acceptodds
Under review as a conference paper at ICLR 2027

Project Before You Distill: Carrying Privileged Chain-of-Thought Across Sensor Gap

Abstract

Large vision–language models provide rich reasoning priors, but physical sensors often lack large paired corpora and intermediate reasoning supervision. Directly distilling a rationale generated from a privileged sensor can introduce claims that are not supported by the deployment sensor. To address this gap, this paper introduces Cross-sensor Rationale Projection (CRP) which treats the privileged rationale as a reasoning proposal and re-derives it against the deployment observation. CRP keeps supported information, abstracts fine details into sensor-supported structure, hedges partially supported claims, and removes unsupported content, followed by a consistency screen. We further analyze information loss across the sensor gap and derive an Estimated Projection Value (EPV), computable even before training the deployment student model, to indicate when projection helps with consistent empirical support. Two purpose-built sensor reasoning benchmarks, RadarQA-300 and AudioQA-300, are constructed for RGB-to-radar and audio-visual-to-audio reasoning. CRP consistently outperforms answer-only supervision, direct privileged-rationale distillation, and deployment-sensor rationale supervision, with gains of and percentage points over answer-only training on RadarQA-300 and AudioQA-300, respectively.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
sensor-grounded chain-of-thought, reasoning, rationale projection, physical-sensor reasoning, radar reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_TPxROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Cross-sensor Rationale Projection (CRP) to address the challenge of directly transferring inference processes generated by privileged sensors to information-constrained deployment sensors. Specifically, CRP first generates an inference chain using information-rich sensors, then transforms it into inference content supported by the target sensor through four operations: Keep, Abstract, Hedge, and Remove. Consistency filtering is then used to construct training supervision for the student model. Furthermore, the authors propose Estimated Projection Value (EPV) to predict the potential gains of cross-sensor inference projection. Experiments are conducted on RadarQA-300 and AudioQA-300 benchmarks. Results show that CRP outperforms answer-only supervision and surpasses direct distillation methods across multiple models and fine-tuning settings.

Critical strengths
  1. This paper points out the differences in information observability in cross-sensor CoT distillation and proposes a solution to the problem of unreliable supervision that may be introduced by direct transfer privileged inference.
  2. Experiments cover different sensor scenarios.
Critical weaknesses
  1. The four inference projection operations lack fine-grained ablation. The paper does not adequately validate the independent contributions of Keep, Abstract, Hedge, and Remove, nor the necessity of using them in combination.
  2. The implementation details of the core projection process are insufficient.
  3. Potential confounding factors remain in the comparison with baseline methods. Although the student model training configuration is consistent, it is unclear whether the number of teacher calls, inference length, selection strategies, and number of effective training samples are consistent across different supervision methods.
  4. There is a lack of direct quantitative evaluation of the consistency of inference evidence. The experiments primarily focus on the accuracy of the final answer and have not yet sufficiently demonstrated that Projected-CoT actually reduces inference statements that the target sensor cannot support.
  5. The experimental scale is limited, and the computational cost and scalability of the method in practical deployment are unclear.
Questions for authors
  1. Is it recommended to ablate the Keep, Abstract, Hedge, and Remove operations to verify their individual contributions and the necessity of their combined use?
  2. Provide more experimental details. Include teacher model configuration, projection decision strategy, and a complete Prompt, and clarify whether the same teacher model is used in each phase.
  3. It is recommended to re-compare CRP with the existing baseline under conditions of matched teacher call budget, CoT length, screening strategy, and training sample size.
  4. Quantitatively analyze inference reliability. Quantitatively compare the quality of CoTs generated by different supervision methods.
  5. More detailed computational efficiency should be reported.
5 comments

We thank Reviewer TPxR for recognizing the sensor-observability problem and the evaluations across radar and audio. We address CRP projection steps (W1), implementation (W2), baseline comparability (W3), claim support (W4), and cost (W5). Revisions are highlighted in blue in the revised paper.

W1 / Q1. Ablation of CRP Projection Steps.

We ablate CRP steps and deletion methods (motivated by Q4f5) on RadarQA-300 with Qwen3-VL-8B LoRA, three seeds (Section 7.4, new Table 7 in Appendix I).

