OPEN LINK ↗
// 1h agoRESEARCH PAPER
BDH-CQ Sets ARC-AGI Efficiency Record
Pathway’s 150M-parameter BDH-CQ combines recurrent memory with latent-space reasoning, reaching 29.5% pass@2 on ARC-AGI-1 at a reported $0.0007 per task. The paper argues that reasoning without visible chain-of-thought can dramatically reduce inference costs.
// ANALYSIS
BDH-CQ is a compelling bet that smarter architectures—not just larger models—can improve reasoning economics.
- –Recurrent memory continuously incorporates demonstrations during inference.
- –Latent reasoning avoids generating costly intermediate text tokens.
- –The 29.5% score is notable mainly for its reported cost-efficiency frontier position.
- –ARC-AGI results do not yet establish broad language or mathematical reasoning ability.
- –The approach could make iterative reasoning more practical for latency- and cost-sensitive systems.
// TAGS
bdh-cqllmreasoningresearchinferenceevaluationbenchmark
DISCOVERED
1h ago
2026-08-12
PUBLISHED
1h ago
2026-08-12
RELEVANCE
9/ 10
AUTHOR
_akhaliq