
4 comments
AuthorsOpenReview ↗

Response to Reviewer niJG: Experimental Validation of Invariance, Gate Reliability, and Scaling to 7B

Thank you for the constructive feedback and recognition of our work. We have addressed your concerns with the experimental results below.

W1. Theoretical basis for invariance. Appendix B makes no formal claim that SrS_r contains causally invariant features. Eq. 12 instead gives a conditional result: when two corrections differ only outside SrS_r and have equal norm, TILR applies the same projected correction, introducing no additional divergence beyond the backbone. Appendix D bounds step-size sensitivity by γ2\gamma^2 of unconstrained refinement. The empirical question is how much real variation lies outside SrS_r. We now measure this on 920 paraphrased GSM8K questions (lower filters more):

Subspace (rank 171)Paraphrase changeReference swap
Calibrated (TILR)0.8620.776
Input PCA0.9770.888
Random1.0021.003

W2. Reliability of the gate. γ measures geometric support for the current reference signal. We directly test whether it tracks reference reliability by progressively corrupting the reference and comparing projection-only refinement with TILR at the same η (GSM8K test, n = 1,319, change vs Coconut in points):

ReferenceProjection onlyTILRHarm removedγ
TILR Original+0.99+1.06no cost0.79
Swapped for under-trained−1.44−0.2384%–
+ noise 1×−2.12−0.6868%0.50
+ noise 3×−13.95−5.5360%0.46

Under noise γ falls to the value of a direction with no preference for the subspace (√(171/768) = 0.47), and the step halves. In a second run where gated and ungated corrections are matched in average step size, 3× noise costs 4.8 points with the gate and 12.7 without. The gate's gains are in robustness, not average accuracy: the hyperparameter sensitivity gap is 10–18× smaller (Table 4).

W3. Small improvement over Refinement. We agree that the mean-accuracy improvement over Refinement is modest; this is consistent with the paper's stated objective of improving robustness while preserving accuracy, and its margin is largest there:

  • worst-case accuracy over the (α, η) neighbourhood: +5.6 / +3.8 / +11.3 points over refinement on GSM8K / GSM-Plus / SVAMP (Table 4);
  • across 9 reference pairs, TILR never falls below the backbone (std 0.41 points); refinement does on 4 of 9 (std 2.21);
  • answered correctly under all five phrasings: TILR 12.0%, refinement 10.2%, Coconut 8.9%.

W4 / Q2. Scale and tasks. Appendix H reports Qwen2.5-Math-1.5B and Llama-3.2-1B, with gains in accuracy, M3 and M4 on all six backbone-benchmark pairs. New: we trained Coconut on Qwen2.5-7B and ran the full pipeline. TILR gives comaparable or slightly better accuracy (GSM8K 33.66% vs 33.51%; SVAMP 69.9% vs 69.7%) while reducing trajectory variance under paraphrasing by 11% (243.8 vs 272.8). For multi-hop inference, ProsQA and ProntoQA are multi-hop deductive tasks, where TILR is the best latent refinement method (Table 1: 90.4%, 94.7%).

Qwen2.5-7BGSM8K acc.SVAMP acc.Trajectory variance, GSM8K paraphrases (M4, ↓)
Coconut33.51%69.7%272.8
TILR33.66%69.9%243.8 (−11%)

Q1. Alignment vs correction gain. Across controlled reference corruptions, lower γ coincides with lower correction reliability. Crucially, this is not explained merely by smaller steps: when gated and ungated updates are matched in average step magnitude, 3× reference noise reduces accuracy by 4.8 points with gating versus 12.7 without.

Q3. End-to-end cost. One RTX 6000 Ada, batch size 1, GSM8K test (200 questions), all reference forward passes included:

Methodms / questionForward passes× Coconut
CoT96.132.73.34
Coconut28.810.31.00
Refinement67.924.32.36
TILR68.724.32.4

The projection and gate add 1% over refinement; TILR is 28.5% faster than CoT. Calibration takes approximately 10 seconds.

