- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The authors propose xRRM, a recurrent reasoning model trained using parallel stochastic trajectories, learned latent perturbations, and computing the gradients using only the candidate closest to a valid solution. Evaluations on Sudoku-Extreme, Pencil Puzzle Bench, and Game-of-24 show improved accuracy and test-time scaling.
Critical strengths
-
At approximately matched training FLOPs, xRRM outperforms the second-best method (URM) by 10 points in single-candidate accuracy on Sudoku-extreme. This substantial improvement is the main reason for my positive assessment.
-
Five-run evaluations strengthen the empirical evidence, supporting the practical value of this simple training strategy.
Critical weaknesses
1 The experiments do not isolate the contribution of learned noise from that of winner-takes-all training, making it difficult to assess the individual contribution of each component.
Questions for authors
- How does xRRM perform with a fixed noise scale, keeping the number of training candidates and the winner-takes-all objective unchanged?
2 comments
I thank the authors for their response. I have decided to maintain my positive rating.
Thanks for the clear suggestion to assess the individual contribution of each component.
W1, Q1: Isolating effects
We ran the suggested control (fixed noise scale) proposed and added it to Table 2 (Sudoku-Extreme), Table 3 (Pencil Puzzle Bench) and Table 4 (Game-of-24):
Following the requests of reviewer qsbA, GY5o, and, SmEt we also added further ablations to the same tables:
Mean over candidates (Mean K): The task loss is averaged over all candidates instead of taking the minimum.
Backprop through a random candidate (Random K)
On Sudoku-Extreme with a single evaluation candidate, xRRM reaches 90.2 ± 1.8%, compared with 86.5 ± 6.3% for fixed noise, 69.1 ± 4.5% for the mean loss, and, 70.8 ± 2.8% for the random-candidate control.
The ablations show that in overall selecting the candidate closest to a solution seems to contribute most to xRRM's performance. An exception to this is Pencil Puzzle Bench, were the learned noise scale seems to be equally important.