Source-linked AI summary
VibeThinker-3B: Exploring the Frontier of Verifiable Reasoning in Small Language Models
Sen Xu, Shixi Liu, Wei Wang, Jixin Min, Yingwei Dai, Zhibin Yin, Yirong Chen, Xin Zhou, Junlin Zhang
TL;DR
Small language models are often thought to face limits on difficult reasoning because frontier capabilities typically require massive parameter scale. This report develops VibeThinker-3B through staged post-training and finds frontier-level performance across demanding verifiable reasoning tasks, while identifying a remaining gap on knowledge-heavy evaluation.
Problem
The report asks whether a strictly 3B-parameter model can overcome the presumed capability boundary of small language models and achieve frontier-level complex reasoning.
Method
VibeThinker-3B uses staged post-training combining diversity-oriented supervised fine-tuning, multi-domain reinforcement learning, offline self-distillation, and instruction alignment.
Results
VibeThinker-3B reaches a performance band comparable to frontier models across verifiable reasoning tasks, including 94.3 on AIME26 and 96.1% acceptance on recent LeetCode contests.
Takeaways & Limitations
The findings motivate a hypothesis that verifiable reasoning can be compressed into compact parameter-dense cores, whereas broad knowledge and general-purpose capabilities require wider parameter coverage.
Takeaways & Limitations
The model still trails the strongest large-parameter systems on knowledge-heavy benchmarks such as GPQA-Diamond.
Abstract
from arXiv · showhide
This technical report introduces VibeThinker-3B, a compact dense model with 3B parameters developed to investigate how far verifiable reasoning can be pushed within a strictly small-model regime. Building upon the Spectrum-to-Signal post-training paradigm, we systematically enhance the model through an optimized pipeline that includes curriculum-based supervised fine-tuning, multi-domain reinforcement learning, and offline self-distillation. Experimental evaluations demonstrate that VibeThinker-3B achieves frontier-level performance on highly demanding verifiable tasks. Specifically, it attains a score of 94.3 on AIME26 (improving to 97.1 with claim-level test-time scaling), an 80.2 Pass@1 on LiveCodeBench v6, and exhibits strong out-of-distribution generalization with a 96.1\% acceptance rate on recent unseen LeetCode contests. This effectively places it in the performance band of first-tier reasoning systems, matching or exceeding flagship models that are orders of magnitude larger, such as DeepSeek V3.2, GLM-5, and Gemini 3 Pro. Furthermore, a score of 93.4 on IFEval confirms that this extreme reasoning enhancement does not compromise strict instruction controllability. Extending our previous 1.5B work, these findings motivate the Parametric Compression-Coverage Hypothesis, which views verifiable reasoning as compressible into compact reasoning cores, while open-domain knowledge and general-purpose competence require broad parameter coverage over facts, concepts, and long-tail scenarios. This perspective suggests that compact models are not merely deployment-efficient substitutes, but a complementary path toward frontier-level performance in parameter-dense capability regimes.
1 Introduction
VibeThinker-3B investigates whether a strictly small model can achieve frontier-level verifiable reasoning through an upgraded post-training pipeline. Its results motivate a view that compact models can specialize in parameter-dense reasoning, while broad knowledge requires greater parameter coverage.
- Motivation: Frontier reasoning is commonly concentrated in models with tens or hundreds of billions of parameters, motivating investigation of small-model alternatives.The introduction contrasts parameter scaling for difficult reasoning tasks with the small-model regime.
- Motivation: VibeThinker-1.5B established stable basic chains of logic, but left the upper bound of small-model reasoning unexplored.The 3B study extends this earlier feasibility result toward stronger reasoning capability.
- Training approach: The upgraded Spectrum-to-Signal pipeline combines stronger synthesis and filtering, curriculum learning, multi-domain coverage, and harder long-horizon reasoning training.Curriculum learning first builds coverage across mathematics, code, STEM, dialogue, and instruction following before emphasizing difficult reasoning samples.
- Evaluation: 97.1 on AIME26 with claim-level reliability assessment and a 96.1% acceptance rate on recent LeetCode contests show gains from test-time scaling and out-of-distribution generalization.CLR also raises HMMT25 to 95.4 and BruMO25 to 99.2.
- Conceptual contribution: The Parametric Compression-Coverage Hypothesis and Reasoning-Knowledge Decoupling Paradigm distinguish compact, parameter-dense verifiable reasoning from parameter-expansive open-domain knowledge.The proposed perspective treats smaller models as specialized complements to large models rather than merely compute-efficient substitutes.
