# Test-Time Scaling

Parallel test-time scaling for large language models normally generates every reasoning path to completion before internal confidence can score it, so compute on low-quality paths is already spent. PreG instead generates short prefixes in parallel, ranks them by internal confidence, and completes only the most promising candidates before aggregating as before. On Qwen3-8B and Llama-Nemotron-4B across math competition sets and GPQA-Diamond, it reaches the accuracy of the original method at a 1000k-token budget with fewer tokens, saving tokens in 54 of 60 configurations and up to 80%.

Parallel test-time scaling for large language models normally generates every reasoning path to completion before internal confidence can score it, so compute on low-quality paths is already spent. PreG instead generates short prefixes in parallel, ranks them by internal confidence, and completes only the most promising candidates before aggregating as before. On Qwen3-8B and Llama-Nemotron-4B across math competition sets and GPQA-Diamond, it reaches the accuracy of the original method at a 1000k-token budget with fewer tokens, saving tokens in 54 of 60 configurations and up to 80%.

Masked diffusion language models can remask generated text, but deciding where to revise by confidence or self-judgment often leaves the error in place. DiVR instead learns an outcome verifier on the generator's own decodes and, at each block boundary, judges whether the partial solution can still reach a correct answer, resampling the whole block when the score is low. On the supervised fine-tuned LLaDA-8B-Base generator, the five-benchmark math and code average rises from 35.3% to 43.4% at the cost of greedy decoding.

Masked diffusion language models can remask generated text, but deciding where to revise by confidence or self-judgment often leaves the error in place. DiVR instead learns an outcome verifier on the generator's own decodes and, at each block boundary, judges whether the partial solution can still reach a correct answer, resampling the whole block when the score is low. On the supervised fine-tuned LLaDA-8B-Base generator, the five-benchmark math and code average rises from 35.3% to 43.4% at the cost of greedy decoding.

Eco-RNA turns inference depth into a selectable resource: one 80M-parameter checkpoint applies a shared core repeatedly, lifting aggregate absolute Spearman correlation on the 70-assay RNAGym fitness benchmark from 0.242 to 0.286 and passing Evo 2 40B at four iterations.

Eco-RNA turns inference depth into a selectable resource: one 80M-parameter checkpoint applies a shared core repeatedly, lifting aggregate absolute Spearman correlation on the 70-assay RNAGym fitness benchmark from 0.242 to 0.286 and passing Evo 2 40B at four iterations.

This work studies hypernetworks conditioned only on an LLM's input query to estimate LoRA, and introduces distributional hypernetworks that output a distribution over LoRAs via an end-to-end loss with a differentiable Monte Carlo approximation; results show that even the mean of the learned distribution can outperform deterministic hypernetworks, that sampling weight updates yields a form of test-time scaling distinct from token sampling with performance improving as more weight samples are considered and remaining stronger than corresponding token-sampling adaptation baselines, and that generated updates can transfer across queries.

This work studies hypernetworks conditioned only on an LLM's input query to estimate LoRA, and introduces distributional hypernetworks that output a distribution over LoRAs via an end-to-end loss with a differentiable Monte Carlo approximation; results show that even the mean of the learned distribution can outperform deterministic hypernetworks, that sampling weight updates yields a form of test-time scaling distinct from token sampling with performance improving as more weight samples are considered and remaining stronger than corresponding token-sampling adaptation baselines, and that generated updates can transfer across queries.

The work proposes SCOUT, an offline multi-agent reinforcement learning framework that separately trains a flow-matching behavioral prior and a decomposed value function, then at test time transports behavioral samples toward high-value regions via Stein variational gradient descent, using the number of transport steps in place of a fixed regularization coefficient; under the individual-global-max (IGM) principle it proves a single-term KL bound on the joint soft-value gap that vanishes as transport converges, with an irreducible additive residual proportional to the IGM violation; empirically it achieves the best average performance across discrete and continuous offline MARL benchmarks and yields improvements in all offline-to-online configurations.

The work proposes SCOUT, an offline multi-agent reinforcement learning framework that separately trains a flow-matching behavioral prior and a decomposed value function, then at test time transports behavioral samples toward high-value regions via Stein variational gradient descent, using the number of transport steps in place of a fixed regularization coefficient; under the individual-global-max (IGM) principle it proves a single-term KL bound on the joint soft-value gap that vanishes as transport converges, with an irreducible additive residual proportional to the IGM violation; empirically it achieves the best average performance across discrete and continuous offline MARL benchmarks and yields improvements in all offline-to-online configurations.

The work introduces architectural sampling, a training-free method that generates candidates through distinct forward computations by reusing selected blocks of decoder layers in a frozen model; across five Qwen checkpoints and twelve multimodal benchmarks it improves pass@9 over standard-path temperature sampling by 6.58 percentage points on average at the same nine-candidate budget, with the strongest gains from reusing early layers, candidate-coverage improvement persisting under greedy decoding, lower lexical overlap among candidates, and higher accuracy when the candidates serve as rollouts for label-free test-time reinforcement learning.

The work introduces architectural sampling, a training-free method that generates candidates through distinct forward computations by reusing selected blocks of decoder layers in a frozen model; across five Qwen checkpoints and twelve multimodal benchmarks it improves pass@9 over standard-path temperature sampling by 6.58 percentage points on average at the same nine-candidate budget, with the strongest gains from reusing early layers, candidate-coverage improvement persisting under greedy decoding, lower lexical overlap among candidates, and higher accuracy when the candidates serve as rollouts for label-free test-time reinforcement learning.

The work introduces Mid-Harness, which samples and verifies candidate actions at the model-harness boundary before forwarding one for execution while keeping the generator and harness unchanged, and studies action-level test-time compute scaling: on TerminalBench-Lite a GPT-5.6 Sol verifier raises TMAX-9B Pass@1 from 50.00% to 68.03% with 8 sampled actions, wider sampling yields little benefit under weak verification, pairwise verification performs best among the evaluated self-verification mechanisms, distilling the stronger verifier's responses into TMAX-9B further improves Pass@1, and combining action scaling with trajectory scaling (Best-of-N and Sequential Refine) reaches higher success at lower estimated token cost.

The work introduces Mid-Harness, which samples and verifies candidate actions at the model-harness boundary before forwarding one for execution while keeping the generator and harness unchanged, and studies action-level test-time compute scaling: on TerminalBench-Lite a GPT-5.6 Sol verifier raises TMAX-9B Pass@1 from 50.00% to 68.03% with 8 sampled actions, wider sampling yields little benefit under weak verification, pairwise verification performs best among the evaluated self-verification mechanisms, distilling the stronger verifier's responses into TMAX-9B further improves Pass@1, and combining action scaling with trajectory scaling (Best-of-N and Sequential Refine) reaches higher success at lower estimated token cost.