In summary, we provide direct evidence addressing each concern:

  • W1: the guarantee is present in the paper (Appendix B, Eq. 12; Appendix D), and the filtering it relies on is now measured: the calibrated subspace filters both perturbations substantially more strongly.
  • W2, Q1: γ tracks reference reliability, drops to chance level under corruption, and the gate then removes 60-84% of the harm.
  • W3: As written in the paper, TILR's margin is largest in: worst-case accuracy, reference stability, consistency across phrasings.
  • W4, Q2: At 7B, TILR gives comparable accuracy as claimed in the abstract and makes trajectories 11% more consistent, and it is the best latent refinement method on the multi-hop tasks.
  • Q3: measured end-to-end: 2.4× Coconut, 28.5% below CoT, 10 s calibration.

We hope these additional results address all your concerns and clarify that TILR's contribution is robust refinement rather than mean-accuracy improvement.

AuthorsOpenReview ↗

Follow-up on Our Response and Additional Results

Dear Reviewer,

Thank you again for your review. We have conducted the additional experiments suggested in your review and incorporated the results into the revised manuscript.

We would appreciate it if you could revisit our response when you have a chance. We would be happy to address any remaining concerns during the discussion period.

I appreciate the authors' responses and supplementary experiments, which have clarified some of the questions. However, the current manuscript still has some shortcomings. Existing evidence supports effective filtering under specific perturbations, but its robustness and transferability are limited. Furthermore, the current manuscript's analysis of the method's applicability conditions and failure scenarios is insufficient. Additionally, based on the presented experimental results, the accuracy improvement on larger-scale models is limited, while the complete inference process still incurs significant additional computational overhead. Considering the above issues and the comments of other reviewers, I maintain the original score.

AuthorsOpenReview ↗

Clarification on Remaining Concerns

Dear Reviewer,

Thank you for revisiting our response and additional experiments. We appreciate your comments. We would like to briefly clarify the remaining points, as the revised manuscript and rebuttal experiments directly evaluate and address each of these concerns.

Remaining concernsEvidence and results directly addressing the concerns
Robustness / transferability remains limitedThe added experiments tests transfer across 7 alternative references, including an independent training seed (fixed SrS_r: 27.63% vs. 27.85% with per-reference recalibration); across benchmarks without recalibration (SVAMP: +1.7 pp, p=.014p=.014; GSM-Plus: +0.8 pp, p=.0004p=.0004); across GPT-2, Llama, and Qwen architectures up to 7B; and across latent-reasoning frameworks. On CODI, TILR reduces trajectory variance by 47-48% while retaining/improving accuracy. These results show that TILR's robustness effect transfers successfully across references, training seeds, benchmarks, architectures, model scale, and latent-reasoning frameworks.
Accuracy improvement at larger scale is limitedThe paper's objective is robustness while preserving accuracy, rather than maximizing accuracy gains. At 7B, the strongest reference provides only 0.7 pp validation headroom; nevertheless, TILR preserves accuracy (33.51% →\rightarrow 33.66%) while reducing paraphrase trajectory variance by 11%. Thus, the 7B result directly validates the central robustness claim at larger scale.
Complete inference incurs significant overheadThe end-to-end experiment directly measures this concern, including all reference forward passes. TILR requires 68.7 ms/question vs. 67.9 ms/question for unconstrained Refinement, i.e., only 1.2% incremental latency. Both methods use the same 24.3 forward passes; therefore, the substantial cost relative to Coconut (28.8 ms) comes from the contrastive reference inference already required by the baseline Refinement method, not from TILR's subspace projection or adaptive gating. TILR is also 28.5% faster than CoT in the same measurement, and calibration is a one-time ~10 s cost. Thus, the revised end-to-end measurement directly quantifies the computational concern and shows that the proposed TILR mechanism itself adds negligible overhead while providing significant robustness improvement.
Applicability conditions / failure scenarios are insufficiently analyzedWe directly experiment the intended failure regime by progressively degrading reference reliability. As the reference deteriorates, the gate responds as designed (γ:0.79→0.46\gamma: 0.79 \rightarrow 0.46) and removes 60-84% of projection-induced degradation. Under matched correction magnitude, 3×3\times reference noise causes only a 4.8 pp accuracy loss with gating versus 12.7 pp without, ruling out smaller updates as the explanation. This experimentally characterizes when gating becomes important and how TILR behaves when its correction signal becomes unreliable. The revised Limitations separately scopes conditions not established by our experiments, including larger distribution shifts, different reasoning domains, calibration dependence, and the distinction between invariance and correctness.

