REVIEW 4 major objections 5 minor 2 cited by
TL;DR: Too Long, Do Re-weighting for Efficient LLM Reasoning Compression
T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read The paper claims that dynamically re-weighting short and long chain-of-thought data during post-training can cut a reasoning model's output length by roughly 40 percent on math benchmarks while keeping accuracy about the same.
desk verdict Dynamic re-weighting of short/long CoT data is a genuinely useful idea with strong empirical support, but the unreferenced AdaR1 short-CoT construction and the averaged compression headline need attention before the claim holds. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is Algorithm 1, a long-to-short dynamic re-weighting loop. At each interval it computes benefits $\lambda_{\text{sys-1}}$ and $\lambda_{\text{sys-2}}$ from the distances between the current proxy model and two reference bounds: the negative average token count of the short-chain-of-thought model (an efficiency upper bound) and the accuracy of the long-chain-of-thought model (a capability upper bound). It updates mixture weights $\alpha_t$ with an exponential rule, renormalizes with smoothing, trains the model on the re-weighted mixture, and repeats; the average of the per-step weights is returned. This lets the data ratio itself adapt to whether the model is currently over-thinking or under-thinking.
What would settle it
Run the TLDR pipeline on DeepSeek-R1-Distill-Qwen-7B with the short chain-of-thought data built by an explicitly described replication of Appendix A.3 and all other settings fixed; the central claim is falsified if average token compression drops well below the reported roughly 40 percent or if accuracy on AIME or AMC falls by more than a few points relative to the original model.
Extended reading notes
Core claim
The paper's central discovery is that short and long chain-of-thought data induce divergent optimization directions, but they can be reconciled by treating their mixture ratio as a trainable parameter rather than a fixed hyperparameter. The proposed TLDR pipeline starts from a long-chain-of-thought reasoning model, constructs System-1 data from short correct solutions to simple GSM8K-style problems and System-2 data from long correct solutions to hard problems, then iteratively updates both model weights and the sampling weights of the two data streams. The re-weighting is driven by two measured gaps: how far the current model's token length is from the short-chain-of-thought model's efficiency, and how far its accuracy is from the long-chain-of-thought model's accuracy. The final checkpoint is chosen for short length subject to a validation-accuracy constraint, yielding models that generate roughly 150 to 250 tokens on simple problems instead of 550 to 770 tokens while retaining most of their accuracy on harder AIME and AMC questions.
Load-bearing premise
The load-bearing premise is that the undeclared AdaR1-style construction—fine-tuning on ten hand-annotated GSM8K answers—produces a short chain-of-thought signal that generalizes across difficulty levels, and if that construction cannot be reproduced or does not transfer, the reported compression and accuracy results would not follow.
Editorial extensions
If this is right
- On DeepSeek-R1-Distill-Qwen 7B the method reports a 44.9 percent average compression ratio with average accuracy moving from 72.2 to 72.8, and on the 14B model a 35.8 percent ratio with average accuracy moving from 71.7 to 73.5.
- Short chain-of-thought built from easy problems compresses token use on hard problems too, so the efficiency gain transfers across difficulty levels rather than only to the training distribution.
- A small proportion of long chain-of-thought from hard problems prevents the accuracy drop caused by short chain-of-thought, while long chain-of-thought from easy problems does not provide the same protection.
- At matched token budgets, the re-weighted model keeps higher accuracy than a budget-controlled reinforcement-learning baseline, suggesting the compression comes from removing redundant deliberation rather than from a hard length cutoff.
- The pipeline needs no hand-built length-diverse datasets or model interpolation; it only samples short and long chain-of-thought and adapts their ratio.
Reading between the lines
- A direct extension would be to test the same dynamic re-weighting on non-math long-chain-of-thought tasks such as code generation, planning, or agent trajectories, where the main obstacle is obtaining a trustworthy short-chain-of-thought signal rather than the re-weighting rule itself.
- A natural next step is instance-level routing: because the reported compression is much larger on easy problems than on hard ones, the learned mixture could inform per-question reasoning budgets instead of a single global checkpoint.
- The quantitative results should be re-measured once the unreferenced short-chain-of-thought construction in Appendix A.3 is replaced with an openly described one; until then, the strongest reading of the roughly 40 percent figure is conditional on that construction being reproducible and generalizing.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes TLDR, a post-training method for compressing the reasoning traces of long-chain-of-thought (CoT) models. The method constructs two data sources — System-1 (short CoT on easy GSM8K-like problems) and System-2 (long CoT on hard s1-style problems) — and dynamically re-weights the mixing ratio during fine-tuning using multiplicative weight updates, guided by validation-set accuracy and token usage. Experiments on DeepSeek-R1-Distill-7B/14B across ASDiv, GSM8K, MATH500, AIME, AMC, and MinervaMath report an average token compression of roughly 40% while approximately maintaining accuracy, and the paper includes comparisons with prompt-based, model-merging, and reward-based baselines, plus ablations on data difficulty and analyses of reasoning patterns.
