{"id":"6f9f6054-d911-41ba-97f2-7aada2539b84","arxiv_id":"2506.09967","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SAE-Tuning, a sparse-autoencoder-guided SFT procedure, elicits RL-comparable reasoning in 1.5B models from CoT-free QA data at about $1 and 20 minutes of training.","lead":"This paper introduces a cheap fine-tuning method that uses sparse autoencoders to transfer reasoning skills from one small language model to another using only question-answer pairs, without chain-of-thought traces. It reports matching the performance of RL-trained models at roughly $1 and 20 minutes of training, and shows the resulting reasoning adapters can be swapped between models at test time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stage II (Eq. 8) trains only a KL-consistency objective with no answer-label supervision, and no null-SAE control exists; the claimed feature-transfer mechanism is therefore untested.","rationale":"The paper's headline is an empirical performance claim, but the mechanism ('SAE-Tuning' as feature transfer) is what makes the method novel and grounds the transparency title. The reader's weakest_assumption already identifies this gap; I agree. My reading of Eq. 8 sharpens it: the loss is only a consistency constraint between the model with and without the SAE, so the trained SAE could be acting as a fixed perturbation whose magnitude, not whose learned content, produces the gain. The existing controls do not rule this out: the SFT baseline changes the loss function, and the pretrained-SAE control still supplies a nontrivial SAE. A null-SAE control is the minimal experiment that would separate feature transfer from KL-regularized training. If the null control reproduces the gain, the central mechanistic claim fails even if the empirical numbers are reproducible. The best-checkpoint selection issue noted by the reader is real but secondary: it affects confidence in the exact 48.06 number, whereas the missing control affects what the number means. Conditional acceptance with a request for this control is appropriate.","tokens_in":28327,"tokens_out":11677,"duration_ms":144533,"concrete_test":"Run the exact Resa-STILL-v5 Stage II protocol (same data, LoRA, hyperparameters, Eq. 8, checkpoint-selection rule of Table 4) with the trained SAE replaced by (i) a frozen randomly initialized SAE of identical architecture and top-k, scaled so its average L2 reconstruction error on the trigger set matches the trained SAE's, and (ii) an identity transform plus isotropic noise of the same RMS. If either null control reaches the 48% range, learned SAE features are not the operative ingredient; if both land near the 39-45% range, the feature content is load-bearing. Also report the SAE-augmented target model's six-benchmark average before adapter training.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. 8 makes the entire Stage II signal a KL divergence between the target model's own next-token distributions with and without the frozen SAE inserted. The verified answer tokens in the elicitation data appear only as conditioning tokens, never as a supervision term. The non-SAE control in Table 3 (STILL-CoT-free-SFT) changes the loss to answer-label cross-entropy, so the 39.00 vs 48.06 gap conflates the SAE with the loss change. The SAE-mode ablations in Table 4 all use nontrivial SAEs; the pretrained SAE, which never saw the trigger data, already reaches 44.99, so much of the gain over SFT is not specific to trigger-trained reasoning features. No control uses a randomly initialized SAE or an identity-plus-noise perturbation with matched reconstruction error, and the SAE-augmented 'teacher' is never benchmarked before adapter training. Thus the observed improvement could be a regularization/consistency effect of learning to be invariant to a fixed activation perturbation, rather than transfer of learned reasoning features.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SAE-Tuning, a two-stage method for eliciting reasoning abilities in 1.5B language models. Stage I trains a sparse autoencoder (SAE) on the activations of a source model at a chosen layer, using verified chain-of-thought-free (CoT-free) question-answer data formatted with a <think>[answer]</think> template. Stage II freezes this SAE, inserts it into a target model, and trains LoRA adapters to minimize the KL divergence (Eq. 8) between the target's output distributions with and without the SAE, after which the SAE is discarded. The paper claims that SAE-Tuning matches or approaches RL-trained counterparts on six reasoning benchmarks at roughly $1 and under 20 minutes of training, that