{"id":"c7c01bf1-2ae4-41bc-9a1f-610b961c6c8c","arxiv_id":"2506.11752","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DART distills chain-of-thought reasoning into non-autoregressive silent-thought tokens, improving accuracy over prior non-autoregressive baselines with negligible latency increase.","lead":"DART trains large language models to answer reasoning questions directly from a small set of fixed 'silent thought' tokens, skipping the step-by-step chain of thought text. On math benchmarks it beats other non-autoregressive reasoning methods with only about one millisecond of added latency.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-position distillation (L_distill on z_N) is the core transfer mechanism, yet it is only validated on short-chain benchmarks; its sufficiency for longer reasoning chains is untested, and the paper's Limitations flags it as potentially suboptimal.","rationale":"The reader identified the same weakest assumption (single-position alignment sufficiency), and I concur that it is the most load-bearing because the method's novelty and the central claim both rest on the sufficiency of the z_N supervision. The other concerns (sourced baselines, no error bars, C selected on GSM8K) are experimental-hygiene issues that would qualify but not overturn the empirical comparison. The proposed test is a direct falsification check: it creates a stress condition (longer chains) where the limited supervision should fail if the concern lands. I recommend keeping the verdict CONDITIONAL; the paper should add this analysis or an equivalent evaluation on longer-chain benchmarks before making broad generalization claims. Since the reader's verdict already reflects this risk, no verdict change is needed.","tokens_in":10262,"tokens_out":10465,"duration_ms":97112,"concrete_test":"Split the GSM8K test set into two subsets using the ground-truth CoT annotations: short-CoT (≤5 reasoning steps) and long-CoT (>5 steps). Run the released DART checkpoint and the No-CoT baseline on both subsets, recording accuracy on each. If the DART-minus-No-CoT gap is significantly smaller on long-CoT examples, or if DART's absolute accuracy drops sharply relative to short-CoT, then the single-position alignment is a real bottleneck and the generalization claim fails; if the gap is preserved, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DART's central claim is that aligning the hidden state of the last separator token z_N between the CoT and ST pathways (Section 3.2, L_distill) is sufficient to transfer autoregressive reasoning into non-autoregressive silent thought. This single-position supervision is the only 'reasoning' channel: the ST tokens receive no token-level or step-level objective, and the model is free to satisfy L_distill by encoding short-range, task-specific features. The paper's Limitations section concedes that the method 'aligns only the activation value of the last word before the answer... may be suboptimal, as it can overlook some information in intermediate tokens.' The empirical validation in Table 1 covers GSM8K, GSM-HARD, SVAMP, and MultiArith, all of which have short CoT traces (typically 2-5 arithmetic steps). No experiment tests tasks with longer, multi-step reasoning chains (e.g., MATH, AIME, or GSM8K examples with more than 5 reasoning steps). If a substantial fraction of reasoning information is carried in intermediate tokens, single-point alignment will underfit long chains, making DART's gain on GSM8K a consequence of capacity or dataset-specific shortcuts rather than robust internalized reasoning. Because the value proposition is that DART is a general efficient-reasoning alternative, this untested sufficiency is the load-bearing risk.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DART, a self-distillation framework that trains a decoder-only LLM with two pathways: a chain-of-thought (CoT) pathway and a silent-thought (ST) pathway in which a fixed number of <st> tokens are inserted before an answer prompt. A lightweight Reasoning Evolvement Module (REM) adapts the key and value projections, and an L1 distillation loss aligns the hidden state of the last separator token between the two pathways. At inference, only the ST pathway is used, so no autoregressive intermediate tokens are generated. Experiments on GSM8K, GSM-HARD, SVAMP, MultiArith, ProsQA, and CommonsenseQA-CoT, plus results with GPT2, Qwen2.5-1.5B, and Llama-3.2-3B, report accuracy and latency. The paper includes ablations, a sensitivity study of the number of ST tokens, a qualitative analysis of decoded ST tokens, and a Limitations section.","tokens_in":10546,"tokens_out":8472,"duration_ms":80235,"significance":"If the reported results are reliable, DART offers a simple and potentially practical way to compress chain-of-thought reasoning into a fixed number of non-autoregressive latent tokens, with a favorable accuracy-latency trade-off on short-chain benchmarks. The REM module is a clean plug-in design with few additional parameters, and the ablations help localize the contribution of each component. The authors also make an honest attempt at out-of-distribution evaluation and include multiple base models. However, the evidence currently supports only short-chain arithmetic and commonsense tasks; the claimed generality of the approach, and the exact position of DART relative to continuous-thought baselines such as CODI and Coconut, are not fully established by the experiments as presented.","major_comments":[{"comment":"The sentence \"DART achieves the best performance among all NAR baselines on GSM8K\" is not decidable from the manuscript as rendered. In §2, Coconut and CODI are described as limited by the autoregressive pattern, which implies they are not NAR baselines, yet Table 1 lists them among the compared methods and its \"Is NAR?