CRP projection steps / deletion methodsAccuracy (%)
All four steps69.6 ± 1.0
Without ABSTRACT47.7 ± 2.6
Without REMOVE57.7 ± 3.1
KEEP + REMOVE55.2 ± 1.2
KEEP-only45.0 ± 2.1
Delete unsupported claim words45.7 ± 5.0
Random sentence deletion53.2 ± 3.2

ABSTRACT contributes most. Omitting it costs 21.9 points, versus 11.9 for REMOVE. Deletion alone is insufficient. KEEP + REMOVE trails all four steps by 14.4 points. Explicit HEDGE adds 0.8 points over the original prompt, which already permits hedging (Appendix G.3).

W2 / Q2. Teacher Models and Projection Implementation.

The projector PdP_d is a prompted LLM that receives the deployment observation, question, ground-truth answer, and privileged rationale, but not the privileged observation. It applies KEEP, ABSTRACT, HEDGE, or REMOVE to each claim. For 1,101 radar examples, Claude Sonnet 5 writes the privileged rationale, projected rationale, and Deployment-CoT. Claude Opus handles the other 207 and selected revisions. Gemini 3.6 Flash writes the audio rationales. Section 4 and new Table B.1 (Appendix B.1) clarify the procedure and teacher assignments. The full radar projection prompt is provided in new Appendix G.1, with other prompts in Appendix G.2.

W3 / Q3. Teacher Budget, Rationale Length, Screening, and Training Samples.

We clarify differences in teacher calls, rationale length, selection procedures, and effective training samples (Table 1, new Table B.1 in Appendix B.1).

Original radar comparisonProjected-CoTDeployment-CoT
Training samples1,3081,308
Teacher calls per example (1,101 Sonnet examples)24.02
Additional model-based screenNoYes
Median rationale length (characters)1,1801,597
Accuracy, Qwen3.5-4B LoRA (%)73.8 ± 1.870.7 ± 3.2

The original Deployment-CoT uses an independent radar-only interpretation and Projected-CoT as references. Both methods undergo consistency checks, which drop 566 candidates. An additional model-based screen rejects 47 of 1,625 Deployment-CoT candidates. Rejected candidates are excluded, while selected rationales with violations are edited or rewritten. Both methods retain the same 1,308 training examples.

We also evaluate three Deployment-CoT variants (three seeds):

ComparisonStudentDeployment-CoT (%)Projected-CoT (%)
Clean radar, one teacher call, no additional screenQwen3-VL-8B, full67.4 ± 1.372.4 ± 1.0
Matched rationale lengthQwen3-VL-8B, LoRA64.7 ± 2.868.8 ± 0.5
Matched input notes (audio)Qwen2.5-Omni-7B, LoRA63.8 ± 1.268.9 ± 1.8

Projected-CoT leads in all comparisons (Section 7.1, Appendix B.1, Appendix J). The clean radar baseline uses one teacher call versus two for Projected-CoT.

W4 / Q4. Independent Evaluation of Claim Support.

In the new Appendix F (Tables 4 and 5), an independent GPT-6.1 evaluator assesses 400 radar training examples across the three supervision methods without seeing the answer, RGB, or method identity. Four human annotators separately assess sampled non-appearance claims. Appearance claims are counted as unsupported by radar.

Radar-supported claimsProjected-CoTDeployment-CoTPrivileged-CoT
Independent model84.6%81.7%21.0%
Human-assisted79.3%—28.1%

Projected-CoT has the highest supported-claim share. Human agreement is 88.1% for paired yes/no judgments. Human-assisted rates include automatically flagged appearance claims as unsupported.

W5 / Q5. Computational Cost and Scalability.

For radar, projection adds one teacher call per example after privileged-rationale generation and no additional work during student training or deployment. Each projection averages 1,803 input tokens, 396 output tokens, and 5.2 seconds. Projecting 1,308 examples takes about 14 minutes with eight parallel requests.

For Qwen3.5-4B, Projected-CoT uses 1.16M supervision tokens per run, compared with 1.50M for Deployment-CoT. Estimated five-run SFT costs are $4.29 for Projected-CoT, compared with $5.52 for Deployment-CoT (new Table B.1, Appendix B.1).

