Technical Summary: VibeThinker-3B

Problem Statement

The field of large language models (LLMs) has increasingly relied on scaling parameter counts to achieve frontier-level reasoning capabilities, often concentrating advanced logical reasoning in models with tens or hundreds of billions of parameters. Consequently, Small Language Models (SLMs) with 3 billion parameters or fewer are frequently viewed as having inherent bottlenecks for complex mathematical derivations and programming tasks, serving primarily as cost-effective fallbacks rather than capable reasoning engines. While previous work (VibeThinker-1.5B) demonstrated that small models could produce stable chains of logic, the upper bound of their reasoning capacity remained unexplored. This report addresses the question: Can a strictly 3B parameter model achieve frontier-level performance comparable to top-tier, orders-of-magnitude larger LLMs on verifiable reasoning tasks?

Methodology

VibeThinker-3B is a 3B dense model built upon the Qwen2.5-Coder-3B base. It employs a systematic, staged post-training pipeline rooted in the Spectrum-to-Signal Principle (SSP), which was introduced in the authors' previous 1.5B work. The pipeline is designed to construct a broad solution space ("Spectrum") during supervised fine-tuning and amplify high-value reasoning signals ("Signal") during reinforcement learning.

1. Supervised Fine-Tuning (SFT)

The SFT stage utilizes a curriculum-based two-stage strategy:

• Stage 1 (Broad Coverage): Trains on a diverse, quality-filtered dataset covering mathematics, code, STEM, general dialogue, and instruction following. This establishes a stable cold-start policy.
• Stage 2 (Hard-Reasoning): Focuses on high-difficulty, long-horizon reasoning samples (filtered to >5K tokens and high error rates) to deepen logical derivation capabilities.
• Data Construction: Features automated query expansion, multi-path reasoning distillation (preserving diverse valid solution traces rather than single standard solutions), and rigorous multi-level quality control (N-gram filtering, LLM-based query assessment, and trace correctness verification).
• Diversity-Exploring Distillation: Intermediate checkpoints are evaluated on domain-specific probing sets to select specialists that maximize valid solution diversity (Pass@K) rather than just minimizing loss. These are merged to create a unified SFT model.

2. Reinforcement Learning (RL)

The core RL algorithm is MaxEnt-Guided Policy Optimization (MGPO), which dynamically weights prompts based on their proximity to the model's current capability boundary (where correctness is neither 0 nor 1).

• Multi-Domain RL: A sequential pipeline applies MGPO to Mathematics, Code, and STEM reasoning. Unlike previous iterations, this stage utilizes a single 64K long-context window from the start to preserve long-horizon reasoning trajectories, avoiding the degradation observed with progressive context expansion.
• Long2Short Math RL: A novel two-stage optimization for mathematics. The first stage optimizes for accuracy; the second stage optimizes for token efficiency by redistributing rewards among correct trajectories to favor shorter, concise reasoning paths without sacrificing accuracy.

3. Offline Self-Distillation & Instruct RL

• Offline Self-Distillation: High-quality reasoning trajectories elicited from Math, Code, and STEM RL stages are filtered and distilled back into the model. A "learning potential" score identifies traces that are correct but not yet well-modeled by the student, prioritizing these for distillation.
• Instruct RL: The final stage aligns the reasoning-enhanced model with complex user instructions using a mixed dataset of format-sensitive and open-ended prompts, employing rule-based validators and rubric-based reward models to ensure strict controllability.

4. Test-Time Scaling

The authors introduce Claim-Level Reliability Assessment (CLR), a test-time scaling strategy for answer-verifiable tasks. Instead of aggregating full reasoning traces, CLR extracts key decision-relevant claims, self-verifies them, and aggregates results based on a non-linear reliability score. This reduces noise and token consumption while boosting performance.

Key Results

VibeThinker-3B demonstrates exceptional parameter efficiency, achieving performance levels comparable to models with hundreds of billions of parameters on verifiable tasks.

• Mathematics:
  • AIME26: 94.3 (97.1 with CLR), matching DeepSeek V3.2 (671B) and Kimi K2.5 (1T).
  • HMMT25: 89.3 (95.4 with CLR).
  • IMO-AnswerBench: 76.4 (80.6 with CLR), outperforming significantly larger models like DeepSeek V3.2 (78.3) and GLM-5 (82.5).
• Coding:
  • LiveCodeBench v6: 80.2 Pass@1, surpassing all small/mid-sized baselines and trailing only top-tier systems like GPT-OSS-120B.
  • Out-of-Distribution (LeetCode): Achieved a 96.1% acceptance rate on recent unseen LeetCode contests (April–May 2026), outperforming GPT-5.2 and Kimi K2.5, and matching Gemini 3 Flash.
• Instruction Following:
  • IFEval: 93.4, confirming that reasoning enhancements do not compromise strict instruction controllability.
• Knowledge:
  • GPQA-Diamond: 70.2 (72.9 with CLR). The model trails top-tier generalists here, highlighting the distinction between reasoning and knowledge-intensive tasks.

Significance and Claims

The paper posits that VibeThinker-3B challenges the conventional wisdom that reasoning capacity is strictly linearly bounded by parameter scale. The authors introduce the Parametric Compression-Coverage Hypothesis, which distinguishes between two types of capabilities:

1. Parameter-Dense Capabilities (Compressible): Verifiable reasoning (math, code) relies on structured search, constraint satisfaction, and error correction. These can be compressed into a compact, reusable reasoning core, allowing small models to reach frontier performance.
2. Parameter-Expansive Capabilities (Coverage): Open-domain knowledge and general-purpose competence require broad coverage over facts, concepts, and long-tail scenarios, necessitating massive parameter scales.

Core Claim: The development of SLMs is not merely a passive compromise for deployment efficiency but a complementary evolutionary trajectory to traditional scaling laws. VibeThinker-3B serves as empirical evidence that a 3B model, when equipped with a rigorous post-training pipeline (diversity-driven SFT, multi-domain RL, and self-distillation), can enter the "frontier reasoning band" of systems 200x–300x larger in parameter count. The paper concludes that the boundary of reasoning capacity for compact models far exceeds conventional expectations, provided the training signals are reliable and the solution space is sufficiently explored.