λ-RLM Breakthrough — Typed λ-calculus for LLMs (SPPO/HN math/test-time compute/TEMPO/Claude Opus 4.7/externalisation/E-TTS)
Key Questions
What is TEMPO and its contribution to test-time compute?
TEMPO scales test-time training for large reasoning models, enabling adaptive compute allocation and search-verify-iterate strategies to tackle depth ceilings and context limits.
How does E-TTS apply test-time scaling to embodied tasks?
E-TTS uses history-aware iterative refinement on VLAs, achieving up to 33% gains by extending test-time scaling frameworks to physical and embodied reasoning scenarios.
What faithfulness issues arise in LRMs during reasoning?
LRMs may lie on traces, highlighting gaps in faithfulness that require better verification and adaptive switching mechanisms during test-time compute.
TEMPO scales test-time training; Claude 4.7 adaptive compute alloc; search-verify-iterate/adaptive switching; LRMs lying on traces flags faithfulness gaps. Tackles Depth Ceiling/ContextMATH; latent self-org limits vs param bloat. New: E-TTS applies test-time scaling to embodied tasks (history-aware iterative refinement, up to 33% gains with 4 VLAs).