We agree that broader evaluation across every possible model, domain, and distribution shift would provide additional evidence. However, these settings are outside the paper's stated scope rather than assumptions required by its claims. Within the stated scope, the revised manuscript now directly addresses the robustness, transferability, scale, computational cost, and failure-regime questions raised above.

If there is a specific stated claim that you believe remains unsupported after these experiments, we would greatly appreciate clarification so that we can address it precisely during the remaining discussion period. Thank you.

Reviewer_mEL1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies inference-time contrastive refinement for continuous latent reasoning models like Coconut. The authors run SVD on hidden-state differences between strong and weak reference checkpoints over 200 calibration samples, project updates onto the top singular subspace, and scale the step size by the squared alignment score.

Critical strengths
  1. Filtering out noisy directions in contrastive latent updates via SVD subspace projection is intuitive and well-motivated.

  2. The proposed projection and alignment-based gating are training-free and easy to plug into existing latent refinement pipelines.

Critical weaknesses
  1. Overclaimed invariance and no cross-benchmark transfer.

a) SVD is only run on a single reference pair and standard training samples, so it only captures high-energy directions rather than invariance. In Table 3, the rank is widely spreading (38–171) with up to 0.804 overlap with input PCA, indicating the subspace is neither low-rank nor separated from surface input features.

b) Each benchmark fits its own Coconut model and SVD subspace, with no test of whether a subspace calibrated on a benchmark transfers to other benchmarks.

  1. Limited baselines and unclear gating gain.

a) All latent baselines build on Coconut, which performs poorly here. Stronger recent methods like Soft Thinking (Zhang et al., NIPS 2025), CODI (Shen et al., EMNLP 2025), and SoftCoT (Xu et al., ACL 2025) are not discussed or compared.

b) In Table 5, adding the adaptive gate over projection-only improves by just 7e-4, showing little benefit from gating.

  1. Limited experimental settings on model archs and tasks.

Experiments only test dense models on mainly GSM8K-style arithmetic tasks, leaving out MoE architectures, harder math like MATH500 and AIME, or other classical coding/reasoning tasks like HumanEval, MBPP, BBH, etc.

  1. Presentation inconsistencies. As far as I understood,

a) Fig. 2 gates memory retention and uses α\alpha in residual blending, whereas Sec. 4.2 explicitly keeps α\alpha fixed.

b) Fig. 2 also uses a linear ηt=ηbaseγt\eta_t = \eta_{base}\gamma_t, whereas Sec. 4.2 defines ηt=min⁡(ηbaseγt2,1)\eta_t = \min(\eta_{base}\gamma_t^2, 1). This makes it unclear which update rule produced the reported results.

Questions for authors
  1. Does a subspace calibrated on GSM8K transfer directly to other benchmarks without per-benchmark recalibration?

  2. Could the authors provide comparison with stronger baselines or backbones mentioned above, and clarify whether the gate brings meaningful gains beyond 7e-4 on other tasks?

  3. How does the method perform on MoE models, and on harder benchmarks or coding/reasoning tasks?

  4. Please clarify which update rule was actually used in the experiments, why it was chosen over the alternative shown elsewhere in the paper, and provide ablations.

