Summary (Overview)

• VibeThinker-3B is a compact 3B-parameter dense model that achieves frontier-level performance on verifiable reasoning tasks (mathematics, coding, STEM), matching or exceeding models orders of magnitude larger (e.g., DeepSeek V3.2 at 671B, GLM-5 at 744B, Gemini 3 Pro).
• Key results: 94.3 on AIME26 (97.1 with Claim-Level Reliability Assessment test-time scaling), 80.2 Pass@1 on LiveCodeBench v6, 96.1% acceptance rate on unseen LeetCode contests, and 93.4 on IFEval (preserving instruction controllability).
• Methodology: Built on the Spectrum-to-Signal post-training paradigm with curriculum-based supervised fine-tuning, multi-domain reinforcement learning (MGPO), Long2Short efficiency optimization, offline self-distillation, and Instruct RL.
• Theoretical contribution: Introduces the Parametric Compression-Coverage Hypothesis — verifiable reasoning is "parameter-dense" (compressible into compact reasoning cores), while open-domain knowledge requires "parameter-expansive" coverage.
• Implications: Proposes the Reasoning-Knowledge Decoupling Paradigm, suggesting SLMs are not merely deployment-efficient fallbacks but a complementary path to frontier performance in specific capability regimes.

---

Introduction and Theoretical Foundation

The paper addresses the prevailing assumption that frontier reasoning ability requires tens or hundreds of billions of parameters. While reinforcement learning has dramatically improved reasoning in large models, small language models (SLMs) ≤3B are generally considered bottlenecked on difficult math and code tasks.

Building on the authors' prior work VibeThinker-1.5B, which demonstrated that extremely small models can produce stable reasoning chains, this work asks: What is the true capability boundary of a strictly 3B model? Can it achieve frontier-level performance comparable to top-tier LLMs?

The theoretical foundation rests on the Spectrum-to-Signal Principle (SSP):

• Spectrum: The SFT stage constructs a broad solution space (diverse reasoning paths).
• Signal: RL amplifies high-value reasoning signals within that space.

The core hypothesis — the Parametric Compression-Coverage Hypothesis — posits that capabilities differ structurally in parameter demands:

• Parameter-dense capabilities (verifiable reasoning): Search, constraint satisfaction, error correction, multi-step composition within structured solution spaces — highly compressible into compact reasoning cores.
• Parameter-expansive capabilities (knowledge-intensive, general-purpose): Broad coverage over facts, concepts, long-tail scenarios — resemble a coverage problem requiring massive parameter capacity.

This explains why VibeThinker-3B matches top-tier systems on math/coding but lags on knowledge-heavy benchmarks like GPQA-Diamond.

---

Methodology

Overall Pipeline (Fig. 3)

Built on Qwen2.5-Coder-3B base, the post-training pipeline unfolds in four stages:

1. Supervised Fine-Tuning (SFT)

Data Construction:

• Data Synthesis & Query Expansion: Seed queries with reliable supervision (math with verified answers, code with unit tests) are rewritten across dimensions (concept composition, constraints, evaluation objectives). Pseudo-labels generated via majority voting from strong teachers.
• Multi-path Reasoning Distillation: Multiple candidate reasoning traces sampled per query (not just one standard solution) to construct a diverse "solution spectrum."
• Multi-level Quality Control:
  1. N-gram-based filtering (removes repetitive/templated/decontaminated samples)
  2. LLM-based query quality filtering
  3. Trace correctness filtering (answer verification, code sandbox execution, LLM majority voting)

Training Process:

• Two-stage curriculum SFT:
  • Stage 1: Broad capability coverage over full dataset, 5 epochs, LR $5 \times 10^{-5}$ → cosine decay to $8 \times 10^{-8}$, sequence packing.
  • Stage 2: Hard-reasoning subset (traces >5K tokens, error rate ≥0.75 via 8 rollouts of VibeThinker-1.5B reference), 2 additional epochs.
• Diversity-Exploring Distillation: Intermediate checkpoints evaluated on Pass@K per domain; domain "specialist" checkpoints merged at parameter level to preserve output diversity.

2. Reinforcement Learning (RL)

Algorithm: MaxEnt-Guided Policy Optimization (MGPO) — core equations preserved:

Empirical group accuracy:

$$p(q) = \frac{1}{G} \sum_{i=1}^{G} \mathbb{I}(r_i = 1) \tag{1}$$

Prompt weighting (maximum-entropy principle, $p_0 = 0.5$):

$$w(q) = \exp\left(-\gamma D_{\mathrm{ME}}(p(q) \| p_0)\right), \quad \text{where } p_0 = 0.5, \gamma > 0 \tag{2}$$

MGPO clipped objective (GRPO-style):

$$\mathcal{J}_{\mathrm{MGPO}}(\theta) = \mathbb{E}_{q, \{y_i\}} \left[ \frac{1}{G} \sum_{i=1}^{G} \frac{1}{|y_i|} \sum_{t=1}^{|y_i|} \min \bigl(\rho_{i,t}(\theta) w(q) A_i, \operatorname{clip}(\rho_{i,t}(\theta), 1-\varepsilon, 1+\varepsilon) w(q) A_i \bigr) \right] \tag{3}$$

Key design decisions:

• Single 64K long-context window (not progressive expansion — high-truncation warm-up was found to disrupt existing high-quality long-horizon reasoning in the stronger 3B initialization).
• Sequential multi-domain RL: Math RL → Code RL → STEM RL, each with domain-specific verifiers.
• Long2Short Math RL: Two-stage efficiency optimization. For correct trajectories $\mathcal{C} = \{i \mid r_i = 1\}$ with brevity score $s_i = 1/L_i$:

$$r_i' = r_i + \lambda \cdot \frac{s_i - \bar{s}}{\max_{j \in \mathcal{C}} |s_j - \bar{s}|}, \qquad i \in \mathcal{C}$$

with $\lambda = 0.2$. The zero-sum property holds:

$$\sum_{i \in \mathcal{C}} (r_i' - r_i) = 0$$

— reshaping relative preference toward concise reasoning without shifting the group-level reward baseline.

3. Offline Self-Distillation

• Rejection sampling with verifiers removes incorrect trajectories.
• Learning-potential filtering scores each verified trace by student-model length-normalized negative log-likelihood:

$$S_{\mathrm{LP}}(q, y) = -\frac{1}{|y|} \sum_{t=1}^{|y|} \log \pi_{\theta_{\mathrm{stu}}}(y_t \mid q, y_{<t}) \tag{4}$$

• Traces ranked within domain-specific length buckets (avoiding length bias); mid-to-high score traces prioritized.

4. Instruct RL

• Mixed instruction dataset (format-sensitive, long-context, general alignment).
• Rule-based validators for explicit constraints; rubric-based reward models for open-ended prompts.
• Reinforces controllability while preserving reasoning.

Test-Time Scaling: Claim-Level Reliability Assessment (CLR)

Two-stage procedure: (1) Generate $K=32$ trajectories, extract $M=5$ decision-relevant claims + final answer per trajectory; (2) Model self-verifies claims, yielding binary verdicts $v_{k,m} \in \{0,1\}$. Nonlinear trajectory-level reliability score:

$$r_k = \left(\frac{1}{M} \sum_{m=1}^{M} v_{k,m}\right)^{M} \tag{5}$$

Answers clustered by equivalence, reliability-weighted aggregation selects final answer:

$$\operatorname{Score}(G) = \sum_{\{k \mid y_k \in G\}} r_k \tag{6}$$

---

Empirical Validation / Results

Core Benchmark Performance (Table 1)

VibeThinker-3B vs. small/mid-sized reasoning models (3B–14B):

VibeThinker-3B also surpasses larger models (20B–560B) on most verifiable benchmarks.

Top-Tier Comparison (Table 2)

With CLR, VibeThinker-3B enters the top cluster on competition mathematics, matching or exceeding flagship models.

Out-of-Distribution Generalization (Table 3): LeetCode Contests (Apr 25–May 31, 2026)

VibeThinker-3B achieves 96.1% (123/128) first-attempt Python acceptance — exceeding GPT-5.2 (95.3%), Doubao Seed 2.0 Pro (94.5%), Qwen3-Max (91.4%), Kimi K2.5 (90.6%), and GLM-5 (76.6%), trailing only GPT-5.3-Codex (100%), Gemini 3.1 Pro (99.2%), and Gemini 3 Flash (96.9%).

Parameter Efficiency (Fig. 2)

On IMO-AnswerBench, VibeThinker-3B (76.4, 3B params; 80.6 with CLR) reaches the performance band of DeepSeek V3.2 (78.3, 671B), GLM-5 (82.5, 744B), and Kimi K2.5 (81.8, 1T).

---

Theoretical and Practical Implications

1. Parametric Compression-Coverage Hypothesis: Verifiable reasoning is fundamentally compressible — its core (search, constraint satisfaction, error correction) can be encoded in a compact reasoning core, whereas knowledge coverage requires broad parameter capacity. This explains the asymmetric performance profile of VibeThinker-3B (frontier-level math/code, gap on GPQA-Diamond).

2. Reasoning-Knowledge Decoupling Paradigm: Large-scale models remain necessary for expansive knowledge breadth; small models can encapsulate high-density reasoning depth. This reframes SLM development as a complementary research trajectory, not a passive compromise.

3. Practical significance:

   • Demonstrates that post-training quality (diverse exploration, reliable verification, curriculum design) can substitute for raw parameter scale in verifiable domains.
   • The single long-context window finding challenges progressive context-expansion strategies for stronger initializations.
   • Long2Short RL shows efficiency optimization can proceed without accuracy loss via zero-sum reward redistribution.
   • CLR provides a parameter-free test-time scaling method that improves answer-verifiable tasks without processing entire verbose trajectories.

---

Conclusion

VibeThinker-3B demonstrates that a strictly 3B model can reach the performance band of frontier LLMs on highly complex verifiable tasks. The authors propose that verifiable reasoning aligns with parameter-dense core compression, while open-domain knowledge requires parameter-expansive coverage. Consequently, the potential of small models for top-tier performance in specific capability domains has been underestimated.

Future directions implied by this work:

• Extending the compact reasoning core to additional verifiable domains (e.g., formal verification, theorem proving).
• Investigating the boundary of the compression-coverage trade-off (at what parameter scale does knowledge coverage become feasible?).
• Further improving knowledge-intensive performance in small models via retrieval or external memory augmentation.
• Scaling the post-training paradigm to other base architectures and parameter sizes.