Projection cost grows linearly with training-set size. Evaluation covers two sensor gaps and five students, while larger-scale validation remains future work.

I appreciate the authors' detailed response and supplementary experiments, which have clarified some of the questions. However, the existing experiments still do not adequately eliminate potential confounding factors in baseline comparisons, and the verification of the reliability of the inference is relatively limited. At the same time, the independent methodological contributions of the overall framework are not clearly defined. The method mainly combines teacher inference generation, text rewriting under target modality constraints, and supervised fine-tuning. Based on the above issues and other reviewers' concerns, I will maintain the original score.

AuthorsOpenReview ↗

Round 2 Response: Part I

We thank Reviewer TPxR for clarifying the remaining concerns. We have conducted additional evaluations of radar support in reasoning rationale during inference (in addition to the support claim evaluation during training, reported in our earlier response) and comparisons between our CRP (claim-based projection) and generic rewriting.

Please also see the revised PDF of the main paper, including updated appendices, for more detailed new results and analyses, with changes highlighted in blue.

1. Baseline comparability. We understand your concern that differences in teacher-call budget, rationale length, screening, and effective training samples may contribute to the observed gains. To address this, the revised paper includes additional comparisons that match specific factors and examine alternative rationale-generation procedures.

The radar comparisons below use the same 1,308 training questions and answers and the RadarQA-300 evaluation, with the student model and training settings fixed within each comparison. Results are averaged over three seeds.

ComparisonStudentDeployment-CoT accuracy (%) ↑Projected-CoT accuracy (%) ↑
Clean radar-only generationQwen3-VL-8B, full fine-tuning67.4 ± 1.372.4 ± 1.0
Matched rationale lengthQwen3-VL-8B, LoRA64.7 ± 2.868.8 ± 0.5

see Table 1 and Section 7.1 and Table 8, Appendix J for more details.

Projected-CoT achieves higher accuracy in both comparisons. These experiments address selected sources of confounding, although teacher-call counts and generation/editing procedures are not jointly matched.

We have now also compared CRP with generic rewriting using the same privileged rationales and two teacher calls per example, with additional rationale editing omitted from both methods. This comparison more directly examines the contribution of the projection procedure, as detailed in point 4 of Part II below.

2. Reliability of reasoning. We understand this concern as whether Projected-CoT reduces reasoning statements that the deployment sensor cannot support. Appendix F of the revised paper provides an independent model evaluation of training rationales (Table 4) and a human-assisted audit (Table 5), as reported in our earlier response. We now extend this evaluation to the students’ own rationales during inference to assess whether the learned reasoning is also supported by radar measurements.

Human-assisted audit of training rationales. Table 5 of the revised paper reports the following radar-supported claim rates:

Training rationalesRadar-supported claims (%) ↑
Privileged-CoT28.1
Projected-CoT79.3

These estimates combine automatic identification of unsupported appearance claims with human judgments on 22 sampled non-appearance claims from Projected-CoT and 32 from Privileged-CoT. Four annotators provide two judgments per claim. Human–human agreement is 88.1%, and agreement between humans and the automated claim evaluator is 83.2%.

New evaluation at inference. We examine the students’ own test rationales on the 259 of 300 test questions with checkable radar point clouds. We define a metric of decisive claim, the statement in the student’s reasoning rationale that most directly supports its answer. A claim is considered to be true when both model judges agree that all checkable statements match the radar measurements.

SupervisionDecisive claims judged true (%) ↑
Clean radar-only Deployment-CoT49.7
Projected-CoT56.7

The two independent model judges agree on 96.3% of decisive-claim assessments. These evaluations provide evidence of radar support in both training and inference rationales generated by the student deployment model.

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

This paper proposes Cross-sensor Rationale Projection (CRP), a framework for transferring reasoning from a privileged sensor to a model that operates using a less informative sensor. Instead of directly distilling the teacher's chain-of-thought, CRP modifies its claims through four operations: keep, abstract, hedge, and remove. A consistency screening step then checks the projected rationale before using it to train the student.

The paper also introduces Estimated Projection Value (EPV), which aims to predict when rationale projection will be beneficial before student training. Experiments on RadarQA-300 and AudioQA-300 show improvements of 5.9 and 12.9 percentage points over answer-only training, respectively. Additional experiments evaluate different models, fine-tuning strategies, and sensor representations.

