lmgame-static-eval

lmgame-static-eval

All 24 declared native panels are available as separately frozen evaluation leagues. Each submitted policy runs one complete selected panel per episode. Panel scores retain their native aggregation; this release defines no score across different panels. The fleet.json index records every panel, population, metric, and fixed reference roster. Qualification and coverage must be established separately for every selected panel.

LMGame-Bench

Twenty-four variants cover Sokoban, 2048, Tetris, and Candy Crush, with the original three-episode panels and step budgets. Each fixes text, vision, or both observations and base or agentic harness. Native prompts, parsers, action handling, game scoring, and per-call 50,000-token budgets remain intact. The suite seed initializes the native three successive episode seeds. Native YAML configurations record the observation and harness settings because the original CLI can discard underscore-named overrides. A private virtual display preserves native rendering. Model requests use the original vLLM profile; accepted responses and inline images remain in replay. Native action mappings rejected or skipped during a turn count once as an invalid response. Performance, environment score, reward, and step means are retained separately. Scores are not pooled across different games. Offline controls establish wiring, not model quality.

Super Mario Bros and Ace Attorney require ROM assets and are absent from this release. This release does not establish full six-game LMGame-Bench coverage.