Significance. If the central claim holds, TLDR offers a practical and relatively simple way to reduce overthinking in o1-like reasoning models, with direct implications for inference cost and latency. The paper has notable strengths: it evaluates at two model scales on several benchmarks, includes a diverse set of baselines, provides ablations that isolate the contribution of System-1 and System-2 data difficulty, and gives a qualitative analysis of trimming exploratory/reflective tokens. However, the evidence as presented is weakened by a reproducibility gap, a checkpoint-selection criterion that is far more permissive than the paper's accuracy-maintenance claim, and an average-based headline that conceals much smaller compression on hard benchmarks. These issues are correctable within the manuscript's scope, so the central idea is defensible but requires substantive revision.
major comments (4)
- [Appendix A.3] The System-1 data construction is the sole source of the concise reasoning signal that drives compression, yet it relies on an unreferenced method: "we adopt the short CoT construction method from AdaR1" and then annotate only 10 GSM8K questions. Since AdaR1 is not cited or described, the pipeline cannot be reproduced or audited from the paper alone, and the generalization of compression to hard benchmarks could be sensitive to those 10 examples. Please provide a full description or citation for AdaR1, and ideally quantify the sensitivity of the final compression and accuracy to the number and choice of annotated examples. This is load-bearing because the abstract's claim that the method "does not rely on sophisticated data annotations" is in tension with hand-annotating short answers.
- [Appendix A.1] The checkpoint selection criterion states: "We ultimately select the checkpoint with the shortest token length among those whose accuracy on the validation set is no less than 30% of that achieved by the original long CoT." This threshold permits checkpoints that have lost up to 70% of original validation accuracy. Such a criterion is inconsistent with the paper's central claim of "maintaining the accuracy of the reasoning." Because the final checkpoint is chosen for maximal token reduction subject only to this weak bound, the test-accuracy results in Table 1 may reflect selection for compression rather than a genuine accuracy-efficiency trade-off. The selection rule should require accuracy within a meaningful margin of the original model (e.g., statistical equivalence or a small bounded drop), and the final selected checkpoint's validation accuracy should be reported.
- [Sections 4.1 and Table 1] The headline "nearly 40%" compression is the A.C.R. averaged over six benchmarks, but per-benchmark compression varies dramatically: for the 7B model, TLDR achieves roughly 81% compression on ASDiv but only about 6.6% on AIME. The abstract and introduction should state this heterogeneity explicitly, and the main tables should include confidence intervals or variance estimates for both accuracy and token length, given that accuracy is estimated from 4–8 samples per question. Without such quantification, the reported average is difficult to interpret, especially on the small AIME set, and the reader cannot assess whether the maintained accuracy is within sampling noise.
- [Section 3.1, Algorithm 1] The algorithm's final line returns "1/T sum_{t=1}^T alpha_t", which is a time-averaged vector of mixture weights, not a trained model or a checkpoint. The text around the algorithm describes selecting a final model, but the pseudocode does not specify how this average weight vector is used, nor how the proxy model parameters are returned. This appears to be a typographical or formal error in the algorithm specification; please correct it and clarify the actual return value (e.g., the final proxy model parameters or a selected checkpoint).
minor comments (5)
- [Abstract and §1] The claim that the method "does not rely on sophisticated data annotations" is contradicted by the hand-annotated 10 short answers in Appendix A.3; please qualify the claim to distinguish between large-scale annotation pipelines and the minimal 10-example seed used here.
- [References] The method name "AdaR1" appears in Appendix A.3 but is not cited or listed in the references; either add the reference or rename the method and describe it fully.
- [Appendix A.1] The training details mention "For the 32B model, we performed SFT on five 8-GPU nodes," but Table 1 reports only 7B and 14B results; please either report the 32B results or remove the mention to avoid confusion.
- [Section 3.2, Eqs. (7)–(8)] The notation θproxy, θs, and θl is used in the λ definitions but is not formally introduced before these equations; please define these models explicitly (e.g., the current proxy model, the System-1 fine-tuned model, and the original long-CoT model).