Critical strengths

1- The paper addresses an important issue in cross-modal reasoning: privileged observations can provide useful reasoning supervision, but some of that information may not be recoverable from the deployment sensor. The proposed projection approach is intuitive and provides a reasonable way to reduce unsupported claims. I also find the idea of estimating when projection is beneficial before training interesting.

2- The authors evaluate the method on two different sensor modalities and several model architectures. CRP consistently improves over Deployment-CoT in the reported experiments, with particularly meaningful gains on RadarQA-300 and AudioQA-300. The supplementary material also provides additional results across fine-tuning configurations and sensor representations, as well as a comparison with GRPO. These experiments provide useful initial evidence for the proposed approach.

Critical weaknesses

1- Although the reported improvements are encouraging, the paper does not sufficiently establish which components are responsible for them. In particular, there is no clear ablation separating the contributions of claim projection, the four projection operations, and consistency screening. The supplementary experiments mainly compare different models and training configurations rather than these individual components. More importantly, there is no quantitative evaluation of whether the projected rationales are actually supported by the deployment sensor. Since the ground-truth answer is provided during rationale generation and screening, it is important to verify that the resulting explanations are genuinely grounded in sensor evidence rather than simply constructed to support the known answer.

2- The supplementary material provides the proof of Proposition 1, but the analysis relies on a simplified scalar regression model and assumes that rationale information loss is directly related to prediction error. It is unclear how well these assumptions represent actual language-model reasoning and fine-tuning. The proposed EPV also relies on estimated claim-level information loss, but the accuracy and reliability of these estimates are not independently validated. Furthermore, the experiments use relatively small, purpose-built benchmarks, and some improvements are modest, particularly the 0.8-point gain for Inkling-Small. The GRPO comparison is also limited to one seed with different training budgets.

3- Some methodological explanations are unnecessarily dense, and Section 4 contains an unresolved reference to "Appendix ??", even though supplementary material is provided. There also appears to be an inconsistency in the InternVL3.5-8B result in Figure 5(a), which shows a 1.1-point improvement, while the main text and supplementary Table 2 report 4.1 points. These issues should be corrected to improve clarity and confidence in the reported results.

Questions for authors

1- Can the authors provide separate ablations for the four projection operations and the consistency screening step? It would also be useful to report how many rationales are rejected by screening and how rejected examples are handled during training.

2- How do the authors verify that the projected rationales are genuinely supported by the deployment sensor? Can they provide an independent evaluation of claim-level correctness and grounding, preferably using human annotations or an independent evaluator?

3- How are the answer-relevant claims and their information-loss estimates determined in practice? Have the authors evaluated the sensitivity of EPV to estimation errors and its predictive performance on independent, held-out settings? This would help establish whether EPV is useful beyond the simplified theoretical model.

4- Can the authors compare CRP against stronger rationale-filtering or evidence-grounding baselines, such as approaches related to the recoverable-privilege methods discussed in the related work? Also, can they clarify the discrepancy between Figure 5(a) and supplementary Table 2 for InternVL3.5-8B, and correct the unresolved "Appendix ??" reference?

2 comments

We thank Reviewer joUM for recognizing the sensor-gap problem, projection, and EPV. We address projection and screening (W1), claim support (W1), EPV (W2), baselines and GRPO (W2), and corrections (W3). Revisions are highlighted in blue in the revised paper.

W1 / Q1. Ablation of CRP Projection Steps and Screening.

We ablate CRP steps and deletion methods (motivated by Q4f5) on RadarQA-300 with Qwen3-VL-8B LoRA, three seeds (Section 7.4, new Table 7 in Appendix I).

CRP projection steps / deletion methodsAccuracy (%)
All four steps69.6 ± 1.0
Without ABSTRACT47.7 ± 2.6
Without REMOVE57.7 ± 3.1
KEEP + REMOVE55.2 ± 1.2
KEEP-only45.0 ± 2.1
Delete unsupported claim words45.7 ± 5.0
Random sentence deletion53.2 ± 3.2

ABSTRACT contributes most. Omitting it costs 21.9 points, versus 11.9 for REMOVE. Deletion alone is insufficient. KEEP + REMOVE trails all four steps by 14.4 points.

