Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2511.02130 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866909885712564224 |
|---|---|
| author | Zabounidis, Renos Golatkar, Aditya Kleinman, Michael Achille, Alessandro Xia, Wei Soatto, Stefano |
| author_facet | Zabounidis, Renos Golatkar, Aditya Kleinman, Michael Achille, Alessandro Xia, Wei Soatto, Stefano |
| contents | We propose Re-FORC, an adaptive reward prediction method that, given a context, enables prediction of the expected future rewards as a function of the number of future thinking tokens. Re-FORC trains a lightweight adapter on reasoning models, demonstrating improved prediction with longer reasoning and larger models. Re-FORC enables: 1) early stopping of unpromising reasoning chains, reducing compute by 26% while maintaining accuracy, 2) optimized model and thinking length selection that achieves 4% higher accuracy at equal compute and 55% less compute at equal accuracy compared to the largest model, 3) adaptive test-time scaling, which increases accuracy by 11% in high compute regime, and 7% in low compute regime. Re-FORC allows dynamic reasoning with length control via cost-per-token thresholds while estimating computation time upfront. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2511_02130 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Re-FORC: Adaptive Reward Prediction for Efficient Chain-of-Thought Reasoning Zabounidis, Renos Golatkar, Aditya Kleinman, Michael Achille, Alessandro Xia, Wei Soatto, Stefano Artificial Intelligence Machine Learning We propose Re-FORC, an adaptive reward prediction method that, given a context, enables prediction of the expected future rewards as a function of the number of future thinking tokens. Re-FORC trains a lightweight adapter on reasoning models, demonstrating improved prediction with longer reasoning and larger models. Re-FORC enables: 1) early stopping of unpromising reasoning chains, reducing compute by 26% while maintaining accuracy, 2) optimized model and thinking length selection that achieves 4% higher accuracy at equal compute and 55% less compute at equal accuracy compared to the largest model, 3) adaptive test-time scaling, which increases accuracy by 11% in high compute regime, and 7% in low compute regime. Re-FORC allows dynamic reasoning with length control via cost-per-token thresholds while estimating computation time upfront. |
| title | Re-FORC: Adaptive Reward Prediction for Efficient Chain-of-Thought Reasoning |
| topic | Artificial Intelligence Machine Learning |
| url | https://arxiv.org/abs/2511.02130 |