Papers

Research notes used for DLManaka's move from distillation to self-play. Implementation status belongs in the self-play roadmap; this page only records what each paper contributes.

Adoption candidates

Regret-Guided Search Control for Efficient Learning in AlphaZero

Yun-Jui Tsai et al., ICLR 2026 — https://arxiv.org/abs/2602.20809

Starts self-play from high-regret positions, where the network's earlier evaluation disagreed strongly with the final outcome. Candidate positions come from trajectories and the MCTS tree and are prioritized for reuse.

Why it matters here: DLManaka already has archived starting positions and a demonstrated need for more targeted data than “more games from the same distribution”. This is the leading data-selection candidate after the cheaper value-target experiments.

Uncertainty-Guided Exploration for Efficient AlphaZero Training

Scott Cheng et al., NeurIPS 2025 — official NeurIPS proceedings.

Finds uncertain positions using label-change rate, branches several MCTS alternatives there, and averages outcomes to reduce value-label variance.

Why it matters here: it targets the same failure mode suspected in DLManaka's noisy z labels. It is more expensive than changing the z/q blend, so it is a later experiment.

Prior work required to understand the current design

Targeted Search Control in AlphaZero for Effective Policy Improvement (Go-Exploit)

Alexandre Trudeau — https://arxiv.org/abs/2302.12359

Earlier search-control work: select suspicious positions from an archive and restart self-play there. RGSC mainly improves how those positions are ranked.

Policy Improvement by Planning with Gumbel

Ivo Danihelka et al., ICLR 2022 — OpenReview.

The basis of DLManaka's current search. Gumbel-Top-k and Sequential Halving produce the improved policy target used at small simulation budgets. The implementation is documented in search design.

On hold

Search-contempt: a hybrid MCTS algorithm for training AlphaZero-like engines

arXiv: https://arxiv.org/abs/2504.07757

Interesting because its experiments are in chess and target cheaper training-time search. It remains on hold because the work is a preprint and DLManaka already has a functioning Gumbel baseline.

Accelerating Monte-Carlo Tree Search with Optimized Posterior Policies (RMCTS)

Keith Frankston, Benjamin Howard — https://arxiv.org/abs/2601.01301

Replaces standard MCTS selection with a regularized posterior-policy computation and breadth-first batching. This competes with Gumbel at the search-algorithm level, so it is a replacement candidate, not an additive feature.

Reference only

ReSCALE — Gumbel and Sequential Halving for LLM reasoning

Leonid Ugadiarov et al. — https://arxiv.org/abs/2603.21162

Not a game-playing paper, but useful supporting evidence that Sequential Halving rather than incidental AlphaZero machinery is a major contributor to the Gumbel-family behavior.

Local PDFs

Some paper PDFs exist under docs/papers/, but the links above are the canonical source. Redistribution terms differ by paper, so a missing PDF should be fetched from its publisher/arXiv page rather than treated as a repository requirement.