ThinkJev: threshold-gated System 1/2 reasoning that stops thinking once confident
RobertTLange · x · 2026-11-07
RobertTLange proposes ThinkJev, a simple way to add explicit reasoning to Jev-style decision models with typed outputs.
- System 1: a hybrid-thinking LLM answers directly, reading option probabilities from next-token logits
- If confidence falls below a dev-tuned threshold τ, System 2 kicks in: generate one reasoning trace, then re-score the same options
- Shared prompt prefix is cached; one trace can be read at multiple token budgets, stopping at the first confident read
The demo figure (Qwen3.5-27B) shows System 1 picking B at 0.35 confidence, triggering thought, then reaching 0.95 confidence on the correct answer C at just 512 of 1402 thought tokens; accuracy climbs with budget across all 8,359 test decisions. The author is skeptical logits are a good certainty measure but finds the idea worth exploring for adaptive-compute decision systems.
More from Research
- Scaling systolic array diagrams from 8x8 to 64x64: 4096 PE matrix multiply visualized — satnam6502 · 2026-11-07
- Aristotle formally verifies FP16 systolic matrix multipliers up to 64x64 in Lean 4 — satnam6502 · 2026-11-07
- Regularization amplifies subliminal learning instead of blocking hidden biases, thread finds — AdtRaghunathan · 2026-11-07
- Your LLM Is Secretly a Reward Model: Revisiting Stanford's DPO Paper — dotey · 2026-11-07
- RareDiseaseBench debuts: 35 reconstructed rare disease cases test AI diagnostic agents — kenbwork · 2026-11-07
- With ETH Refuted, GNN Chains Tackle the Still-Hard Graph Alignment Problem — marc_lelarge · 2026-11-07