The screen checks that rationales end with the ground-truth answer and meet wording and length rules, dropping 566 candidates. An additional model-based screen rejects 47 of 1,625 Deployment-CoT candidates. Dropped candidates are excluded from training, while selected rationales with violations are edited or rewritten (Section 4, new Table B.1 in Appendix B.1).

W1 / Q2. Independent Evaluation of Claim Support.

In the new Appendix F (Tables 4 and 5), an independent GPT-6.1 evaluator assesses 400 radar training examples across the three supervision methods without seeing the answer, RGB, or method identity. Four human annotators separately assess sampled non-appearance claims. Appearance claims are counted as unsupported by radar.

Radar-supported claimsProjected-CoTDeployment-CoTPrivileged-CoT
Independent model84.6%81.7%21.0%
Human-assisted79.3%—28.1%

Projected-CoT has the highest supported-claim share. Human agreement is 88.1% for paired yes/no judgments. Human-assisted rates include automatically flagged appearance claims as unsupported.

W2 / Q3. EPV Theory, Estimation, and Reliability.

The new Appendix H (Theorem 1) extends Proposition 1 from a scalar model to a multicomponent model. Under matched supervision noise, the difference in expected estimation error depends on the information lost from each component and its contribution to the answer. When all components have the same information-loss fraction, the result reduces to Proposition 1.

Frozen judges score claims as supported (0), partial (0.5) or unsupported (1), estimating EPV\mathrm{EPV} without training students (Appendix A, new Appendix G.4).

For all 300 radar test questions, full EPV from recovered RGB correlates with the sensor-gap estimate δ^2\hat\delta^2 used in Figure 4(a) (Pearson r=0.980r=0.980). Both remain indicative of the observed accuracy advantage of Projected-CoT over Deployment-CoT (new Table A.1):

Full EPV bandSamplesQwen3.5-4BQwen3-VL-8B
< 064−1.04−1.04
0–0.189+2.25+3.00
0.1–0.479+4.22+6.33
> 0.468+7.84+9.31

In 1,000 judge-error simulations, we perturb claim-level scores based on two-judge disagreements, recompute sensor-gap or EPV groups, and keep student predictions fixed (new Table A.2):

Group or comparisonMedian margin
Radar: δ^i2≤0.01\hat\delta_i^2 \le 0.010.00
Radar: 0.01<δ^i2≤0.10.01 < \hat\delta_i^2 \le 0.1+2.49
Radar: 0.1<δ^i2≤0.40.1 < \hat\delta_i^2 \le 0.4+3.45
Radar: δ^i2>0.4\hat\delta_i^2 > 0.4+7.14
Audio: EPV > 0.1 vs. ≤ 0.1+6.92

These results suggest that sensor-gap and EPV estimates remain indicative of projection benefits without student training, despite judge errors. Theorem 2 (new Appendix H) shows that judge errors can weaken the difference between EPV groups but do not reverse which group benefits more on average, under its assumptions. The same fixed sensor-gap bands also generalize to a second student model, Qwen3-VL-8B (−6.6, +4.0, +5.7, +12.7 points, new Table A.3).

W2 / Q4. Benchmark Scale, GRPO, and Baselines.

Projected-CoT leads in seven settings across two 300-question benchmarks, though Inkling-Small gains only 0.8 points (Tables 1 and 2). For Inkling-Small, gains are −2.6 on natural and +8.3 on EPV-selected questions (Figure 4(b)). We add a new three-seed GRPO experiment, matching SFT epochs and optimizer steps. GRPO reaches 53.1 ± 1.3% versus 72.4 ± 1.0% for Projected-CoT SFT (updated Appendix E).

We also add a C2KD-style baseline. On Qwen3-VL-8B LoRA, it reaches 51.0 ± 2.0% versus 68.8 ± 0.5% for Projected-CoT (new Table 8, Appendix J).

W3 / Q4. Corrections.

Figure 5 corrects InternVL3.5-8B (+4.1 in (a)) versus radar heatmaps (+1.1 in (b)). "Appendix ??" is fixed to Appendix G. Section 4 is shortened.