the extracted reasoning abilities generalize across datasets and transfer as modular adapters across models in the Qwen family, and that layer-wise counts of SAE features activated at think tokens align with layer-wise reasoning performance, providing transparency.","tokens_in":28655,"tokens_out":10411,"duration_ms":109960,"significance":"If the empirical results hold, the cost and data-efficiency claims are significant for democratizing reasoning-model research, and the open release of code, weights, and training logs is a concrete strength. The modular-adapter result, where LoRA adapters trained on Qwen or Qwen-Math transfer to R1-Distill without retraining, is a potentially useful finding for compositional reuse. However, the interpretive layer of the paper—feature transfer and transparency—is currently under-supported: the mechanism in Eq. (8) is not isolated by controls, and the Section 5 GMM analysis does not establish prediction. The headline quantitative claims are also built on small benchmarks, best-checkpoint selection, and a cost baseline that is not matched to the direct RL counterpart. The work will interest the reasoning and interpretability communities, but the evidence as presented supports the empirical method more strongly than it supports the feature-transfer and transparency interpretations.","major_comments":[{"comment":"The headline efficiency gains (\">2000x cost, >450x time\") are computed against STILL-3-1.5B-Preview and DeepScaleR-1.5B-Preview, not against the RL-trained counterparts (Tina-STILL, Tina-DeepScaleR) to which the performance comparison is actually made. By Figure 1's own numbers, Tina-STILL costs $60 and 60 hours, so Resa-STILL-v5's $1 and 0.33 hours represent roughly a 60x cost reduction and a 180x time reduction against the matched counterpart, not the >2000x and >450x stated in the Abstract. Please report the efficiency comparison against the direct RL counterpart and qualify the headline claims accordingly.","section":"Section 3.2, Figure 1, Table 2"},{"comment":"The central mechanistic claim that Stage II transfers learned reasoning features is untested. The only non-SAE control (STILL-CoT-free-SFT, Table 3, 39.00 avg) differs from SAE-Tuning in both the presence of the SAE and the loss function (answer-label cross-entropy vs. the KL divergence in Eq. 8). Moreover, Resa-STILL-v2 with a pretrained SAE that never saw the trigger data reaches 44.99 avg, capturing 3.07 of the 9.06-point gap over SFT. Without controls using a randomly initialized SAE, an identity-plus-noise perturbation with matched reconstruction error, or a pure KL-consistency baseline without any SAE, the observed gains could be attributed to a consistency regularizer rather than to the SAE's learned reasoning features. Please add such controls or substantially soften the feature-transfer interpretation.","section":"Section 2.1, Eq. (8), Table 4"},{"comment":"The transparency analysis fits a 3-component GMM to the 26 layer-wise feature counts and to the 26 layer-wise performance scores, then reports that the fitted means and weights align. This does not establish prediction. Both distributions are indexed by the same ordered layer indices, so similar component locations are expected for any smooth trend; the feature counts are small integers (0-5), the performance scores are means of six small benchmarks with no uncertainty, and a 3-component GMM has roughly eight free parameters on 26 points. The text also inconsistently states that the feature-count GMM is fit to the base model's counts (Section 5) while Table 6's caption says the counts are for Tina-STILL. Please specify the exact input data, provide a permutation test or held-out layers to show the alignment exceeds chance, and give bootstrap confidence intervals on the GMM means. As written, the claim that the structure is a robust predictor is not supported.","section":"Section 5, Table 6, Figure 3"},{"comment":"All reported Resa scores are best-checkpoint selections, and checkpoint-to-checkpoint variation is large (e.g., Resa-DeepScaleR-v1 ranges from 40.87 to 48.95 across checkpoints in Appendix Table 9). The headline comparison of 48.06 vs. 48.16 is within the sampling noise of the 30-problem AIME sets, and no seeds or error bars are reported. In addition, the Abstract mixes configurations: the 43.33 AIME24 and 90 AMC23 examples come from Resa-STILL-Tina-50-step (Trained-from-Scratch), which uses a lightly RL-trained source, not the base-model source used for the >97% retention claim. Please specify a fixed checkpoint-selection rule, report bootstrap confidence intervals or per-benchmark