7 comments
AuthorsOpenReview ↗

Response to Reviewer mEL1: Addressing Subspace validity, reference robustness, cross-domain transfer, and CODI

Thank you for the detailed review and for finding the projection intuitive, well motivated, and easy to integrate. Several concerns are already tested and addressed in the submission; below we clarify these points and provide new experiments directly targeting the remaining questions. These experiments use the paper's GPT-2 Coconut backbone on GSM8K-Aug with its reference pair and validation-selected (α,η)(\alpha,\eta).

W1. TILR's Invariance claim.

As Appendix B states, we make no formal claim that SrS_r contains causally invariant features. Operational invariance is measured directly: TILR reduces reference sensitivity by 53-73% (RQ2) and improves paraphrase agreement by 11.4% while reducing trajectory variance on every benchmark (RQ3). Thus, invariance is established by measured robustness under reference and input perturbations; RQ4 separately characterizes the geometry underlying these effects.

W1a. Does SrS_r capture invariance, or merely high-energy/input directions?

SrS_r is computed from strong-minus-weak reference differences, not from latent states. The high-energy subspace of the states is precisely the input-PCA control already tested in RQ5/Table 5 (23.4% accuracy vs. 27.7% for TILR).

We now directly measure how much perturbation-induced variation each rank-171 subspace admits on 920 paraphrased GSM8K questions (1.0 = no filtering; lower is better):

SubspaceParaphrase changeReference swap
Calibrated SrS_r0.8620.776
Input PCA0.9770.888
Random1.0021.003

Thus, the calibrated subspace filters both perturbations substantially more strongly than input PCA or a rank-matched random subspace. This difference also appears in trajectories: with equal average correction magnitude and no gate, M4 is 262.1262.1 using SrS_r versus 269.2269.2 using input PCA.

Regarding rank, our claim is spectral concentration rather than rank collapse: only 10/768 directions capture 76--84% of contrastive energy, with a task-dependent tail (r0.90=38r_{0.90}=38--171; Table 3).

W1a. Is the result tied to one reference pair?

No. We evaluated nine additional reference pairs, including three references from a separately trained model:

9 reference pairsStd. accuracyWorstBelow backbone (26.61%)
Refinement2.21 pts22.67%4/9
TILR0.41 pts27.14%0/9

We then kept the paper's SrS_r fixed and swapped with seven good references. TILR remains above the backbone with every reference (26.84--28.20%). The fixed subspace averages 27.63%, versus 27.85% when recalibrated per reference, only 0.22 points apart. One calibration therefore serves multiple references.

W1b / Q1. Does a calibrated subspace transfer across tasks?

Yes. SVAMP, GSM-Plus, and GSM-Symbolic already use the GSM8K-Aug-trained backbone and transferred (α,η)(\alpha,\eta) in the submission; previously only SrS_r was recalibrated. We now transfer the GSM8K-Aug subspace unchanged:

Fixed GSM8K-Aug SrS_rTILR vs. backbonevs. AdaAnchor
SVAMP+1.7 pts, p=0.014p=0.014p=0.018p=0.018
GSM-Plus+0.8 pts, p=0.0004p=0.0004p=0.001p=0.001

Thus, the learned directions remain useful outside their calibration domain without recalibration.

W2a. What about other latent-reasoning methods such as CODI?

TILR is an inference-time correction applied to a trained latent reasoner, whereas CODI/SoftCoT modify how that reasoner is trained. We therefore tested whether TILR transfers to CODI-GPT-2 (Shen et al., EMNLP 2025), trained using the official recipe:

TILR −- CODI accuracyGSM8KSVAMPGSM-Plus
Backbone below reference+2.0, p=.013p=.013+4.4, p<.001p<.001+0.8, p=.0004p=.0004
Backbone near reference+0.5, p=.036p=.036+1.0, p=.031p=.031+0.2, p=.044p=.044

TILR improves CODI significantly where the reference provides headroom and never significantly hurts it when the backbone approaches reference performance. More importantly, the robustness mechanism transfers: relative to unconstrained refinement, TILR reduces GSM8K paraphrase trajectory variance from 3011 to 1562 (-48%) and from 1440 to 770 (-47%) across the two CODI backbones. Refinement's divergence grows over the latent steps, while TILR's stays flat. We will add CODI to the paper and discuss all three methods in the related work.