Thank you again for your helpful feedback. We have provided additional analyses and clarifications in the rebuttal and revised manuscript (please check the revised pdf for the main paper with highlighted changes in blue), with additional details in the appendices. Could you let us know whether these additions address your concerns? If any specific issues remain, we would appreciate your guidance so that we can focus our remaining efforts on them.

Reviewer_Q4f5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the sensor-gap problem in chain-of-thought supervision: a vision-language teacher has access to a privileged observation during training, but the student deploys on a different sensor, and directly distilling the teacher's rationale leaks details the deployment sensor cannot observe. The proposed Cross-sensor Rationale Projection (CRP / Projected-CoT) decomposes the privileged rationale into atomic claims and re-adjudicates each against the deployment observation with four operations — KEEP, ABSTRACT, HEDGE, REMOVE — followed by a consistency screen and standard SFT. Two purpose-built benchmarks (RadarQA-300, AudioQA-300) are evaluated across five student backbones and matched/mismatched sensor representations.

Critical strengths
  1. The problem is real and well-motivated: privileged CoT supervision contains sensor-unobservable claims (clothing, color, layout invisible to radar), and the fact that Privileged-CoT is worse than Answer-only on RadarQA-300 is a concrete motivation.
  2. The Propose → Project → Screen mechanism is clean and interpretable; the claim-level operations map naturally onto what a deployment sensor can and cannot support, and the worked example of a radar rationale being re-derived from point clouds is genuinely instructive.
  3. The EPV framing is practical — it estimates before student training where projection should help — and the empirical check (advantage growing monotonically with estimated sensor-gap loss, flipping sign at the predicted threshold) is the most convincing analysis in the paper, with a consistent advantage sign across two sensor gaps, five backbones, and two fine-tuning regimes.
Critical weaknesses
  1. The entire method hinges on the projection function P_d and the screener V, but the main text never states what model performs the claim-level KEEP/ABSTRACT/HEDGE/REMOVE: a prompted LLM given the deployment observation, a fine-tuned judge, or human annotation, and what its per-claim accuracy is. If P_d is an oracle with privileged access the method collapses to supervised filtering; if it is a weak prompted model, its own errors confound every result.
  2. The paper explicitly differentiates itself from C2KD and PULSE in Related Work, but the main results compare only against three strawman variants (Answer-only, Privileged-CoT, Deployment-CoT); without C2KD numbers, the claim that CRP differs meaningfully from existing cross-modal distillation is asserted rather than demonstrated.
  3. There is no ablation of the projection mechanism itself: ablations vary backbones, fine-tuning regimes, and sensor representations, but never isolate which operation carries the gain. A KEEP-only, REMOVE-only, or random-claim-masking control is needed to show that claim-level projection — not merely filtering some claims — drives the improvement.
  4. The Deployment-CoT baseline itself is opaque: it requires a teacher that emits rationales from the deployment sensor alone, and the paper does not state how those deployment-side rationales are produced or how good they are. Since their quality sets the bar CRP must clear, a weak radar/audio rationale generator could make CRP's win partly a weak-baseline artifact.
  5. The abstract leads with gains over Answer-only (5.9 / 12.9 points), while the more relevant margin over Deployment-CoT is smaller and variable (+0.8 to +6.6).
Questions for authors

See weakness for details

2 comments

We thank reviewer Q4f5 for recognizing the sensor-gap problem, the interpretable Propose → Project → Screen mechanism, and the EPV analysis. We address five concerns: projection and claim support (W1), C2KD comparison (W2), projection-step ablations (W3), Deployment-CoT quality (W4), and abstract emphasis (W5). Revisions are highlighted in blue in the revised main paper.

W1. Projection, screening, and claim support.

The projector PdP_d is a prompted LLM (Claude Sonnet 5 for radar; Gemini 3.6 Flash for audio). It receives xdx^d, question, ground-truth answer, and privileged rationale zpz^p, but not privileged sensor xpx^p. Claude Opus generated 207 radar examples and selected revisions. The screen VV checks the final answer and rejects rationales mentioning the generation process, RGB-only attributes, frame-by-frame descriptions, or invalid lengths. Details are clarified in Section 4, the new Table B.1 (Appendix B.1), and the new Appendix G.