standard errors, and align each Abstract claim with the exact configuration in Table 4.","section":"Section 3.2, Tables 3-4, Appendix B"},{"comment":"The paper states in Section 2.1 that the <think>[Answer]</think> template is hypothesized to activate the source model's latent reasoning abilities and promises a detailed analysis in Section 5, but Section 5 does not ablate the template in the training pipeline; it only counts SAE features that fire on think tokens in a separate prompt. The role of the thinking tokens is therefore untested, which matters because the method's CoT-free claim depends on this design choice. Please include an ablation that removes or alters the think tokens in the trigger/elicitation data, or explicitly acknowledge that the template's contribution is not isolated.","section":"Section 2.1, Section 5"}],"minor_comments":[{"comment":"Equation (3) contains a grammatical error: \"we only the top k features\" should read \"we keep only the top k features.\"","section":"Eq. (3)"},{"comment":"The sentence \"In the following, We now formalize\" has a capitalization and duplication issue; please rephrase.","section":"Section 2.1"},{"comment":"The column header \"GPQAMinerva\" in Table 4 is missing a space; it should read \"GPQA\" and \"Minerva\" as separate columns.","section":"Table 4"},{"comment":"Figure 1 is difficult to parse as rendered, with unclear axis labeling and bar values; please redraw with explicit units and legible value labels.","section":"Figure 1"},{"comment":"The reference \"Intelligent Internet. II-Thought: A large-scale, high-quality reasoning dataset, 2025\" and the in-text citation \"Internet, 2025\" are incomplete; please provide the full dataset documentation or a URL.","section":"References"},{"comment":"The phrase \"we cut off the first and final layer\" is informal; consider \"we exclude the first and final layers.\"","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical efficiency claim is plausible and potentially valuable, but the current evidence does not support the full set of headline claims. The missing mechanism controls (random SAE, no-SAE KL consistency) and the statistical treatment of the GMM alignment are load-bearing for the paper's transparency and feature-transfer contributions. I would encourage the editor to request the additional experiments rather than desk-reject, given the low cost of the proposed procedure and the open-source artifacts. The title and Abstract currently overpromise relative to the evidence; the authors should be asked to align claims with the controlled results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — the short version: this is a cheap and open recipe for eliciting reasoning without CoT traces, and the modular-adapter result is genuinely interesting. But the paper's central mechanistic story — that a frozen SAE transfers 'reasoning features' into the target — is not actually tested by the experiments as designed.\n\nWhat's new: SAE-Tuning runs SFT with a frozen SAE inserted at one layer and a plain KL-consistency loss, using only question-answer pairs formatted with <think>[Answer]</think>. Training cost is around $1 and 20 minutes on two GPUs. The authors show that a LoRA adapter trained this way on Qwen or Qwen-Math can be attached to R1-Distill at test time and give competitive scores. That result is surprising and worth following up on. The paper is also unusually open: code, weights, training logs, and a cost breakdown.\n\nThe soft spots are real. First, the Stage II objective (Eq. 8) is a KL divergence between the target's own outputs with and without the SAE. No answer-label supervision appears in that objective. The non-SAE control (STILL-CoT-free-SFT) uses standard cross-entropy on answers, so the 39.00 vs 48.06 gap conflates the presence of the SAE with a change in loss function. There is no random-SAE or identity-perturbation control, and the pretrained SAE that never saw the trigger data already reaches 44.99. The improvement from trigger-trained features over a generic SAE is therefore small, and the 'feature transfer' mechanism is untested. Second, all headline numbers are best-checkpoint selections without error bars; AIME has 30 problems, so 48.06 vs 48.16 is noise. Third, the transparency analysis is post hoc: a 3-GMM is fit to both feature counts and performance scores from the same 26 models, and the best layer (18) has zero detected 'reasoning features.' That weakens the claimed correlation.\n\nThe abstract also mixes numbers from different configurations (e.g., 43.33% AIME from a lightly RL-trained source vs. the end-to-end 48.06% average), which is easy to misread.