These results show that TILR is neither tied to a single reference pair, calibration task, nor Coconut training procedure: the learned subspace transfers across references and tasks, while the refinement mechanism transfers to a separately developed latent-reasoning framework.

AuthorsOpenReview ↗

Response to Reviewer mEL1: Addressing Gate, Scale, Task Scope, Update-rule concerns

W2b / Q2. Does the gate contribute beyond projection?

Table 5 shows that projection provides most mean-accuracy gain under the standard reference. The gate instead suppresses unreliable corrections. We test this under controlled reference degradation at fixed η\eta:

ReferenceProjection onlyTILRHarm removedγ\gamma
Paper's reference+0.99+1.06no cost0.79
Under-trained−1.44-1.44−0.23-0.2384%--
+ noise 1×1\times−2.12-2.12−0.68-0.6868%0.50
+ noise 3×3\times−13.95-13.95−5.53-5.5360%0.46

Under corruption, γ\gamma approaches the isotropic rank-171 baseline 171/768=0.47\sqrt{171/768}=0.47, suppressing the update. This is not merely a smaller-step effect: at matched average correction magnitude, 3×3\times noise causes a 4.8-point loss with gating vs. 12.7 without; other matched-size ungated corrections lose 11.811.8--20.020.0 points.

Without induced corruption, gating reduces paraphrase trajectory variance by 4% (262.1 →\rightarrow 252.2), accounting for 36% of TILR's total reduction over an equal-size unprojected correction (279.8 →\rightarrow 252.2). TILR also has the lowest variance among all equal-size corrections tested. Its hyperparameter-sensitivity gap is 10--18$\times$ smaller than Refinement's on GSM8K, GSM-Plus, and SVAMP (Table 4), consistent with Appendix D. Thus, the gate's contribution is robustness rather than additional mean accuracy under a reliable reference.

W3 / Q3. Scale and task scope

The manuscript already evaluates GPT-2, Qwen2.5-Math-1.5B, and Llama3.2-1B (Appendix H), with TILR improving accuracy, M3, and M4 across all six larger-backbone/benchmark combinations. It also includes multi-hop deduction (ProsQA: 90.4%; ProntoQA: 94.7%) and perturbation benchmarks GSM-Plus and GSM-Symbolic.

New Qwen2.5-7B results:

Qwen2.5-7BGSM8K Acc.SVAMP Acc.GSM8K M4 ↓\downarrow
Coconut33.51%69.7%272.8
+ TILR33.66%69.9%243.8 (-11%)

The strongest available 7B reference is only 0.7 points above the backbone on validation, leaving limited corrective headroom. TILR therefore preserves accuracy while reducing paraphrase trajectory variance by 11%. Together with CODI (Part 1), TILR now spans two latent-reasoning frameworks and four base-model settings. MoE, MATH500, AIME, and code require latent reasoners trained for those settings; we will identify these as future work.

W4 / Q4. Which update rule was used, why, and does it matter?

Every reported result uses ηt=min⁡(ηγt2,1)\eta_t=\min(\eta\gamma_t^2,1) with fixed α\alpha, as stated in Eq. 8, Section 4.2, and Algorithm 1. Fixed α\alpha recovers the unmodified backbone as γ→0\gamma\rightarrow0; γ2\gamma^2 more strongly suppresses low-alignment corrections and yields Appendix D's sensitivity bound; the cap limits the effective step. Figure 2 is an earlier version and will be corrected in the revised manuscript.

