acceptodds
Under review as a conference paper at ICLR 2027

Lightweight Reasoning Transfer via Chain-of-Thought Vector Projection

Abstract

Transferring reasoning capabilities from large language models to smaller ones often requires extensive teacher supervision or costly student adaptation. We study projected cross-model Chain-of-Thought (CoT) Vector transfer, a two-stage approach that keeps both models frozen. First, we learn a task-level CoT Vector in the teacher using dataset-provided solution traces and final-answer supervision. We then train a lightweight linear projector to map the fixed vector into a smaller student’s residual space. Student-side training uses only final-answer cross-entropy on a small subset of task training examples; it requires neither teacher outputs nor intermediate solution traces. We select the student injection layer and intervention scale on a labeled support set and report results on the disjoint remainder of each evaluation split. On GSM8K and MATH, the projected vector improves base students by 10.37–21.98 accuracy points across the evaluated teacher–student settings. On four code-generation benchmarks, the Qwen2.5-7B-Instruct student gains 3.23–4.67 pass@1 points, while the base Qwen2.5-7B student gains up to 9.65 points. These results show that a teacher-side reasoning vector can be reused across smaller students with different residual dimensions through projector-only adaptation.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
chain-of-thought, reasoning distillation, representation transfer, knowledge distillation, large language models

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

This paper proposes a method for knowledge distillation (KD) leveraging the Chain-of-Thought (CoT) vectors. The method consists of two stages: in the first stage, CoT vectors or the compact representations of CoT are learned from the teacher model. The vectors are learned by switching a learnable vector on/off at each layer depending on the presence of the ground-truth CoT along with the query and the answer in the prompt. In the second stage, the student is trained to reason with the frozen teacher's CoT vectors. The method is evaluated on math and coding benchmarks to show the performance gain achieved by the proposed KD method.

Critical strengths
  1. The method is straightforward to implement and reproduce. The paper is clearly written and easy to follow.

  2. The proposed method achieves significant gain as compared with the baseline student (no KD) for various reasoning tasks.

Critical weaknesses
  1. Limited technical contribution. The proposed method heavily relies on the original work on CoT vector [1] (cited in the paper). For example, the first stage of the proposed method is essentially identical to the procedure in [1]. Overall, this method simply transfers CoT vector method to KD without additional technical novelty or significant insights. The level of contribution appears limited considering the publication standard of ICLR conference.

  2. Limited comparison with existing baselines. The evaluation section mostly reports the performance gain comparing with or without the proposed KD. This is more of a self-comparison (ablation). This work completely lacks comparison with existing KD methods (e.g., widely studied baselines like SeqKD, Generalized KD (GKD), DistiLLM-1,-2, Speculative KD, or recent methods concerning distilling long reasoning traces from teachers like P-ALIGN [2] or ORION [3]), although the work deals with traditional topic and benchmarks. Thus, it is difficult to judge whether the proposed method is a genuine enhancement of existing KD framework.

[1] Li, Li, et al. "Cot vectors: Transferring and probing the reasoning mechanisms of llms." International Conference on Learning Representations. Vol. 2026. 2026.

[2] Liu, Zhenghao, et al. "Long-chain reasoning distillation via adaptive prefix alignment." Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

[3] Wu, Zhuoyang, et al. "Enhancing Long-Chain Reasoning Distillation through Error-Aware Self-Reflection." Findings of the Association for Computational Linguistics: ACL 2026. 2026.

Questions for authors

Please see weaknesses.

3 comments

Thank you for your constructive evaluation. We address your specific concerns and questions below.

  • Technical Contribution, Practical Utility & Privacy Protection:

    We agree that Stage 1 builds upon prior work (Li et al., 2026). However, our work introduces a highly practical and secure paradigm: achieving cross-scale reasoning representation transfer while keeping both teacher and student backbones completely frozen. By training only a lightweight linear projector on a tiny subset of data using final-answer supervision, our method avoids the high computational costs and complex pipeline requirements of traditional KD.

    Crucially, this design offers exceptional privacy protection for the teacher model. Because student-side adaptation requires no teacher logits, no hidden-state alignment, and no intermediate solution traces, the teacher model's proprietary weights, internal representations, and text generations remain entirely unexposed.