- [Table 3] The averaged accuracy column in the 7B ablation (first block) reports 60.1 for the original model, while other tables report 69.3 for the same model; please check whether this is a transcription error or a different aggregation over the four benchmarks and make the aggregation consistent across tables.
Circularity Check
No significant circularity; the ~40% compression claim is an empirical result on held-out benchmarks. Minor background self-citations are not load-bearing, but the unreferenced AdaR1 short-CoT construction in Appendix A.3 is a reproducibility caveat.
full rationale
The paper's headline claim, nearly 40% token reduction while maintaining accuracy, is an empirical measurement reported in Table 1 for DeepSeek-R1-Distill-7B/14B on held-out benchmarks such as ASDiv, GSM8K, MATH500, AIME2024, AMC, and MinervaMath, not a quantity derived tautologically from the training objective. Equations (5)-(6) define the System-2 accuracy bound as the original long-CoT model's accuracy and the System-1 token bound as the short-CoT model's token length; these serve as design references for the dynamic re-weighting, and the final result is obtained by training under Algorithm 1, selecting a checkpoint by the validation rule in Appendix A.1, and then evaluating on separate test benchmarks. The central evaluation is therefore external to the fitted bounds, and I find no self-definitional or fitted-input-called-prediction step. The paper cites its own survey [28] and RedStar [50] for background System-1/System-2 framing and cognitive-pattern analysis; neither citation carries the central compression claim, so these are minor, non-load-bearing self-citations. One flagged limitation is missing support: Appendix A.3 states, "For short CoT, to avoid inconsistencies in the system prompt format, we adopt the short CoT construction method from AdaR1. We annotate 10 randomly selected questions from GSM8K using the instruct model, then fine-tune the long CoT model to overfit on them," but "AdaR1" does not appear in the reference list. This makes the System-1 data construction difficult to audit and is a reproducibility risk, but it is not circular because the reported compression and accuracy are measured on held-out benchmarks rather than derived from that construction. Overall, the method paper is self-contained as an empirical study; no step reduces to its own inputs by construction.
Assumptions & free parameters
free parameters (5)
- alpha smoothing parameter c =
1e-4
- alpha learning rate eta =
not reported
- initial mixture alpha_0 =
(1/2,1/2)
- validation accuracy threshold =
30% of original
- number of annotated short CoT examples =
10
assumptions (3)
- domain assumption Validation accuracy and token length on AIME 1983-2023 are a reliable proxy for tuning the data ratio that generalizes to the test benchmarks.
- domain assumption Fine-tuning the long CoT model on 10 annotated short answers (the AdaR1 method) induces a robust System-1 behavior that can be used to generate training data.
- ad hoc to paper Multiplicative weight updates with the given lambda definitions converge to a good trade-off in practice.
Cite this review
Pith. "Pith review of TL;DR: Too Long, Do Re-weighting for Efficient LLM Reasoning Compression." pith.science (2026). https://pith.science/paper/KQUOS7KN
@misc{pith2026250602678,
author = {Pith},
title = {Pith review of: TL;DR: Too Long, Do Re-weighting for Efficient LLM Reasoning Compression},
year = {2026},
howpublished = {\url{https://pith.science/paper/KQUOS7KN}},
note = {Machine review of arXiv:2506.02678}
}
read the original abstract
Large Language Models (LLMs) have recently achieved remarkable progress by leveraging Reinforcement Learning and extended Chain-of-Thought (CoT) techniques. However, the challenge of performing efficient language reasoning--especially during inference with extremely long outputs--has drawn increasing attention from the research community. In this work, we propose a dynamic ratio-based training pipeline that does not rely on sophisticated data annotations or interpolation between multiple models. We continuously balance the weights between the model's System-1 and System-2 data to eliminate redundant reasoning processes while preserving the model's reasoning capability. We validate our approach across models on DeepSeek-R1-Distill-7B and DeepSeek-R1-Distill-14B and on a diverse set of benchmarks with varying difficulty levels. Our method significantly reduces the number of output tokens by nearly 40% while maintaining the accuracy of the reasoning. Our code and data will be available soon.
Forward citations
Cited by 2 Pith papers
-
SwS: Self-aware Weakness-driven Problem Synthesis in Reinforcement Learning for LLM Reasoning
SwS uses failures during RL training to synthesize targeted math problems, improving reasoning accuracy on eight benchmarks.
-
Towards Concise and Adaptive Thinking in Large Reasoning Models: A Survey
A comprehensive review that categorizes methods for shortening and adaptively triggering chain-of-thought reasoning in large language models.
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.