2 Methods
VibeThinker-3B is built from the Qwen2.5-Coder-3B foundation through a staged post-training pipeline grounded in the Spectrum-to-Signal Principle. The pipeline combines data synthesis, diversity-oriented supervised fine-tuning, multi-domain reinforcement learning, offline self-distillation, and instruction-oriented alignment, with an optimized two-stage curriculum strategy.
- Overall Pipeline: VibeThinker-3B starts from Qwen2.5-Coder-3B, a compact 3B dense foundation model.
- Overall Pipeline: The sequential post-training pipeline combines data synthesis, diversity-oriented supervised fine-tuning, multi-domain reinforcement learning, offline self-distillation, and instruction-oriented alignment.
- Overall Pipeline: The framework continues the Spectrum-to-Signal Principle as its organizing post-training paradigm.
- Supervised Fine-Tuning: The optimized SFT stage upgrades data synthesis and filtering while introducing a two-stage curriculum learning strategy.
2.1 Supervised Fine-tuning
The supervised fine-tuning phase builds a quality-controlled, multi-domain cold-start policy and expands reasoning coverage through synthesized queries and multi-path distillation. A curriculum-based two-stage procedure then progresses from broad capability coverage to harder, longer-horizon reasoning while preserving solution diversity.
- Dataset construction: The multi-domain supervised dataset covers math, code, STEM reasoning, general chat, and instruction following to establish a stable cold-start policy for later reinforcement learning.
- Data construction: Data synthesis expands query coverage by seeding only mathematical queries with explicit credible answers or rationales and programming queries with reliable tests or executable evaluation rules.
- Quality control: Multi-level quality control removes repetitive or contaminated samples, filters invalid queries, and stratifies the remaining data by reasoning-chain length and problem difficulty for curriculum SFT.
- Curriculum-based two-stage SFT: The two-stage curriculum begins with broad capability coverage and behavioral cold start, then shifts toward higher-difficulty, longer-horizon reasoning using a hard-reasoning subset filtered by length and difficulty.
- First-stage optimization: 5 × 10−5 initial learning rate, 8 × 10−8 minimum learning rate, 128 global batch size, 5 epochs, and 5% linear warmup define first-stage optimization.
2.2 Reinforcement Learning
VibeThinker-3B uses MGPO to concentrate reinforcement learning on prompts near the model’s capability boundary, applying it across mathematics, code, and STEM tasks with domain-specific verification. Its sequential curriculum first strengthens accuracy through long-form reasoning, then improves efficiency by favoring concise correct trajectories.
- MGPO weighting: MGPO emphasizes prompts with intermediate correctness, where correct and incorrect rollouts coexist near the model’s capability boundary.Prompts with near-zero correctness are too difficult, while near-one prompts are saturated; the weighting therefore prioritizes sufficient uncertainty.
- Multi-domain RL: MGPO is applied to mathematics, code, and STEM reasoning using final-answer verification, sandbox execution with test cases, and answer-plus-option verification, respectively.All domains use the same policy-optimization framework but rely on different reward sources and verification mechanisms.
- Sequential curriculum: The sequential Reasoning RL pipeline progresses from Math RL to Code RL and finally STEM RL, targeting symbolic derivation, executable logic, and generalized logical reasoning.The stages respectively emphasize long-horizon derivation and search, boundary-case handling and constraint satisfaction, and transfer of the underlying reasoning ability.
- Long2Short Math RL: VibeThinker-3B uses a two-stage Long2Short Math RL strategy that first optimizes accuracy with standard MGPO before introducing length-aware rewards for efficiency.The first stage allows fully unfolded reasoning and diverse solution paths; the second reshapes preferences among correct trajectories toward brevity.
- Long2Short Math RL: λ = 0.2 controls the maximum redistribution magnitude, while centered shifts preserve the group-level reward baseline and leave rewards unchanged when correct trajectories have equal length.The zero-sum redistribution changes relative preferences among correct solutions without systematically shifting the mean reward.
2.3 Offline Self-Distillation
Offline self-distillation extracts verified, high-quality reasoning trajectories from Math, Code, and STEM reinforcement-learning checkpoints, then distills them into a unified student model through supervised fine-tuning. Selection uses learning-potential scoring within domain-specific length buckets to emphasize traces that are correct but insufficiently modeled by the student.