We evaluated all eight combinations of fixed/gated α\alpha, γ/γ2\gamma/\gamma^2, and capped/uncapped updates on GSM8K, SVAMP, and GSM-Plus at the same (α,η)(\alpha,\eta). Every variant is within 0.45 accuracy points, every paired 95% CI contains zero, and every variant remains above the backbone. Figure 2's variant differs by only −0.45/+0.20/−0.01-0.45/+0.20/-0.01 points. Thus, the discrepancy is just a presentational error.

Summary

  • Subspace: filters paraphrase/reference variation more strongly than PCA/random and transfers across references/tasks.
  • Generality: TILR transfers to CODI and reduces M4 by 47--48% over Refinement.
  • Gate: removes 60--84% of projection-only harm; at matched magnitude, 3×3\times noise loss falls from 12.7 to 4.8 points.
  • Scale: at 7B, accuracy is preserved while M4 decreases 11%.
  • Update rule: all results use Eq. 8/Algorithm 1; the eight-variant ablation confirms conclusions are insensitive to Figure 2's presentational error.

We hope these results address the soundness concerns and would appreciate reconsideration of the initial assessment in light of this evidence.

AuthorsOpenReview ↗

Follow-up on Our Response, Clarifications and Additional Results

Dear Reviewer,

Thank you again for your review. We have provided additional experiments and clarifications in our rebuttal addressing the questions raised in your assessment and have incorporated these updates into the revised manuscript as well.

We would appreciate it if you could revisit our response when you have a chance. We would be happy to address any remaining concerns during the discussion period.

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

Latent reasoning refinement can be sensitive to hyperparameters, reference checkpoints, and semantically equivalent reformulations of an input. The paper proposes TILR, which projects contrastive updates into a calibrated subspace and gates their magnitude according to subspace alignment. Experiments on arithmetic and deductive reasoning benchmarks report improved latent-method accuracy and more consistent answers and trajectories under input perturbations, with supporting results on additional backbones.

Critical strengths
  1. The evaluation considers reference sensitivity and paraphrase consistency alongside average accuracy.

  2. Random-subspace, PCA, knockout, and component controls provide useful tests of the refinement mechanism.

Critical weaknesses
  1. SVD identifies dominant variation in the reference models' differences, which may reflect task-specific patterns rather than invariant reasoning structure. The current evidence supports useful filtering more clearly than the stronger invariance interpretation.

  2. The full method does not consistently outperform the strongest subspace controls, and projection alone nearly matches it in the component ablation. The additional benefits of the chosen subspace and adaptive gate are therefore not fully isolated.

  3. The subspace is recalibrated for each benchmark and reference pair, with most evaluation centered on small, task-trained models. The additional backbones are helpful, but transfer of a fixed subspace across reference changes or domains remains untested.

Questions for authors

If the calibration subspace is kept fixed while the reference pair changes, how do accuracy and paraphrase consistency change?

2 comments
AuthorsOpenReview ↗

Response to Reviewer dtUJ: Experimental Validation of Subspace Filtering, Gate Effects, and Cross-Reference Transfer

Thank you for the careful review, and for noting that the evaluation goes beyond average accuracy and that the controls test the mechanism. We conducted experiments with the GPT-2 Coconut backbone on GSM8K-Aug to address your concerns.

W1. Filtering vs. invariance. As stated in Appendix B, we make no formal claim that SrS_r contains causally invariant features. Operational invariance is measured directly through reference sensitivity (RQ2) and semantic reformulations (RQ3). We now test the mechanism directly: on 920 paraphrased GSM8K questions, we measure how much perturbation-induced variation each subspace admits (lower filters more):

Subspace (rank 171)Paraphrase changeReference swap
Calibrated (TILR)0.8620.776
Input PCA0.9770.888
Random1.0021.003

The calibrated subspace filters both sources of variation substantially more strongly than input PCA or a rank-matched random subspace. On task specificity, the GSM8K-Aug subspace also transfers unchanged to SVAMP and GSM-Plus (W3). We will state this operational definition of invariance explicitly in the camera-ready revision.