  • Inclusion of Additional KD Baselines (LoRA, SFT, DistiLLM, P-ALIGN, and ORION):

    Following your feedback, we have added extensive comparisons with LoRA, SFT, and DistiLLM under a matching 200-example budget on MATH, where our method consistently outperforms them (59.13% vs. 48.98% / 47.94% / 49.20%).

    Regarding recent long-chain reasoning distillation methods such as P-ALIGN and ORION, we attempted to implement them under our base-student setting but found they experience severe bottlenecks during the data-construction phase (yielding a survival rate of only ≈1%\approx 1\%). Specifically, P-ALIGN requires uninstructed base students to execute meta-tasks like self-assessing prefix sufficiency (e.g., generating [ENOUGH]/[NOT_ENOUGH] markers), while ORION requires step-by-step error reflection. Base students fail to follow these auxiliary formatting instructions, preventing the successful creation of a training set. We have duly cited both works in our related work section.

Reviewer_xZGiOpenReview ↗

Thanks

Thank you for the responses. They do not fully address my concerns, in particular the one on the technical novelty or how this work extends beyond CoT vectors. Thus I would like to keep my rating. I gave the score of 1 because the rating scale is coarse (only 4-point scale); in finer scales, I would rate this work a bit lower than weak reject, but not downright reject. I simply didn't have choice this time. I hope the authors could understand this aspect and do not take my rating as too much of discouragement.

Thank you for your candid follow-up.

We respect your perspective on technical novelty, and would like to briefly highlight two aspects where our work extends beyond prior CoT vectors and conventional distillation:

Bridging Incompatible Residual Spaces: Prior CoT vector transfer assumes identical residual dimensions between models, whereas transferring across model scales (e.g., 72B →\to 7B or 7B →\to 1.5B) requires aligning distinct representation spaces. Showing that a lightweight linear projection trained solely with final-answer cross-entropy—without teacher logits, hidden-state alignment, or intermediate traces—can preserve task-level reasoning guidance across dimensions provides a simple yet non-trivial empirical insight into cross-scale activation transfer.

Practical Utility in Low-Resource & Base-Model Settings: Unlike rationale-based or reflection-based distillation pipelines (which often require instruction-following capabilities to generate or filter intermediate trajectories), our two-stage design works directly on uninstructed base models with just 200 final-answer examples, while keeping the teacher's internal weights and generations completely unexposed during student adaptation.

Thank you again for your time, constructive critiques, and thoughtful engagement throughout the review process.

Reviewer_vhn5OpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes transferring reasoning capabilities from large to small LLMs via “projected CoT vector transfer.” Stage 1 learns a task-level CoT vector in the frozen teacher’s activation space. Stage 2 trains a bias-free linear projector that maps this fixed teacher vector into the student’s residual stream, using only 200 training examples and final-answer cross-entropy. The injection layer and scale are selected on a labeled support set drawn from each official evaluation split, and results are reported on the disjoint remainder. On Qwen2.5-Math teacher–student pairs, the method improves the base students by 10.37–21.98 points on GSM8K and MATH, while the 1.5B-Instruct student shows essentially no change (−0.09 / +0.34). On code benchmarks, the Qwen2.5-7B-Instruct student gains 3.23–4.67 pass@1 points on all four benchmarks. The base 7B student gains 9.65 points on HumanEval but at most 0.80 on the other three, with zero gain on LiveCodeBench and DS-1000.