- Offline trajectory distillation: Verified trajectories from Math, Code, and STEM RL checkpoints are distilled into a unified student model through supervised fine-tuning.Data filtering extracts high-quality offline reasoning patterns and enables more stable integration of multi-domain capabilities.
- Learning-potential Filtering: Rejection sampling with domain-specific verifiers removes incorrect trajectories before learning-potential scoring estimates each correct trace’s distillation value.The score is computed as the length-normalized negative log-likelihood of a verified teacher trajectory under the student model.
- Learning-potential Filtering: Higher learning-potential scores identify verified traces that are less well modeled by the student and therefore offer greater distillation value.This prioritizes correct reasoning patterns that can most improve the student model.
- Length-aware selection: Priorities are computed within domain-specific length buckets, excluding extremely short traces and filtering extreme high-score outliers.These safeguards reduce bias from sequence length, abnormal tokens, format errors, distributional shifts, and noisy samples.
2.4 Instruct RL
Instruct RL converts the reasoning-enhanced checkpoint into a more reliable user-facing model. It trains on diverse instruction data and applies specialized rewards for explicit constraints and open-ended prompts.
- Training objective: Instruct RL converts the reasoning-enhanced checkpoint into a more reliable user-facing model.The procedure is designed for user-facing reliability.
- Training data: The mixed instruction dataset includes format-sensitive prompts, long-context instructions, and general alignment examples.
- Reward design: For explicit constraints, rule-based validators reward format, ordering, item count, keyword constraints, and task completion.
- Reward design: For open-ended prompts, rubric-based reward models evaluate helpfulness.
3 Evaluation
VibeThinker-3B reaches first-tier performance on many verifiable reasoning tasks across mathematics, coding, and instruction following, with claim-level test-time scaling further improving answer-verifiable results. However, broad knowledge-heavy benchmarks remain a relative weakness, while recent LeetCode contests show strong coding generalization.
- Core benchmark performance: 94.3 on AIME26 and 91.4 on AIME25 establish strong competition-mathematics performance for VibeThinker-3B.It also reaches 93.8 on BruMO25, exceeding several much larger parameter models.
- Core benchmark performance: 80.2 on LiveCodeBench v6 and 93.4 on IFEval show that the gains extend beyond mathematics without compromising instruction controllability.VibeThinker-3B surpasses all models in Table 1 on LiveCodeBench v6 and also achieves 38.6 on OJBench and 74.5 on IFBench.
- Test-time scaling: 97.1 on AIME26 with CLR demonstrates that claim-level reliability assessment improves answer-verifiable mathematics without updating model parameters.CLR also raises AIME25 to 96.7, HMMT25 to 95.4, BruMO25 to 99.2, IMO-AnswerBench to 80.6, and GPQA-Diamond to 72.9.
- Limitations: 72.9 on GPQA-Diamond after CLR still trails the strongest large-parameter systems, exposing a clear limitation on knowledge-heavy evaluation.The reported improvement is from 70.2 to 72.9, contrasting with stronger performance on competition mathematics and executable coding.
- LeetCode generalization: 96.1% overall acceptance on recent LeetCode contests, from 123 of 128 first-attempt Python submissions, demonstrates strong out-of-distribution coding generalization.The evaluation covers weekly and biweekly contests from Apr. 25 to May 31, 2026, using Python-only one-shot generation.
4 Conclusion
VibeThinker-3B is presented as a compact 3-billion-parameter reasoning model that performs strongly on challenging verifiable-reasoning benchmarks and generalizes robustly to out-of-distribution LeetCode evaluations. These findings motivate the Parametric Compression-Coverage Hypothesis, which distinguishes compressed verifiable reasoning from scale-dependent broad capability coverage.
- 4 Conclusion: VibeThinker-3B, with only 3 billion parameters, performs strongly on AIME26, HMMT25, IMO-AnswerBench, and LiveCodeBench v6.The report also evaluates robust generalization on out-of-distribution LeetCode tasks.
- 4 Conclusion: VibeThinker-3B demonstrates robust generalization on out-of-distribution LeetCode evaluations.Together with its benchmark results, this places the model in a performance band comparable to representative frontier LLMs.
- 4 Conclusion: The Parametric Compression-Coverage Hypothesis proposes that verifiable reasoning is encoded through parameter-dense core compression.By contrast, open-domain knowledge and general-purpose capabilities depend more heavily on the broad coverage afforded by model scale.