W2. Isolating the subspace and the gate. To separate direction selection from update magnitude, we matched the mean ℓ2\ell_2-norm of the applied correction across conditions and compared 920 questions ×\times 5 phrasings:

Correction, equal average step sizeTrajectory variance under paraphrasing (↓\downarrow)All five phrasings correct
TILR (calibrated + gate)252.212.0%
Calibrated, no gate262.111.8%
Input PCA, no gate269.211.7%
Unprojected, no gate279.811.1%
Random, no gate281.610.0%
  • Subspace isolated: at matched correction magnitude, the calibrated directions yield lower trajectory variance than input-PCA, unprojected, or random directions, and the highest all-five-correct rate (12.0% vs. 11.1% unprojected and 10.0% random).
  • Gate isolated: on the same calibrated subspace, gating preserves all-five-correct performance (11.8% →\rightarrow 12.0%) while reducing trajectory variance by 3.8% (262.1 →\rightarrow 252.2). Under 3×3\times reference noise, it reduces accuracy loss from 12.7 to 4.8 points (62% less degradation).

These controlled results clearly isolate the roles of the two components: the learned subspace selects robust correction directions, while the gate suppresses unreliable corrections.

W3 / Q. A fixed subspace across references and domains. We kept the paper's learned subspace fixed and replaced the good reference with seven alternatives, including three checkpoints from a separately trained model (GSM8K test; all-five-correct measured on the 920 paraphrased questions):

Good referenceAccuracyAll five phrasings correct
Coconut backbone (none)26.61%8.9%
TILR Original (epoch 22)27.67%12.0%
Epoch 2628.20%10.9%
Epoch 3027.98%10.3%
Epoch 4627.22%10.8%
Seed 2, epoch 1428.20%9.9%
Seed 2, epoch 3427.29%9.3%
Seed 2, epoch 5026.84%9.8%

With every reference, both accuracy and the share of questions answered correctly under all five phrasings remain above the Coconut backbone. The fixed subspace averages 27.63%27.63\% accuracy, compared with 27.85%27.85\% when recalibrated for each reference, a difference of only 0.220.22 points.

Across domains, we apply the GSM8K-Aug subspace unchanged to SVAMP and GSM-Plus. Under the corresponding transfer evaluation, it yields gains of +1.7+1.7 points on SVAMP (p=0.014p=0.014) and +0.8+0.8 points on GSM-Plus (p=0.0004p=0.0004). Thus, the learned subspace remains useful outside the calibration domain without recalibration. Also, Appendix H reports Qwen2.5-Math-1.5B and Llama-3.2-1B; on a new Qwen2.5-7B run, TILR maintains comparable accuracy (GSM8K 33.66% vs 33.51%) and reduces trajectory variance under paraphrasing by 11% relative to the backbone.

In summary, we provide direct evidence addressing each concern:

  • W1: the claim is operational invariance to input and reference perturbations; the calibrated subspace filters both substantially more strongly than input PCA or a rank-matched random subspace.
  • W2: equal-step controls isolate the contributions of the learned subspace and gate: the calibrated directions reduce trajectory variance, and gating provides additional stabilization while substantially limiting damage from corrupted references.
  • W3 / Q: one fixed subspace remains effective across seven reference checkpoints, differs by only 0.220.22 accuracy points from per-reference recalibration, and transfers without recalibration to SVAMP and GSM-Plus.

We hope these additional results address the reviewer's concerns and clarify the robustness and transferability of the learned refinement subspace.

AuthorsOpenReview ↗

Follow-up on Our Response and Additional Results

Dear Reviewer,

Thank you again for your review. We have provided additional experiments in our rebuttal that directly address all your questions and have incorporated these results into the revised manuscript.

