ThinkBooster 用 GPT-6 Sol 等模型做推理搜索

marfin @marfinxx原推发布于 2026.09.25

在 X 上打开原推

this is pure f*cking magic Claude Opus 5.5, GPT-6 Sol, and Grok 4.7 broke the inference scaling limit Single-pass autoregression has hit a physical wall. When each reasoning step carries a 90% success probability, a 20-step engineering chain resolves correctly only 12.1% of the time: P(success) = 0.90^20. Adding pre-training parameters cannot overcome exponential error compounding or unguided sampling biases Researchers from MBZUAI introduced ThinkBooster to unify test-time compute scaling across frontier models. Orchestrating Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and Grok 4.7 across dynamic tree search lifts HumanEval+ pass@1 by +9.5% to 88.8% and boosts KernelBench CUDA correctness to 30.0% while slashing inference compute by 68.2% Complexity triage → Adaptive MCTS rollout (GPT-6 Sol) → Step entropy scoring (GPT-6 Luna) → Cutoff pruning (<0.45) → Formal synthesis (Claude Opus 5.5 & Grok 4.7) Four operational principles separating test-time search engines from naive prompt retries: 1. Search Topologies vs Naive Resampling Generating 64 independent full-sequence completions wastes compute on identical early mistakes. ThinkBooster replaces best-of-N with Monte Carlo Tree Search, using Upper Confidence Bounds (UCT = 0.94) to concentrate token budgets along verified high-reward trajectories 2. Domain-Agnostic Entropy vs PRM Collapse Trained Process Reward Models overfit to synthetic math problems and collapse on complex software. Domain-agnostic step entropy H(t) and predictive distribution gaps evaluate intermediate code correctness with zero training overhead and zero parameter cost 3. Dynamic Budget Forcing & Early Tripwire Pruning Prefixing deliberative control tokens like "\nWait, let me double check..." forces models into extended reasoning mode on difficult CUDA kernels. Trajectories dropping below a 0.45 quality threshold trigger instant branch termination, eliminating runaway hallucination loops 4. Heterogeneous Frontier Model Orchestration Frontier models operate as specialized nodes in a single inference tree: GPT-6 Sol drives parallel hypothesis exploration, GPT-6 Luna executes ultra-low-cost step evaluation, Grok 4.7 falsifies edge cases, and Claude Opus 5.5 locks production code into persistent KV cache Treating inference as a linear token stream caps autonomous capability. True reasoning requires transforming autoregressive generation into a verifiable graph search that prunes dead ends before committing a single invalid token Read the complete breakdown in the article below ↓

这条是引用着另一条写的：x.com/marfinxx/status/2084595353060258071