Critical strengths
  1. The paper clearly states what it does and does not do. The two-stage decomposition (teacher-side vector learning, then student-side projector training) is simple and well motivated by prior work on task vectors and activation steering. The teacher vector is learned once and can be reused across students with different residual dimensions.
  2. The student side trains only a bias-free linear projector (~5.5M or ~29.4M parameters) on 200 examples for two epochs. It needs no teacher logits, no hidden-state alignment, and no intermediate solution traces. Inference adds a single precomputed vector. This is much lighter than full-model distillation on the student side. (The teacher-side cost is a separate matter; see Weakness 3.)
  3. The paper reports fail-to-pass and pass-to-fail transitions for the 1.5B-Instruct student (Table 9), shows failure cases (Cases 3 and 4 in Appendix C; the base-student case in Appendix D), and reports zero gain for the base 7B student on LiveCodeBench and DS-1000. It also lets the reader see that the 5.00-point support-set gain for the 1.5B-Instruct student on GSM8K does not persist on the held-out set. The case studies are well chosen and informative.
Critical weaknesses
  1. With v_T fixed, the objective constrains only Pv_T, so the method may reduce to learning a single student-side vector under a different parameterization. The only direct-vector ablation (Table 6) covers one setting and uses different scales (0.10 vs. 0.20), so it cannot attribute the +21.98 vs. −0.71 gap to teacher information. Random-vector or cross-dataset-vector controls are missing. Gains also concentrate in base students with degenerate outputs (Appendix C, Cases 2 and 5; SVAMP 26.2→70.1), suggesting output repair rather than reasoning transfer.
  2. There is no comparison with LoRA, soft-prompt tuning, or CoT distillation under a comparable student-side budget. The efficiency claim also ignores teacher-side cost: one vector per dataset and candidate block, trained on the full training split with solution traces.
  3. The layer and scale are selected on labeled examples from the evaluation split. This is not leakage, but the max over a layer × scale grid is optimistic, as seen in the 1.5B-Instruct GSM8K gain (+5.00 support vs. −0.09 held-out). All results are single greedy runs without confidence intervals. From the reported percentages, I infer that the held-out code sets are small (e.g., about 62 problems for LiveCodeBench), so +3.23 is roughly two problems.
Questions for authors

See above

1 comment

Thank you for your thorough review and insightful comments. We address each point below.

  • Expressivity and Direct-Vector Controls:

    We acknowledge that with vTv_T fixed and PP unconstrained, PvTPv_T spans the student's residual space. However, direct optimization of a student-side vector uSu_S or using norm-matched random vectors fails to replicate our success (achieving 37.46% and 42.63% on MATH, respectively, compared to 59.13% with the true teacher vector). This demonstrates that the success of our method stems from the information carried by the teacher vector rather than mere parameter capacity.

  • Output Repair vs. Reasoning Transfer:

    We agree that repair is a real component of the gain, and we quantify it. We define a zero-shot output as degenerate if no answer can be extracted or the generation collapses into repetition. On the GSM8K official test set (1,319 questions, per-question predictions), 270 (20.5%) zero-shot outputs are degenerate; of these, 258 (19.6%) are degenerate and wrong, while the remaining 12 degenerate outputs still extract a correct answer. Injection repairs 156 of the 258 degenerate-and-wrong outputs (60.5%), and the overall degenerate-output rate falls from 20.5% to 2.4%. This is the repair component. But it is not the whole story: on the remaining 1,049 questions whose zero-shot outputs are not degenerate, accuracy rises from 55.7% to 68.5% — a net gain of 134 questions (+12.8 pp; 221 gains vs. 87 losses). Roughly half of the total +290-question gain therefore comes from questions that needed no repair at all.

    Furthermore, Appendix C Case 1 (GSM8K #500) illustrates the non-repair component. The baseline output is complete and legal: it sets up the correct equation D+(D+7)+2D=43D + (D + 7) + 2D = 43, but wrongly concludes D=10D = 10 and emits 17\boxed{17}. After injection, the student explicitly expands 4D+7=43  ⟹  4D=36  ⟹  D=94D + 7 = 43 \implies 4D = 36 \implies D = 9 and answers 1616. The baseline failure is an intermediate algebraic computation inside an otherwise well-formed solution—not repetition, a missing answer, or a boxed-format error. The intervention thus improves computational correctness within already well-formed solutions, not only repairs degenerate ones.

  • Comparison with SFT, LoRA, and Advanced Distillation Baselines:

    We evaluated matched-budget baselines on MATH using the same 200 examples and 2 epochs. LoRA, SFT, and DistiLLM achieve 48.98%, 47.94%, and 49.20%, respectively—significantly below our 59.13%, despite receiving intermediate reasoning supervision that our method avoids.

