@marfinxx: This Google DeepMind paper is f*cking brilliant A new research paper proves that turning verifiers into generative next…

This Google DeepMind paper is f*cking brilliant A new research paper proves that turning verifiers into generative next-token predictors unlocks superhuman reasoning accuracy The result: chain-of-thought verification rationales, inference-time compute scaling, and seamless generalization from grade-school math to competition problems candidate solutions → CoT rationale generation → token probability scoring → majority-voted ensemble → optimal solution select Google DeepMind unified verification and generation under standard next-token prediction: → Verification as Next-Token Prediction Replaces arbitrary scalar reward heads with the direct probability of generating a Yes or No token. Verification becomes native autoregressive text generation rather than artificial classification → Chain-of-Thought Verification (GenRM-CoT) The verifier generates an explicit step-by-step critique before predicting correctness, catching hidden calculation bugs and hallucinations missed by direct scoring → Inference-Time Compute Scaling Samples multiple independent verification rationales and averages the predicted Yes probabilities across majority votes, filtering out single-path errors and outperforming prompt-based LLM judges → Easy-to-Hard Generalization A 9B generative verifier trained only on grade-school math successfully generalized to high-school competition problems on MATH500, boosting solve rates from 28.0% to 44.6% → Hypothesis Verification & Empirical Results Google DeepMind hypothesized that generative critique with test-time compute scaling dramatically outperforms discriminative reward modeling. Practical validation on complex multi-step reasoning pipelines confirmed the findings: task accuracy surged from 73.0% to 93.4% on GSM8K while requiring 2.5x fewer candidate samples Treating verification as generative reasoning turns reward modeling into an autoregressive self-correction engine Read the complete breakdown in the article below ↓