Curriculum Mining

Source

• Raw Markdown: paper_curriculum-mining-2026.md
• Official blog/demo: Curriculum Mining
• Official code and checkpoints: the-puzzler/selfplay
• Original X status: https://x.com/i/status/2081445835636654096
• Canonical X status: https://x.com/MozarellaPesto/status/2081445835636654096
• Linked inspiration: OmniReset
• Local X conversation snapshot: papers/curriculum-mining-2026/x_thread_mozarellapesto_2081445835636654096.json
• Local normalized thread: papers/curriculum-mining-2026/x_thread_mozarellapesto_2081445835636654096.md
• Local blog snapshots: papers/curriculum-mining-2026/source_blog.html, papers/curriculum-mining-2026/source_blog_article.html, and papers/curriculum-mining-2026/source_blog_article.md
• Local pinned code snapshot: papers/curriculum-mining-2026/source_code_readme.md at commit 853743229b8ef842678327b26ad5a5963049306d

Status And Credibility

This is a July 26, 2026 author blog, code/checkpoint release, and X thread by Matteo Peluso, not a peer-reviewed paper. It is current and unusually inspectable for a weekend project: the release includes final checkpoints, a compute-matched Othello checkpoint, exact architecture-port checks, result CSV/SVG artifacts, and verification scripts. No independent replication was verified, the code/checkpoints were not executed during this ingest, and only one training run per game is reported. No repository license file was visible at the pinned revision, so code reuse terms remain unspecified.

Treat the source as an important mechanism-generating project and a controlled claim against published reference networks, not as current board-game state of the art.

Core Claim

The source asks whether search-free self-play PPO can train strong raw policy networks for Othello, Hex, and Go from sparse terminal rewards. Its central mechanism is to reset each game into a random legal mid-game position, letting the current policy learn first from short, near-terminal problems and then propagate value backward toward earlier states.

The author calls this curriculum mining:

broad candidate reset distribution
  -> current learner succeeds on the locally learnable subset
  -> new behavior makes new reset states learnable
  -> the competence frontier moves toward the opening

A rolling pool of frozen opponent checkpoints addresses a separate instability: half of the games use one of 12 prior policies, and only learner moves contribute to the loss.

Exact Social Claim And Narrowing

The X root says the policies beat Pgx’s published Gumbel AlphaZero policy networks with no MCTS, teacher, or reward shaping, using self-play PPO, sparse terminal rewards, and comparable or fewer environment interactions. A same-author follow-up makes the necessary boundary explicit: the comparison is against the raw published policy networks, not against MCTS-assisted AlphaZero at inference.

Two later same-author replies should remain separate from the reported experiment:

• GRPO is said to have looked workable during informal testing, but no controlled GRPO result appears in the release.
• POET, quality-diversity methods, and dominated novelty search are proposed as ways to manage a population of start states more efficiently than uniform random resets; this is a future direction, not a result.

Method

The release keeps the architecture aligned with Pgx’s published AZNet v0 and changes the objective:

Component	Pgx published baseline	Curriculum Mining
Policy architecture	Gumbel AlphaZero AZNet v0	Flax port of the same architecture, reported as 1,793,995 trainable parameters
Training target	MCTS search-improved policy/value targets	PPO from sparse terminal $\pm 1$ returns
Initial states	Standard game starts inside the baseline recipe	Random legal positions generated by up to 40–50 random plies
Opponent process	Gumbel AlphaZero self-play/search	Current learner plus a 12-checkpoint frozen opponent ring
Evaluation	Published reference checkpoints	Greedy raw policy versus greedy raw policy, no search on either side

The release says the port reproduces baseline outputs with maximum difference zero and 100% greedy-move agreement. That check is present in code but was not rerun during this ingest.

Reported Results

The pinned README reports 10,240 head-to-head games per comparison, using five evaluation seeds, 1,024 random openings, and both seat assignments:

Game	Score versus published raw baseline	W / D / L
Othello 8×8	$0.671\pm 0.005$	66.0 / 2.3 / 31.8
Othello, compute-matched checkpoint	$0.579\pm 0.005$	56.8 / 2.1 / 41.0
Hex 11×11	$0.835\pm 0.004$	83.5 / – / 16.5
Go 9×9	$0.817\pm 0.004$	81.7 / – / 18.3

Training-run variance is not characterized: the release reports one training run (seed 0) per game. Evaluation seeds and opening clusters do not replace independent training seeds.

Compute Accounting

The source counts each AlphaZero search simulation as an environment interaction. Under that symmetric simulator-call ledger:

• Curriculum Mining final policy: 1,048,576,000 interactions.
• Pgx Othello/Hex published checkpoint: 865,075,200 interactions.
• Pgx Go 9×9 published checkpoint: 1,730,150,400 interactions.
• Curriculum Mining Othello iteration-3000 checkpoint: 393M interactions and score 0.579.

Thus the final PPO run uses about $1.21\times$ the Othello/Hex interaction count but $0.61\times$ the Go interaction count. The root’s comparable or fewer wording is game- and checkpoint-dependent.