In the new Appendix F (Tables 4 and 5), an independent GPT-6.1 evaluator assesses 400 radar training examples across the three supervision methods without seeing the answer, RGB, or method identity. Four human annotators separately assess sampled non-appearance claims. Appearance claims are counted as unsupported by radar.

Radar-supported claimsProjected-CoTDeployment-CoTPrivileged-CoT
Independent model84.6%81.7%21.0%
Human-assisted79.3%—28.1%

Projected-CoT has the highest supported-claim share. Human agreement is 88.1% for paired yes/no judgments. Human-assisted rates include automatically flagged appearance claims as unsupported.

W2. Comparison with C2KD.

We add a C2KD-style baseline on RadarQA-300, using 1,308 training examples, a Qwen3-VL-8B LoRA student, and three seeds.

MethodAccuracy (%)
Projected-CoT68.8 ± 0.5
C2KD-style51.0 ± 2.0

Projected-CoT leads by 17.8 points (new Table 8, Appendix J). No official C2KD implementation is public. Our baseline uses soft-label distillation and rank-based selection, but omits additional answer cross-entropy, non-target-class and bidirectional distillation, and proxy modules. See the new Appendix J for more details.

W3. Ablation of CRP Projection Steps.

We evaluate CRP projection-step ablations and deletion-based methods on RadarQA-300 with Qwen3-VL-8B LoRA over three seeds, without consistency screening (Section 7.4; new Table 7, Appendix I).

CRP projection steps / deletion methodsAccuracy (%)
All four steps69.6 ± 1.0
Without ABSTRACT47.7 ± 2.6
Without REMOVE57.7 ± 3.1
KEEP + REMOVE55.2 ± 1.2
KEEP-only45.0 ± 2.1
Delete unsupported claim words45.7 ± 5.0
Random sentence deletion53.2 ± 3.2
Random word-run deletion45.9 ± 2.7

ABSTRACT contributes most: omitting it reduces accuracy by 21.9 points, while omitting REMOVE costs 11.9 points. Deletion alone is insufficient: KEEP + REMOVE trails the four-step projection by 14.4 points, and selective or random deletion performs worse. Explicitly specifying HEDGE adds only 0.8 points relative to the original prompt (68.8 ± 0.5%), within one standard deviation. Prompts and further evaluations are in the new Appendix G.3.

W4. Deployment-CoT Generation and Baseline Quality.

The new Table B.1 (Appendix B.1) details Deployment-CoT generation. For 1,101 radar examples, Claude Sonnet 5 re-examines the radar using an initial radar-only reading and Projected-CoT as references, without RGB. This uses 4.02 teacher calls per example versus two for Projected-CoT, plus an additional screen. Claude Opus generated the remaining 207 examples. Audio Deployment-CoT uses Gemini 3.6 Flash with audio, question, and answer, but no video. An independent evaluation finds 81.7% of radar Deployment-CoT claims supported versus 84.6% for Projected-CoT (Table 4).

We test a clean radar Deployment-CoT without projected-rationale references or extra screening, plus audio Deployment-CoT from Gemini 3.8 Flash:

Sensor / studentDeployment-CoTProjected-CoTGain (points)
Radar, Qwen3-VL-8B full (original)68.0 ± 4.272.4 ± 1.0+4.4
Radar, Qwen3-VL-8B full (clean)67.4 ± 1.372.4 ± 1.0+5.0
Audio, Qwen2.5-Omni-7B LoRA (original)62.3 ± 3.368.9 ± 1.8+6.6
Audio, Qwen2.5-Omni-7B LoRA (newer)65.2 ± 0.268.9 ± 1.8+3.7

Projected-CoT remains better in both revised comparisons (Table 1; Section 7.1; Appendices B.1 and G.2).

W5. Accuracy margins in the abstract.

We revise the abstract to lead with the gain over Deployment-CoT (up to 6.6 points), before the 5.9- and 12.9-point gains over Answer-only on radar and audio. Projected-CoT exceeds Deployment-CoT in all seven evaluated model and fine-tuning configurations, with gains of 0.8–6.6 points (Tables 1 and 2).