We would appreciate it if you could revisit our response when you have a chance. We would be happy to address any remaining concerns during the discussion period.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Review Discussion and Complete Experimental Evidence

Dear Area Chair, Senior Area Chair, and Program Chairs,

We thank the reviewers for their feedback and discussion. The questions raised during review and subsequent discussion are now directly addressed through the analyses in the manuscript and additional experiments, with consistent results across references, benchmarks, architectures, model scale, and latent-reasoning frameworks. These additional results and clarifications are incorporated into the revised manuscript, and the corresponding experimental scripts are uploaded for reproducibility. We summarize the complete evidence below.

Question evaluated during review/discussionEvidence and resolution
Is the robustness effect specific to the learned subspace?Direct perturbation tests show that SrS_r filters substantially more paraphrase/reference variation (0.862/0.776) than rank-matched input PCA (0.977/0.888) or random subspaces (1.002/1.003). Under equal-magnitude corrections, TILR achieves the lowest trajectory variance (252.2 vs. 262.1-281.6), ruling out generic dimensionality reduction or smaller updates as explanations.
Is the invariance effect directly supported and transferable?RQ2 reduces reference-pair sensitivity by 53-73%; RQ3 improves answer agreement by 11.4% and reduces trajectory variance relative to Refinement on every evaluated benchmark. A fixed SrS_r further transfers across 7 references, including an independent training seed, with only a 0.22 pp gap from recalibration, and transfers without recalibration to SVAMP (+1.7 pp, p=.014p=.014) and GSM-Plus (+0.8 pp, p=.0004p=.0004). Thus, direct invariance evidence and cross-benchmark transfer are independently established.
Does adaptive gating provide a distinct robustness benefit and behave appropriately under failure?Under progressive reference degradation, γ\gamma decreases from 0.79 to 0.46, and gating removes 60-84% of projection-induced degradation. Under matched correction magnitude, 3×3\times noise causes only a 4.8 pp loss with gating versus 12.7 pp without, directly characterizing its role when the correction becomes unreliable.
Are the comparisons and generalization sufficiently broad?The evaluation includes Coconut (2024), Refinement (2026), AdaAnchor (2026), and CoT, plus PCA/random/knockout/gating/equal-magnitude controls that test the principal alternative explanations. The main evaluation spans six mathematical and deductive benchmarks, including ProsQA and ProntoQA. Additional results extend across GPT-2, Llama3.2-1B, Qwen2.5-1.5B, Qwen2.5-7B, and CODI, where TILR reduces trajectory variance by 47-48% while retaining/improving accuracy.
Does the robustness effect persist at larger scale?On Qwen2.5-7B, the strongest reference provides only 0.7 pp validation headroom. TILR nevertheless preserves accuracy (33.51% →\rightarrow 33.66%) while reducing paraphrase trajectory variance by 11%, directly validating the stated robustness-without-accuracy-loss objective at larger scale.
Are the conclusions sensitive to update design or computational cost?Across 8 update-rule variants, all remain within 0.45 accuracy points, all paired 95% CIs include zero, and all remain above the corresponding backbone. Refinement and TILR use the same 24.3 forward passes and require 67.9 vs. 68.7 ms/question, so TILR adds only 1.2% incremental latency; calibration is a one-time ~10 s cost.

Importantly, the paper's invariance claim is explicitly scoped throughout: invariance is operational and directly evaluated through stability across reference choices and semantically equivalent inputs, without making a causal-invariance claim.

Taken together, the complete evidence directly addresses the substantive questions raised throughout review and discussion. The robustness effect survives controls for generic dimensionality reduction and update magnitude, transfers across references, an independent training seed, benchmarks, architectures up to 7B, and CODI, and remains robust across update-rule choices. Controlled reference degradation further establishes the distinct role of adaptive gating.

These results consistently support the paper's central claim: TILR introduces a geometric approach to robust latent refinement that improves stability to reference choice, input reformulation, and hyperparameter variation while preserving reasoning performance, without retraining. The revised manuscript carefully scopes this claim to operational invariance and explicitly separates broader extensions from what is required to establish the contribution here.

We respectfully ask that the final assessment consider the complete evidence now available in the manuscript, author responses, revision, and experimental results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.