Resource-rational compute allocation for LRMs

Develop a principled cost–performance framework that adapts compute allocation and halting policies for large reasoning models based on instance difficulty and epistemic uncertainty, thereby addressing the open question of efficient reasoning control.

Background

Inference-time scaling improves accuracy but introduces over-thinking for easy instances and under-thinking under aggressive truncation. The authors argue for adaptive compute allocation policies but note the lack of a principled trade-off framework. Solving this would guide LRMs to reason only as long as warranted by marginal utility.

References

However, generalizing these approaches into a principled cost-performance trade-off remains an open question.

— A Survey of Reinforcement Learning for Large Reasoning Models  (2509.08827 - Zhang et al., 10 Sep 2025) in Section 7.4 Teaching LRMs Efficient Reasoning

The open problem is a controller that decides how much structure to impose, and when to change it, rather than committing to a single transition style for an entire trajectory.

— Complex Problem Solving in Large Language Models: A Statistical Control Survey and Diagnostic Framework  (2609.20973 - Cai et al., 17 Sep 2026) in Section 5.2, Adaptive Transition, Search, and Rollback Control

Several directions remain open. An important one is to jointly allocate computation between candidate generation and verification.

— Test-Time Scaling via Budgeted Multi-Attribute Verification  (2609.34322 - Xue et al., 28 Sep 2026) in Section Conclusion and Future Work

A more principled strategy would select tree sizes at each decoding round using calibrated estimates of acceptance benefit derived from transition scores, together with a cost model reflecting hardware and serving load. We leave developing and evaluating such a policy to future work.

— UBTree: Parallel Tree Drafting via Unigram and Bigram Models for Speculative Decoding  (2609.39972 - Liang et al., 30 Sep 2026) in Section 6, Limitations, Appendix A.6 (page number unavailable)

Finer-grained and trajectory-aware termination remains an open direction. Future work can study policies that use richer information from the generated prefix, support finer-grained stopping decisions, and evaluate a broader range of models and tasks. Such studies can also incorporate deployment overheads and examine how closely online decisions can approach the offline references.

— Constructing Pareto Frontiers of Optimal Reasoning Termination Methods for Overthinking in Reasoning Models  (2610.13570 - Guan et al., 9 Oct 2026) in Appendix E, “Limitations and Further Directions”