\n\nWho this is for: practitioners who want a low-cost way to elicit reasoning, and interpretability researchers who will find the mechanism claim under-supported. It deserves a serious referee, but the right outcome is major revision: add proper null SAE controls, report variance, and reframe the feature-transfer language as speculative. I would not cite it yet as evidence for modular reasoning features, but it is a useful data point on cheap elicitation.","headline":"A cheap, open, and potentially useful recipe for CoT-free reasoning elicitation whose headline mechanistic claim—SAE feature transfer—is not actually tested by the experiments as designed.","tokens_in":29096,"tokens_out":2625,"would_cite":false,"duration_ms":29070,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SAE-Tuning elicits 1.5B-model reasoning from question-answer pairs alone, matching RL-trained models for roughly $1.","keywords":["sparse autoencoders","SAE-Tuning","reasoning elicitation","chain-of-thought-free supervision","low-rank adapters","KL divergence distillation","reasoning feature extraction","interpretability"],"falsifier":"Train the identical LoRA setup with the same $\\mathrm{KL}$ objective but replace the frozen SAE with a randomly initialized, frozen autoencoder of the same shape, or with an identity map; if the resulting model matches the reported gains, the reasoning-feature transfer story is wrong, while a collapse toward the 39% CoT-free SFT baseline would support the claim that the SAE's learned features are doing the work.","tokens_in":1524,"feed_emoji":"🧠","tokens_out":3358,"duration_ms":83685,"temperature":0.7,"pith_summary":"The paper tries to establish that strong reasoning can be elicited from a small base language model without reinforcement learning and without chain-of-thought traces, by first training a sparse autoencoder on verified question-answer pairs and then using that frozen autoencoder to guide supervised fine-tuning. The claimed payoff is cost and transparency: a 1.5B model reaches about the same average score as its RL-trained counterpart for roughly $1 and 20 minutes of training. The authors also argue that the extracted reasoning ability generalizes across datasets, transfers across models in the same family as a portable adapter, and leaves a layer-wise feature distribution that predicts reasoning performance.","feed_headline":"A $1 SAE recipe matches RL-trained reasoning in a 1.5B model","feed_subtitle":"Verified question-answer data, no reasoning traces, trains in 20 minutes and rivals costly RL baselines.","key_machinery":"The load-bearing object is the sparse autoencoder (SAE) inserted into the target model at a fixed layer, combined with the $\\mathrm{KL}$-divergence objective. The SAE is a dictionary of 65,536 features trained with a Top-$k$ encoder ($k=32$) to reconstruct the source model's layer-12 MLP activations; freezing it and requiring the target model's output distribution to match with and without its insertion forces the target's LoRA parameters to become compatible with features that encode reasoning. The <think>...</think> wrapper around a bare final answer is the trigger hypothesized to activate those features, and the trained SAE is removed at test time, leaving the reasoning ability in the model's own parameters.","core_discovery":"SAE-Tuning transfers reasoning ability from a source model to a target model using only CoT-free verified question-answer data. In Stage I, an SAE is trained on activations of a chosen source layer, capturing latent features triggered by a template that places the final answer inside <think> tokens. In Stage II, the frozen SAE is inserted after the same layer of the target model and low-rank adapters are trained to minimize the $\\mathrm{KL}$ divergence between the target's output distribution with and without the SAE; after training, the SAE is removed. The central result is that this procedure applied end-to-end to the base R1-Distill model as its own source (Resa-STILL-v5) scores 48.06% average on six reasoning benchmarks, essentially matching the RL-trained Tina-STILL at 48.16%, while a lightly RL-trained source yields peaks such as 49.46%.","pith_inferences":["The $\\mathrm{KL}$ objective itself could be acting as a regularizer; a decisive control would replace the frozen SAE with a randomly initialized frozen autoencoder of the same shape, or with an identity map, and check whether the reported gains survive.","The reasoning-as-adapter result suggests that reasoning ability and factual knowledge may be composable independently; a natural extension is testing