  • Configuration Selection, Decoding Protocol, and Benchmark Scope:

    Grid-Selection Optimism & Zero-Tuning Generalization. While support-set grid search introduces optimization optimism (as seen in the 1.5B-Instruct GSM8K support vs. held-out variance), our cross-dataset transfer results (Table 5) confirm that the learned configuration is robust rather than overfitted: transferring the GSM8K configuration directly to SVAMP and ASDiv without any target-dataset tuning yields massive gains of +43.90 pp and +41.86 pp, respectively.

    Decoding Protocol. Regarding single runs, in standard evaluation protocols for math and code reasoning (such as the Qwen2.5-Math official protocols), single-run greedy decoding (temperature = 0) is the established academic practice to ensure strict reproducibility and deterministic comparison.

    Code Benchmark Consistency. While individual held-out code splits are compact, our evaluation spans four distinct code-generation benchmarks (HumanEval, BigCodeBench, LiveCodeBench, and DS-1000), where the 7B-Instruct student demonstrates consistent positive gains across all four benchmarks (+3.23 to +4.67 pass@1), reflecting reliable cross-task improvement.

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

This paper studies how to transfer reasoning ability from a large teacher to a smaller student through reasoning vectors. The method has two stages. First, a task-level CoT Vector is learned in the teacher, following the Learnable CoT Vector objective of Li et al. (2026). Second, the vector is kept fixed, and a bias-free linear projector is trained to map it into the student's residual stream. The main contribution is extending prior same-width CoT Vector transfer to students of different scales and residual dimensions through projector-only adaptation. Experiments cover Qwen2.5-Math teacher–student pairs on GSM8K and MATH, with cross-dataset transfer to SVAMP and ASDiv, and Qwen2.5 pairs on four code-generation benchmarks.

Critical strengths
  • Clear writing and well-stated contributions. This paper is easy to follow. And the contributions are stated clearly, and the paper is explicit about what each stage uses and does not use.
  • Extensive experimental results are presented. The evaluation covers multiple teacher–student pairs on math and code generation. It also includes cross-dataset transfer, and complementary ablation studies.
Critical weaknesses
  • Limited novelty. The teacher-side stage reuses the existing work of Li et al. (2026). Mapping steering vectors or activations across models with linear or affine maps has also been studied. The new component is mainly a linear projector trained with final-answer cross-entropy, which is a relatively incremental extension.

  • Ablations do not sufficiently justify the design.

    a) Because vTv_T is fixed and nonzero, PvTPv_T can represent any vector in the student's residual space. Projector training and direct student-vector learning therefore have the same expressivity. A control with a random or mismatched (e.g., other task) vTv_T is needed to show that the teacher vector matters.

    b) There is no comparison with simple student-side baselines using the same 200 examples, such as LoRA/SFT or few-shot prompting.

Questions for authors

1, Can you expand the exp in Table 6 to other benchmark and model variants? So we can have a better understanding of this observation. 2, I would expect a further analysis regardign the effect of reasoning vector and the projection part, with comparison of 1)randomly initialized vector [frozen] with projection, and 2) replace the fronzen reasoning vecto + proejction with trainable vector [same parameter size with projection], and tune it with the same recipe.

3 comments

Thank you for the constructive feedback and for recognizing the clarity of our writing and the breadth of our experiments. We address your specific concerns and questions below.

  • Extension of Direct Student-Vector Comparisons & Ablations:

    We have extended the direct student-vector baseline and random vector controls to additional benchmarks (MATH and HumanEval), following the same data and epoch budget (200 training examples, 2 epochs).

    Math reasoning (7B-Instruct teacher → 1.5B base student). On MATH, direct student-vector training achieves 37.46% (compared to 38.08% for the baseline), while our projected teacher CoT Vector reaches 59.13%.

    Code generation (32B-Instruct teacher → 7B-Instruct student). On HumanEval, direct student-vector training yields 71.93% (baseline: 72.81%), whereas our projected teacher vector achieves 77.19%.

  • Analysis of the Reasoning Vector and Projection Components:

    We evaluated two specific controls to isolate the contribution of the teacher-side CoT Vector (Table 7).