This is not a matched-total-compute comparison. It does not equate neural-network forward/backward FLOPs, wall-clock time, accelerator utilization, memory traffic, or hyperparameter-search cost. The Pgx checkpoints are published reference networks from a NeurIPS 2023 systems paper, not current state-of-the-art game-playing agents.

Why It Matters For Dynamic Curriculum JEPA

Dynamic Curriculum Learning For JEPA currently centers online example valuation: score windows by latent surprise, gradients, probe alignment, or metagradient value. Curriculum Mining adds a complementary axis: problem-start selection.

For latent-state time-series or JEPA training, the transferable interface is:

trajectory or event stream
  -> propose context endpoints, target start points, and horizon lengths
  -> mix locally solvable transitions with harder frontier transitions
  -> refresh the mixture as the model's latent state improves

The reset analogue is strongest when a simulator can generate valid states. For offline multivariate time series, the model cannot reset the real system; it can only resample valid observed windows, vary context/target geometry, or use a trusted simulator. Any generated restart must preserve temporal order, state reachability, exogenous variables, event history, and action/control-input semantics.

The project also sharpens the curriculum objective. Maximum novelty or maximum difficulty is not enough. A useful restart population should balance:

• learnability — short enough credit or prediction paths to produce signal now;
• novelty — states or transitions not already mastered;
• coverage — support across regimes, actions, rare events, and normal behavior;
• validity — physically, causally, and temporally reachable states;
• retention — enough familiar starts or opponent snapshots to prevent drift and forgetting.

Relation To OmniReset

Curriculum Mining explicitly credits OmniReset for the reset mechanism. The shared idea is broad initial-state coverage, but the evidence and reward contracts differ:

• Curriculum Mining uses random legal board states and sparse terminal rewards.
• OmniReset constructs manipulation-centric simulator resets and uses a fixed task-agnostic reward with dense reach, distance, and smoothness terms plus sparse success and safety terms.
• Curriculum Mining’s frozen opponent ring is an additional self-play stabilization mechanism, not part of OmniReset.

Limitations And Gotchas

• Blog/code/X release, not peer reviewed.
• One training run per game and no independent replication verified.
• Raw-policy versus raw-policy evaluation does not test MCTS-assisted playing strength.
• Reference-network wins are not state-of-the-art claims.
• Environment-interaction accounting is not total training-compute accounting.
• The opponent-pool mechanism is called essential but is not isolated by a captured public ablation.
• Legal random board positions are easy to generate compared with valid physical states or causal temporal windows.
• The finite game state spaces only approximate open-ended search.
• The social claims about GRPO and population-based start-state search are informal follow-ups.
• Code reuse terms are unclear because no license file was visible at the pinned revision.

Foundation TSFM Relevance

Agenda slot	Verdict	Evidence	Missing pieces
Data diversity, curriculum, and long tail	adjacent	Shows a concrete learner-dependent curriculum induced by a broad reset distribution: near-terminal states become learnable first and competence expands toward longer-horizon starts.	Needs public JEPA or multivariate time-series experiments, valid temporal-window generation, support-preservation probes, and matched-compute baselines.
Data valuation	adjacent	Treats each reset state as a candidate intermediate problem whose value depends on current learnability.	The release does not estimate marginal JEPA training effect, gradient value, or downstream rare-state utility.
Latent-state prediction	adjacent	Suggests varying context/target start points and horizon lengths so latent prediction moves from local to longer-horizon state transitions.	No representation-learning or latent-state experiment is reported.
Control and counterfactuals	insufficient evidence	The source trains game policies in known simulators with actions and terminal outcomes.	It does not learn an action-conditioned world model or validate alternative-action consequences in a learned dynamics model.
Benchmark hygiene	warning	The author explicitly narrows the win to raw policy networks and publishes interaction-count accounting and caveats.	Needs multiple training seeds, opponent-pool/reset ablations, stronger current baselines, and matched FLOPs/wall-clock accounting.

Links Into The Wiki

• Dynamic Curriculum Learning For JEPA
• OmniReset
• Implicit Curriculum Hypothesis
• World Models
• Robotics Time-Series Modeling
• Foundation Time-Series Model Research Agenda

Open Questions

• Does reset/start-state curriculum add value beyond surprise-based window reweighting for JEPA?
• Should a JEPA curriculum maintain a population over start time, context length, target horizon, regime, and action history rather than one scalar example score?
• Can learnability be estimated without biasing the sampler toward easy repeated states?
• Which novelty or quality-diversity objective preserves rare regimes without overselecting invalid or noisy transitions?
• Does a checkpoint/teacher bank analogue stabilize a changing JEPA sampler, or would it freeze stale representation geometry?
• How much of the reported board-game gain comes from random legal resets versus the frozen opponent ring, and do the two mechanisms interact differently across Othello, Hex, and Go?
• Can the reported board-game results be replicated across training seeds and stronger raw policy baselines under matched FLOPs and wall-clock cost?