\" column carries no visible values in the text. If Coconut and CODI are counted as non-autoregressive, DART's 42.6% is below Coconut's 50.6% and CODI's 55.6%, contradicting the claim. If they are excluded, the claim should say so explicitly, because the abstract's phrase 'existing non-autoregressive baselines' is then misleading. Please clarify the taxonomy and the exact scope of the claim.","section":"§4.2, Table 1"},{"comment":"The CODI and iCoT accuracies are marked as taken from Shen et al. (2025) rather than re-run under the same training conditions, and no error bars or multiple-seed results are reported for any method. Since the central quantitative claim is a 10.1-point gain over No-CoT, a single run without variance cannot establish that this difference is significant. Please report multiple seeds with standard deviations, or at least state clearly that the comparison is a single-run qualitative indicator. If the CODI and iCoT checkpoints were not produced with the same base model, dataset, and training epochs as DART, the comparison in Table 1 is not controlled.","section":"§4.2, Table 1"},{"comment":"The only distillation signal used to transfer reasoning is L_distill on the hidden state of the last separator token z_N. The paper's own Limitations section concedes that this single-position alignment may overlook information in intermediate tokens. No experiment tests tasks with longer reasoning chains: all math benchmarks have short CoT traces of roughly two to five arithmetic steps, and ProsQA and CommonsenseQA-CoT are similarly short. Since the introduction and abstract frame DART as a general alternative to CoT for complex tasks, the sufficiency of single-position alignment for longer chains should be demonstrated, for example with a dataset such as MATH or with a breakdown of GSM8K by number of reasoning steps. Without such evidence, the generalization claim is not supported by the current experiments.","section":"§3.2, Limitations"},{"comment":"The sensitivity analysis for the number of ST tokens C is performed on GSM8K, which is the same benchmark used for the main comparison and for reporting the headline accuracy. This risks selecting hyperparameters on the test set. The analysis should be conducted on a held-out development set or clearly described as model selection rather than as evidence of robustness. The same concern applies to the fixed choices λ = 20, d = 128, and α = 32, which are reported without sensitivity checks.","section":"Figure 2, §4.2"}],"minor_comments":[{"comment":"There are several typos: 'SV AMP' should be 'SVAMP' in Table 1 and §4.1; 'DARt' appears in §4.2; and 'Ldistll' appears in the overall objective LDART = LCoT + LST + λLdistll instead of 'Ldistill'.","section":"Throughout"},{"comment":"The derivation drops the softmax and scaling factor and then continues as if the simplified expression were exact. Please state explicitly that this is a rough approximation intended only to motivate the REM design, and that no formal error bound is claimed.","section":"Appendix A"},{"comment":"The statement that 69.9% of translated ST tokens match ground-truth CoT words needs a precise definition of 'match': exact token identity, case-insensitive, stemmed, and whether the comparison is position-wise or set-based. Without this definition the number is hard to interpret.","section":"§4.3, Qualitative Analysis"},{"comment":"The CoT inference time on CommonsenseQA-CoT (1471 ms) is much larger than on ProsQA (882 ms) despite similar benchmark sizes; a brief description of the decoding protocol (e.g., max new tokens, batch size, and whether times include prompt processing) would help the reader interpret the latency numbers.","section":"§4.2, Table 4"},{"comment":"The implementation details say that the configuration is 'consistent with the configuration used for LoRA' to fine-tune the CoT pathway, but it is not clear which parameters are trainable in each pathway and whether REM replaces or supplements the LoRA adaptation. Please specify the LoRA rank, target modules, and the exact role of REM relative to LoRA for the ST pathway.","section":"Appendix C"}],"recommendation":"major_revision","confidential_remarks":"The central idea is interesting and the paper is clearly written, but the main comparative claim needs disambiguation relative to continuous-thought baselines, and the single-seed, short-chain evidence does not yet support the generality claimed in the abstract. The paper would be strengthened by a controlled comparison with CODI/Coconut under identical settings, multiple runs, and at least one long-chain benchmark."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a solid, honest paper about a new non-autoregressive reasoning method. It deserves a serious referee, but do not overread the generalization claims — the transfer mechanism is a single-position hidden-state alignment, and the paper's own Limitations section flags that it may miss information in intermediate tokens.