Research note 02 / PDF
LittlePly as a chess agent
Early draft v0.2, 10 November 2026, not peer reviewed. CM-14 and the harness that turns predictions into play: a five-page account of the combined agent, dated benchmark and Lichess evidence, and measured lessons from CM-15 and CM-16. It separates model learning from system performance and treats the auxiliary-objective explanation as a working hypothesis.
Read the research draft (PDF, 5 pages)
Preserved earlier draft: v0.1, 6 November 2026 (PDF, 3 pages). Its proposed evaluation predates the completed 50-game full-system result. See the dated evidence and protocol differences.
1. A game becomes a sequence
CM-14 reads the complete legal move history from the standard starting position. Each UCI move—including its promotion piece—is one token. A beginning-of-game token separates games. A 512-token context is a limit, not permission to silently forget older moves.
2. Training learns policy and outcome estimates
The transformer learns from both players' eligible moves. Teacher forcing supplies earlier real moves, never future outcomes. Packed games use separate causal attention blocks. A separate three-class normalization predicts White win/draw/loss at selected positions; those estimates are model predictions, not calibrated promises.
3. Rules turn predictions into a legal move
A chess rules library enumerates legal moves. The model ranks them. This keeps actual play legal even when its raw top prediction is illegal, so raw prediction legality and legal played moves must be reported separately.
4. The harness decides how to use the model
The online harness adds uncertainty-guided lookahead, persistent search and inference caches, bounded Stockfish assistance, and clock-aware execution. It checks history and legality before acting, then records the decision. Browser play keeps the simpler model-and-legal-filtering path. These are different agents around the same learned weights, so their results answer different questions.
5. Evaluation tests the model and the agent
Loss, raw legality, Sunfish results and tournament standings answer different questions. Keep opponent, opening set, helper, clocks and failure accounting fixed before attributing gains. A small score or one tournament is not a FIDE rating.
Exact current architecture · Experiment evidence · Run the recipe
Explore earlier ideas
The historical learning notebook preserves three-token and state-input explanations. Their token counts and training objectives describe those models, not CM-14.
Related research
A reading map, checked against primary sources on November 4, 2026. These papers ask different questions: human alignment, strength, planning and interpretability are not interchangeable. Citations do not validate LittlePly’s scores or imply wholesale implementation. “Possible future experiment” means unimplemented reading, not an announced run.
Learning from human games
Human-aligned Chess with a Bit of Search — Zhang et al., 2024.
Allie models human game sequences, including thinking time and resignations, and uses time-adaptive Monte Carlo tree search for skill alignment.
Connection to LittlePly: Related work. LittlePly models moves, but does not implement Allie’s human-time prediction or adaptive MCTS.
Aligning Superhuman AI with Human Behavior: Chess as a Model System — McIlroy-Young et al., 2020.
Maia trains on human games to predict decisions at particular skill levels. Predicting human choices is a different objective from maximizing playing strength.
Connection to LittlePly: Related work. LittlePly’s human-game supervision does not establish Maia-style human alignment.
Maia-2: A Unified Model for Human-AI Alignment in Chess — Tang et al., 2024.
A unified model integrates player skill through skill-aware attention, rather than training independent models for separate rating bands.
Connection to LittlePly: Possible future experiment: explicit skill conditioning. It is not implemented in CM-14.
Transcendence: Generative Models Can Outperform The Experts That Train Them — Zhang et al., 2024.
Studies when transcript-trained autoregressive chess models can outperform their data-generating players, including the role of low-temperature sampling.
Connection to LittlePly: Related work for imitation and decoding. It does not show that LittlePly exceeds its training players.
Learning and planning
Amortized Planning with Large-Scale Transformers: A Case Study on Chess — Ruoss et al., 2024.
ChessBench supplies Stockfish annotations for comparing behavioral cloning, state-value and action-value prediction. The work studies distilling search into a policy without explicit inference-time search.
Connection to LittlePly: Possible future experiment: engine action-value supervision. CM-14’s game-result WDL diagnostic is not that target.
Policy improvement by planning with Gumbel — Danihelka et al., 2022. Author-hosted paper.
Uses sampling without replacement and sequential halving to improve planning with few simulations; its policy-improvement guarantees rely on the stated value assumptions.
Connection to LittlePly: Possible future experiment, not deployed. LittlePly’s bounded Stockfish mate checks are not Gumbel policy improvement.
A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play — Silver et al., 2018. Author-hosted overview.
Combines self-play reinforcement learning with a neural network guiding Monte Carlo tree search across three games.
Connection to LittlePly: Foundational related work. LittlePly’s supervised training and external Stockfish assistance do not implement AlphaZero’s self-play loop.
Implicit Search via Discrete Diffusion: A Study on Chess — Ye et al., 2025.
DiffuSearch models future trajectories through discrete diffusion and compares them with one-step and explicit-search policies in its own chess experiments.
Connection to LittlePly: Possible future architecture experiment. CM-14 is an autoregressive next-move model, not a diffusion planner.
What a model represents and predicts
Emergent World Models and Latent Variable Estimation in Chess-Playing Language Models — Karvonen, 2024.
Uses probes and activation interventions to investigate board-state and player-skill representations in a chess language model.
Connection to LittlePly: Related diagnostic work. LittlePly’s legal-move performance alone does not prove that it has the same internal representations.
Evidence of Learned Look-Ahead in a Chess-Playing Neural Network — Jenner et al., 2024.
Investigates future-move representations and causal evidence of look-ahead inside Leela Chess Zero’s transformer network on selected positions.
Connection to LittlePly: Related diagnostic work, not evidence of learned search in LittlePly. Such a claim would require separate interventions and controls.
Win-Draw-Loss evaluation — Leela Chess Zero / crem, 2020.
Explains why predicting win, draw and loss separately contains more information than a single expected-score number, and describes UCI WDL output.
Connection to LittlePly: Implemented concept, different implementation: CM-14 normalizes three fixed-White outcome logits separately from move logits. It is not LCZero’s network or search, and its probabilities are not established as calibrated.
Use each paper’s own evaluation protocol. Human-move accuracy, puzzle scores, online ratings and our Sunfish development matches cannot be placed on one common strength scale.