# Test-Time Compute

Test-time search for multi-step LLM reasoning has long traded between two imperfect signals: post-hoc voting keeps diversity but cannot steer generation toward correct paths, while beam-style verifier guidance intervenes early but over-concentrates and amplifies verifier errors. VGSC multiplies step-level correctness with a future-outcome prediction into an acceptance weight, applies bounded rejection-resampling to candidate steps during decoding, and keeps multiple trajectories for final self-consistency voting. Across five reasoning benchmarks with the Qwen2.5 Math 1.5B Instruct generator and Qwen2.5 Math PRM 7B verifier, VGSC reaches 63.1% average accuracy at verifier-trust exponent 10, above the strongest non-VGSC baseline BS+SC-Future at 60.7%.

Test-time search for multi-step LLM reasoning has long traded between two imperfect signals: post-hoc voting keeps diversity but cannot steer generation toward correct paths, while beam-style verifier guidance intervenes early but over-concentrates and amplifies verifier errors. VGSC multiplies step-level correctness with a future-outcome prediction into an acceptance weight, applies bounded rejection-resampling to candidate steps during decoding, and keeps multiple trajectories for final self-consistency voting. Across five reasoning benchmarks with the Qwen2.5 Math 1.5B Instruct generator and Qwen2.5 Math PRM 7B verifier, VGSC reaches 63.1% average accuracy at verifier-trust exponent 10, above the strongest non-VGSC baseline BS+SC-Future at 60.7%.

This work systematically studies how test-time compute affects predictions of pretrained tabular foundation models along three axes—adaptation, aggregation, and context construction: it introduces DiagScale, a diagonal query-key similarity update training only 0.003-0.03% of parameters that matches full fine-tuning gains across three independently pretrained backbones; on a broader pool of 96 configurations, greedy selection reduces error by 2.4% relative to the default predictor while uniform averaging increases error; attention-guided retrieval improves TabPFN-3 on some large tables, whereas the context expansion methods tested yield no consistent improvement.

This work systematically studies how test-time compute affects predictions of pretrained tabular foundation models along three axes—adaptation, aggregation, and context construction: it introduces DiagScale, a diagonal query-key similarity update training only 0.003-0.03% of parameters that matches full fine-tuning gains across three independently pretrained backbones; on a broader pool of 96 configurations, greedy selection reduces error by 2.4% relative to the default predictor while uniform averaging increases error; attention-guided retrieval improves TabPFN-3 on some large tables, whereas the context expansion methods tested yield no consistent improvement.

The work trains 15 masked-diffusion codec TTS models (19-133M parameters, 3 seeds) on 2,000 hours of speech and sweeps refinement steps T in [1,16] at inference, measuring zero-shot synthesis on 174 held-out speakers via ASR word error rate and speaker verification; refinement closes 86.2% of the intelligibility range but only 46.4% of the identity range (a 1.86x asymmetry), Best-of-K search recovers identity with 64.6-79.0% win rates across four independent encoders, a separable B(d)B(T) model fits significantly better than substitution models (Delta AICc=+69.3), and 62% of the remaining identity deficit lies in the codec rather than the generator.

The work trains 15 masked-diffusion codec TTS models (19-133M parameters, 3 seeds) on 2,000 hours of speech and sweeps refinement steps T in [1,16] at inference, measuring zero-shot synthesis on 174 held-out speakers via ASR word error rate and speaker verification; refinement closes 86.2% of the intelligibility range but only 46.4% of the identity range (a 1.86x asymmetry), Best-of-K search recovers identity with 64.6-79.0% win rates across four independent encoders, a separable B(d)B(T) model fits significantly better than substitution models (Delta AICc=+69.3), and 62% of the remaining identity deficit lies in the codec rather than the generator.