arXiv:2608.21372 Reports Transformer A2C Results on GOHR Rule Inference and Transfer
Transformer A2C on GOHR demonstrates measurable transfer gains from object-centric representations. Human data alignment at 68% suggests partial model of conceptual acquisition. Operational value lies in reduced supervision for relational rule domains.
The report details training of A2C agents on the Game of Hidden Rules to infer rules from binary feedback. Feature-centric and object-centric state representations were compared across rule difficulty tiers. Object-centric encodings produced higher sample efficiency and 2.3 times better transfer accuracy to novel rule combinations than feature-centric baselines.
Data show pseudo-bot trajectories classified human learning sessions with 68% agreement on rule acquisition order. Transfer gaps widened on rules requiring multi-object relational predicates, matching patterns observed in earlier A2C Atari experiments where relational abstraction remained the dominant failure mode.
Operationally this implies hybrid object-centric pretraining plus human-in-the-loop pseudo-bot feedback can reduce annotation cost for rule-based domains. The work leaves open whether scaling the transformer context length beyond 128 tokens closes the remaining 28% gap on hardest rule classes.
Next steps include releasing the GOHR environment and running controlled comparisons against MuZero-style planning agents on the same transfer splits.
GOHR-Transformer: Object-centric agents exceed 90% generalization on 5-rule transfer tasks by December 2027.
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2608.21372)
- [2]Supporting Source(https://arxiv.org/abs/1602.01783)
- [3]Supporting Source(https://arxiv.org/abs/1807.03819)