Tiny Model, Big Logic: Large-Model Reasoning Ability in VibeThinker-1.5B (arxiv.org)

VibeThinker-1.5B is a dense 1.5-billion-parameter model trained under a new Spectrum-to-Signal Principle (SSP) that combines a Two‑Stage Diversity‑Exploring Distillation (SFT) to generate a wide spectrum of candidate solutions and a MaxEnt‑Guided Policy Optimization (RL) phase to amplify the correct signal. Trained for a reported total cost of $7,800, the model reportedly matches or exceeds much larger systems: it outperforms closed-source Magistral Medium and Claude Opus 4, rivals open-source GPT OSS‑20B Medium, and even beats the 400× larger DeepSeek R1 on several math benchmarks (AIME24 80.3 vs. 79.8; AIME25 74.4 vs. 70.0; HMMT25 50.4 vs. 41.7). On LiveCodeBench V6 it scores 51.1 versus Magistral Medium’s 50.3, a dramatic uplift from its base model (e.g., AIME gains from 6.7→80.3). The paper’s key technical claim is that diversity‑driven optimization — deliberately producing many plausible solution trajectories, then using an entropy‑aware RL step to select and reinforce the true signal — can elicit strong reasoning in small models without brute‑force scaling. If reproducible, this lowers training and inference cost barriers and could democratize research-grade reasoning models. However, the result hinges on replication and thorough benchmarking (including dataset and evaluation transparency), so the community will watch closely for code, checkpoints, and independent validations.