Thank you again for your constructive review. We hope that we have addressed your comments in the rebuttal and revised manuscript (please check the revised pdf for the main paper with highlighted changes in blue), with additional details in the appendices. We would appreciate your feedback on whether these revisions resolve your concerns and would be happy to clarify any remaining points.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Shared Response to All Reviewers (Round 2)

Dear Reviewers,

Thank you for your comments and suggestions. Following our earlier responses, we further evaluate whether the students' inference-time reasoning is supported by radar measurements, when the students receive radar inputs without privileged RGB. In particular, we examine whether correct answers are accompanied by claims supported by the radar measurements, a fine-grained sensor grounding, rather than just the final answer accuracy, during inference. The training rationale evaluation was included in the previous response.

1. Radar support of inference-time reasoning.

Specifically, we compare the inference rationales of students trained using CRP (claim-based projection) and generic rewriting, respectively. Both methods start from the same privileged rationales. Generic rewriting uses an LLM to adapt these rationales to radar-based reasoning without explicitly specifying CRP's four projection steps: KEEP, ABSTRACT, HEDGE, and REMOVE. Both methods use the same two-call teacher protocol without additional rationale editing, as described in our follow-up response to Reviewer TPxR, for a fair comparison. This comparison examines whether explicit sensor-aware claim-level projection improves the support of the generated rationales beyond general rewriting.

For each method, we train five Qwen3-VL-8B students with full-parameter fine-tuning using seeds 11, 21, 31, 41, and 51. An answer-blind evaluator extracts a decisive claim from each student's generated rationale, and two independent model judges assess the claim against radar measurements using the following criteria:

  • Judged true: Both judges agree that all checkable statements in the claim match the radar measurements.
  • Classified as not checkable: The claim contains no statement that the checking rules can verify, the required measurements are unavailable, or the judges disagree.

Results over five seeds on the same 259 test questions are shown below. Percentages are over assessed decisive claims.

Measure, mean ± SD (%)CRPGeneric rewriting
Decisive claim judged true ↑55.4 ± 2.751.4 ± 2.8
Correct answer and decisive claim judged true ↑43.9 ± 3.240.1 ± 2.6
Decisive claim classified as not checkable ↓20.3 ± 0.924.7 ± 1.7
Correct answer and decisive claim classified as not checkable ↓14.8 ± 1.217.8 ± 1.9
Decisive claim judged true when both methods answer correctly ↑63.2 ± 3.757.8 ± 2.1

↑ Higher is better. ↓ Lower is better.

The first four rows use the same 259 test questions across five seeds. The fifth row considers only questions where both student models give correct answers (153–162 questions per seed, totaling 781 question–seed pairs). In this both-correct subset, CRP has a higher rate of decisive claims judged to be supported by radar input in all five seeds.

These results provide additional evidence that CRP improves radar support of student-generated reasoning at inference time, even when both methods predict the correct answer. This suggests a benefit of explicit claim-based projection over generic rewriting in this experimental setting, beyond differences in final-answer correctness.

2. Example of inference-time reasoning and radar evidence.

Appendix K, Figure 7 of the updated main paper shows the actual radar input (five radar frames in a top-down view, colored by time) and the corresponding RGB frames. The complete student rationales from CRP and generic rewriting are reproduced in Appendix K. RGB is shown for reference and is not provided to either student at inference time.

Both students correctly answer “toward the radar's left,” but their rationales describe different positions of the radar point cluster.

DescriptionCRP rationaleGeneric rewritingRadar measurement
Final lateral position−1 to 0 m−2 to −1 mMedian: −0.39 m

The measured lateral medians across the five frames are +0.82, +0.68, +0.43, 0.00, and −0.39 m, supporting the leftward-motion conclusion in both rationales. CRP's final-position estimate agrees with the measured median, whereas generic rewriting places the final cluster farther left.

These additional results provide evidence of radar support in the students' inference-time reasoning, beyond final-answer accuracy. They also show that unsupported details can remain in otherwise correct rationales.

We hope these additional analyses and the illustrative example help clarify the evidence for radar support in student-generated reasoning and the contribution of claim-based projection. We would appreciate your consideration alongside our earlier responses and the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.