whether the same adapter transfers to larger models or to models outside the architecture family when the SAE dimensions can be matched.","The tri-modal GMM correlation is an aggregate pattern, and the paper notes single-layer feature counts do not predict performance; a predictive test would use the identified layer-clusters to select hookpoints on held-out data and measure whether the resulting Resa models outperform layer-choice by heuristic.","If the mechanism is genuinely about surfacing latent reasoning features, SAE-Tuning could plausibly extend to other verifiable skills beyond math and science, such as code correctness or instruction following, using the same answer-only data format."],"forward_implications":["Standard SFT on the same CoT-free data without the SAE scores only 39.00% average, so the SAE is presented as the active ingredient rather than the answer-only template alone.","Training an SAE from scratch on the trigger dataset is about as effective as fine-tuning a pre-trained SAE, so SAE-Tuning needs no pre-existing SAE and no expensive SAE pre-training corpus.","Reasoning features extracted from the STILL dataset transfer to other datasets (DeepScaleR, Open-S1, II-Thought, OpenR1), reaching performance on par with models trained via RL on those datasets.","Adapters trained on Qwen or Qwen-Math can be attached to R1-Distill at test time without retraining, yielding scores around 47.5-47.9%, supporting the decomposition strong reasoning equals abstract reasoning ability plus foundational knowledge.","The layer-wise count of reasoning features forms a tri-modal distribution whose Gaussian mixture components align with the Gaussian mixture of achieved reasoning performance, offering a data-driven heuristic for choosing the SAE hookpoint layer."],"supporting_citations":[{"why":"Supplies the Tina models, the RL-trained source models and baselines whose reasoning performance SAE-Tuning replicates or matches.","marker":"Wang et al. 2025a"},{"why":"Provides R1-Distill, the base model that serves as target model and also as source in the end-to-end self-source experiments.","marker":"DeepSeek-AI 2025"},{"why":"Supplies the STILL trigger and elicitation dataset used in the main proof-of-concept and source-model ablation runs.","marker":"RUCAIBox STILL Team 2025"},{"why":"Supplies the DeepScaleR dataset and preview model used as an alternative source and elicitation dataset for cross-dataset and cost comparisons.","marker":"Luo et al. 2025"},{"why":"Provides the Top-k SAE training recipe that Stage I of SAE-Tuning follows for feature extraction and reconstruction.","marker":"Gao et al. 2024"}],"fun_headline_variants":["20-minute, $1 SAE tuning rivals RL-trained reasoning","SAE trick: $1 and 20 min to match RL-trained reasoning","Transparent reasoning from SAEs: >97% of RL at 1/2000th cost","SAE-tuning rivals RL-trained reasoning for $1 in 20 minutes","SAE reasoning modules transfer across models without retraining"],"cache_read_input_tokens":31360,"weakest_assumption_plain":"The mechanism assumes that writing a bare final answer inside <think> tokens activates the source model's latent reasoning features, and that matching the output distribution with a frozen SAE inserted transfers those features into the target's LoRA parameters; no experiment isolates whether gains come from feature transfer rather than from the $\\mathrm{KL}$ objective acting as a regularizer.","fun_headline_variants_meta":{"raw":{"variants":["20-minute, $1 SAE tuning rivals RL-trained reasoning","SAE trick: $1 and 20 min to match RL-trained reasoning","Transparent reasoning from SAEs: >97% of RL at 1/2000th cost","SAE-tuning rivals RL-trained reasoning for $1 in 20 minutes","SAE reasoning modules transfer across models without retraining"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00154,"raw_usage":{"total_tokens":6194,"prompt_tokens":1015,"completion_tokens":5179,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":5082}},"tokens_in":631,"tokens_out":5179,"duration_ms":39815,"temperature":1.0,"reasoning_tokens":5082,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:36:42.746996+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical LoRA setup with the same $\\mathrm{KL}$ objective but replace the frozen SAE with a randomly initialized, frozen autoencoder of the same shape, or with an identity map; if the resulting model matches the reported gains, the reasoning-feature transfer story is wrong, while a collapse toward the 39% CoT-free SFT baseline would support the claim that the SAE's learned features are doing the work.","supporting_citations":[],"review_version":1}