\n\nWhat is actually new: DART replaces autoregressive continuous thoughts (Coconut, CODI) with a fixed set of non-autoregressive <st> tokens, and uses a low-rank REM to align the hidden state of the last separator token with the CoT pathway. That specific combination is new. The experiments are broad: four math benchmarks, two other reasoning tasks, three base models, ablations, latency/FLOPs tables. The ablation study is genuinely useful — it shows that distilling on z_N works better than on answer tokens or the whole sequence, which is evidence the design choice is not arbitrary. The qualitative decoding of ST tokens (69.9% match to CoT words) is a nice touch, even if it is only suggestive.\n\nSoft spots: the headline 10.1% gain is over No-CoT, not over a strong reasoning baseline; the gap to autoregressive CoT (42.6 vs 58.8 on GSM8K) remains large. The CODI and iCoT accuracies are taken from a prior paper rather than re-run under identical training conditions, so the comparison is not fully controlled. There are no error bars or multiple seeds. The choice of C (number of ST tokens) is tuned on GSM8K, the same benchmark as the headline result. The theoretical derivation is approximate and is used only as motivation, which is fine. The stress-test concern — that single-position alignment may not transfer to longer reasoning chains — is real but untested; the paper does not evaluate on tasks with longer CoTs, and the Limitations concede the point.\n\nWho this is for: people working on efficient reasoning for LLMs in latency-sensitive deployments, and researchers studying implicit reasoning in hidden states. It is a useful data point, not a breakthrough.\n\nRecommendation: send to peer review. Ask for re-run of the missing baselines, multiple seeds, and at least one evaluation with longer reasoning chains. The core idea is worth the referee time.","headline":"A solid, honest non-autoregressive reasoning paper with a useful new method, a narrow validation, and a clear need for more controlled baselines and longer-chain tests.","tokens_in":11103,"tokens_out":3705,"would_cite":true,"duration_ms":34961,"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":"DART distills chain-of-thought reasoning into silent thought tokens, achieving the best non-autoregressive accuracy on GSM8K with only about 1 ms extra latency.","keywords":["non-autoregressive reasoning","silent thought","self-distillation","chain-of-thought","hidden-state alignment","efficient inference","math reasoning","LLM fine-tuning"],"falsifier":"A decisive check: train DART on GSM8K-Aug but attach the distillation loss to a random token instead of the final separator $z_N$, then evaluate on GSM-HARD. If the 10.1-point GSM8K gain persists, the stated mechanism is not the active one; if the out-of-distribution accuracy collapses to the No-CoT level, then aligning only $z_N$ is insufficient to generalize.","tokens_in":1693,"feed_emoji":"⚡","tokens_out":2176,"duration_ms":96588,"temperature":0.7,"pith_summary":"Chain-of-thought reasoning makes LLMs accurate but slow because every reasoning step is generated one token at a time. DART proposes to distill that step-by-step process into a short sequence of silent thought tokens that the model processes in a single non-autoregressive pass, producing the answer directly. The authors show that aligning the hidden state of the last separator token between the chain-of-thought and silent-thought pathways transfers most of the reasoning gain. On GSM8K, DART beats all other non-autoregressive baselines by 10.1 accuracy points at essentially the same inference latency as direct answering. A sympathetic reader would take this as evidence that implicit, non-autoregressive reasoning can be a practical alternative to explicit chain-of-thought.","feed_headline":"Silent thought replaces chain-of-thought at near-zero latency cost","feed_subtitle":"Turns chain-of-thought into silent tokens; beats fast baselines on GSM8K by 10.1%.","key_machinery":"The central mechanism is the Reasoning Evolvement Module (REM), a plug-in set of low-rank key and value projection modifications paired with a distillation loss. At each decoder layer, REM adapts the attention keys and values for the sequence consisting of the question and the silent tokens, and the distillation loss forces the hidden state of the last separator token in the silent-thought pathway to match that of the chain-of-thought pathway. This makes the silent tokens carry an approximation of the hidden-state shift that the chain-of-thought intermediate tokens would have produced. The shared separator tokens, fixed to the answer prompt, give the alignment a stable anchor.","core_discovery":"The paper claims that an LLM can internalize autoregressive chain-of-thought reasoning into non-autoregressive silent thought by training two pathways at once. In the CoT pathway the model predicts the intermediate reasoning steps and the answer; in the silent thought pathway the model receives a fixed number of special tokens after the question and produces only the answer. A lightweight Reasoning Evolvement Module inserted at each decoder layer rewrites the attention key and value projections so that the hidden state of the last separator token in the silent pathway converges to the corresponding hidden state in the CoT pathway. Once trained, only the silent pathway is active at