Exploitation mechanisms beyond selection and synthesis

Develop exploitation mechanisms beyond selection and synthesis that can convert high-quality candidate pools into substantially better final responses.

Background

The experiments show that additional inference compute steadily improves the quality of candidate pools, but existing exploitation strategies recover only a fraction of the available headroom. External selection is limited by poor verifier correlation, tree search suffers diversity collapse, and Fusion—despite being the strongest consistently improving method—captures only about 40% of available quality. The authors therefore leave open the development of exploitation methods beyond selecting or synthesizing among candidates.

References

Second, developing exploitation mechanisms beyond selection and synthesis: the high unrealised headroom represents a substantial opportunity.

— Test-Time Scaling in the Wild: Why Exploitation, Not Exploration, Is the Bottleneck  (2608.18931 - Romano et al., 19 Aug 2026) in Section Future work, Section 1 / Conclusion

A prompted, untrained $\leq$4B aggregator is in the regime where GSR's scaling account predicts little refinement ability \citep[\S4;][]{wang2025gsr}; whether a net recombination benefit appears with training, scale, or iteration (\S5) remains an open question.

— Selection, Recombination, or a Fresh Solve? A Candidate-Free Control for Single-Pass Test-Time Aggregation  (2608.18379 - Farmanfarmaian, 18 Aug 2026) in Section 1, Introduction; Section 4, Related work