    Direct trainable student vector (uSu_S). Learns a trainable vector directly in the student's residual stream without a teacher vector or projector under the same 200-example budget, yielding 37.46% on MATH and 71.93% on HumanEval.

    Norm-matched random vector + projection. Replaces the teacher-side CoT Vector with a frozen random vector in the teacher's residual space (rescaled to match the CoT Vector's Euclidean norm) and trains a linear projector. This achieves 42.63% on MATH and 70.18% on HumanEval.

  • Discussion on Expressivity, Novelty & Privacy:

    While unconstrained linear projection PvTPv_T theoretically spans the student's residual space due to a fixed non-zero vTv_T, our empirical controls prove that parameter capacity and direct student-side optimization alone are insufficient. Under an identical training budget, directly training a student vector (uSu_S) fails to improve over the baseline (37.46% vs. 38.08% on MATH; 71.93% vs. 72.81% on HumanEval), and replacing the teacher vector with a norm-matched random vector collapses the performance from 59.13% / 77.19% down to 42.63% / 70.18%.

    Regarding novelty and practical utility, while the teacher-side vector learning builds upon Li et al. (2026), our core contribution lies in achieving cross-scale representation transfer between incompatible dimensions. Furthermore, our framework offers a strong guarantee for teacher model privacy: because student-side adaptation requires no teacher logits, no hidden-state alignment, and no intermediate solution traces, the teacher's internal parameters and generations remain completely unexposed during student adaptation. This provides a secure, decoupled paradigm for capability transfer in privacy-sensitive deployments.

  • Comparison with Additional Baselines (LoRA, SFT, DistiLLM):

    We have added direct comparisons with LoRA, SFT, and DistiLLM under the same 200-example budget on MATH. Our method significantly outperforms these student-side tuning baselines (e.g., achieving 59.13% compared to 48.98% for LoRA, 47.94% for SFT, and 49.20% for DistiLLM), demonstrating the clear advantage of leveraging the projected teacher CoT Vector over conventional student adaptation.

I thank the authors for the rebuttal. Two concerns raised are note fully addressed in the rebuttal, so I will keep my score unchanged.

  1. The random-vector control does not isolate teacher information. Taking v_T fixed, only u = P v_T matters, and ∂L/∂P = g v_Tᵀ. Under standard optimizers with zero or i.i.d. initialization of P, the trajectory of u depends on the norm of v_T but not on its direction. Thus, the projector is effectively a direct student vector with a rescaled learning rate. The gap between the teacher vector and the norm-matched random vector (59.13 vs. 42.63) is therefore more likely due to differences in effective learning rate or training setup than to teacher information.

  2. The direct-vector baseline seems under-tuned. It falls below the no-intervention student in both settings, eventhough the zero vector is in its feasible set.

Thank you for the follow-up and for sharing your mathematical perspective on ∂L/∂P=gvT⊤\partial \mathcal{L}/\partial P = g v_T^\top.

Optimization dynamics vs. norm scaling: While u=PvTu = P v_T determines the injected vector, the resulting trajectory in practice is not merely a scalar learning-rate rescaling of vTv_T's norm. Because the injected vector interacts with layer activations across sequence positions and updates occur under adaptive/momentum-based optimization and fixed training budgets, using a norm-matched random vector versus the teacher CoT vector leads to markedly different empirical outcomes under the exact same training recipe (59.13% vs. 42.63% on MATH). We view this empirical gap as evidence that the learned teacher vector provides more useful initialization/direction for transfer than an arbitrary norm-matched vector.

Fairness and reproducibility of the direct-vector baseline: We ensured a strictly fair comparison by evaluating all ablations under the exact same unified protocol—using the identical 200-example training split, 2-epoch budget, the same learning rate (1×10−41\times10^{-4} on MATH and HumanEval) and support-set configuration selection procedure, with fixed random seeds so that every experiment is fully reproducible. Even though the zero vector lies in the feasible set, optimizing uSu_S directly from scratch under this tight low-data regime without a teacher prior can exhibit slight fluctuations relative to the uninjected baseline, which further highlights why leveraging a teacher-derived vector is advantageous when adaptation data is scarce.

We appreciate your time and constructive engagement throughout the review process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.