TL;DR
Reasoning models often continue "thinking" long after they have internally arrived at the correct answer—a phenomenon the authors call "slack." By applying a simple self-training recipe that prioritizes the shortest correct reasoning traces, Concise-9B achieves the same accuracy as its base model (Qwen3.5-9B) while using 34% fewer tokens.
What Problem Does This Paper Address?
Modern "thinking" models (like DeepSeek-R1 or the Qwen-Think series) use test-time scaling: they generate long chains of thought to improve accuracy on complex tasks. However, this "thinking" is expensive. Previous research suggested that these models often "overthink," continuing to generate tokens even after the problem is solved.
The authors set out to quantify exactly how much of a reasoning trace is "slack" (tokens generated after the final answer is already fixed) and determine if a model can be trained to eliminate this waste without sacrificing its reasoning capabilities.
Methods
The researchers developed a recipe called concise self-training, which they applied to Qwen3.5-9B. The process involves:
- On-policy Rejection Sampling: The model generates multiple reasoning traces for a set of training problems.
- Shortest-Correct Selection: Instead of just keeping any correct answer, the fine-tuning process specifically selects the two shortest correct samples per problem.
- Answer-Synthesis Targets: To maintain robustness, 30% of the training data consists of "synthesis passes," where the model is shown multiple previous (potentially conflicting) attempts and asked to re-derive the final answer.
- Forced Readouts (Measurement): To measure slack, the authors used a "forced readout" technique. Every 2,048 tokens during a thinking trace, they forced the model to stop and provide a final answer. This allowed them to identify the "fixed point"—the moment after which every subsequent readout yielded the same final answer.
Datasets
The model was trained on a pool of 8,252 verifiable problems from MATH, GSM8K, MMLU, and GPQA.
Evaluation was performed on:
- Development Suite: 498 problems from GPQA-Diamond, MATH, and MMLU.
- Held-out Sets: A massive collection of 8,983 problems, including high-level competition math (AIME, HMMT), SuperGPQA, and MMLU-Pro.
Key Insights and Ideas
- Slack is massive: In the base Qwen3.5-9B model, a staggering 54% to 64% of thinking tokens are generated after the answer has already stabilized.
- Training beats inference-time rules: You might think you could just stop the model early at inference time (e.g., "stop if the last three internal answers agree"). However, the authors found this "hindsight" approach is brittle and causes a significant drop in accuracy (up to 9.2 points). Training the model to be concise is far more effective.
- Efficiency through speed, not just cutting: Concise training doesn't just truncate the end of a trace; it actually makes the reasoning more efficient. The model reaches the correct answer earlier in the trace than the base model does.
- The "Cheap Commit" Signal: The stabilization of forced readouts provides a reliable signal for "adaptive sampling." If a model's internal answer hasn't changed across the last two checkpoints, it's a safe bet to stop, saving compute without needing a second full sample.
Key Results
- 34% Token Savings: At a 32K thinking budget, Concise-9B generates only 0.66x the tokens of the base model while maintaining equivalent accuracy (within ±1 point).
- Maintained Accuracy: On the development suite, single-sample accuracy and majority voting remained equivalent to the base model despite the drastically reduced token count.
- Slack Reduction: Concise training removed 86% of the targeted slack. In Concise-9B, the "after-fixed" portion of the trace dropped to 41-48%.
Limitations and Future Work
- The Performance Ceiling: While Concise-9B is excellent for most tasks, it loses some "ceiling" on the absolute hardest competition math problems. When allowed to think up to 80K tokens, the base model eventually finds 25 more correct answers (out of 1,276) that the concise model misses. This suggests that for the most extreme edge cases, "verbose" thinking still has value.
- Base Model Verbosity: The recipe's success depends on the base model. When replicated on Olmo-3-7B-Think (which is already quite terse), the token savings were much smaller (1–4%).
- Measurement Resolution: The slack measurement is conservative because it relies on a 2,048-token grid; the actual answer might fix even earlier between those points.
Future work could explore more granular stopping signals or training methods that preserve the high-end reasoning "ceiling" while still eliminating slack on easier problems.
Summary updated