inference, and the silent tokens evolve into informative embeddings as they pass through the layers. The authors report the best accuracy among non-autoregressive baselines on GSM8K, consistent gains on out-of-distribution math benchmarks, and improvements across several decoder-only model sizes.","pith_inferences":["The authors do not test this, but the same single-position alignment should transfer to other tasks that have a stable separator token before the answer; if so, DART becomes a general recipe for silent reasoning rather than a math-specific trick.","Because the decoded silent tokens mostly match ground-truth chain-of-thought words when the answer is correct, DART could be used as a probe: the interpretability of silent tokens may serve as a diagnostic for whether distilled reasoning is actually being used rather than memorized.","A natural pressure test, which the paper names but does not run, is to generate chain-of-thought data with a larger model and distill it into a smaller one; if the gain survives, DART could act as a reasoning-preserving model-compression tool.","The reported latency is single-query latency on one GPU; an extension to batched serving and throughput would clarify whether the near-zero overhead persists at scale."],"forward_implications":["If DART works as claimed, an LLM can answer reasoning questions in roughly the same wall-clock time as direct answering while keeping a large share of chain-of-thought accuracy.","Since no autoregressive reasoning tokens are emitted during inference, per-query compute on math benchmarks drops to the cost of one forward pass over the question and silent tokens.","The method transfers across model scales, with accuracy gains reported on small and larger decoder-only models, not just the primary 1-billion-parameter model.","Because only the silent pathway is used at inference, DART can be added to existing decoder-only LLMs as a fine-tuned plug-in with a small number of additional parameters.","Gains on out-of-distribution benchmarks such as GSM-HARD, SVAMP, and MultiArith suggest the distilled reasoning behavior generalizes beyond the exact training distribution."],"supporting_citations":[{"why":"Supplies the hidden-state shift analysis that justifies aligning the last separator token's hidden state between the two pathways.","marker":"Dai et al. 2023"},{"why":"Defines chain-of-thought prompting, the autoregressive reasoning process that DART distills and the primary baseline it replaces.","marker":"Wei et al. 2022"},{"why":"Supplies GSM8K, the primary math benchmark whose augmented chain-of-thought version is used for training and evaluation.","marker":"Cobbe et al. 2021"},{"why":"Provides Coconut, the continuous-thought baseline whose autoregressive latency DART removes and whose GSM8K accuracy it exceeds.","marker":"Hao et al. 2024"},{"why":"Provides CODI, the one-stage distillation baseline DART compares against and the source of some reported baseline numbers.","marker":"Shen et al. 2025"},{"why":"Provides iCoT, an implicit-chain-of-thought baseline that reasons in hidden space and is outperformed by DART.","marker":"Deng et al. 2024"},{"why":"Provides PauseFT, the pause-token baseline DART improves on at the same non-autoregressive latency.","marker":"Goyal et al. 2024"},{"why":"Supplies the Llama-3.2 model family used for the primary experimental results and cross-model robustness checks.","marker":"Dubey et al. 2024"}],"fun_headline_variants":["Distill chain-of-thought into silent tokens for faster reasoning","Silent thought tokens replace chain-of-thought at lower cost","LLM distills reasoning into silent thought to slash latency","Non-autoregressive silent thought beats CoT on speed"],"cache_read_input_tokens":13184,"weakest_assumption_plain":"The load-bearing premise is that matching the hidden state of the single final separator token between the chain-of-thought and silent-thought pathways transfers the whole reasoning process; the paper's own limitations section acknowledges that this one-position supervision may miss information in intermediate tokens.","fun_headline_variants_meta":{"raw":{"variants":["Distill chain-of-thought into silent tokens for faster reasoning","Silent thought tokens replace chain-of-thought at lower cost","LLM distills reasoning into silent thought to slash latency","Non-autoregressive silent thought beats CoT on speed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000578,"raw_usage":{"total_tokens":2712,"prompt_tokens":918,"completion_tokens":1794,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":1725}},"tokens_in":534,"tokens_out":1794,"duration_ms":13505,"temperature":1.0,"reasoning_tokens":1725,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:03:34.814692+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check: train DART on GSM8K-Aug but attach the distillation loss to a random token instead of the final separator $z_N$, then evaluate on GSM-HARD. If the 10.1-point GSM8K gain persists, the stated mechanism is not the active one; if the out-of-distribution accuracy collapses to the No-CoT level, then aligning only $z_N$ is insufficient to generalize.","supporting_citations":[{"cited_title":"Chi, Quoc V","cited_arxiv_id":null,"evidence_quote":"Defines chain-of-thought prompting, the autoregressive reasoning process that DART distills and the primary baseline it replaces."}],"review_version":1}