Researchers propose ThinkJev to dynamically trade off speed and accuracy in typed decisions
ThinkJev combines a fast 'System 1' and a slower 'System 2' from a single hybrid-thinking model, enabling calibrated decisions without retraining.
1 source · cross-referenced
- ThinkJev introduces a two-system approach using one hybrid-thinking model to balance speed and accuracy in typed decisions.
Researchers at arXiv cs.AI describe ThinkJev, a method that pairs a fast 'System 1'—which reads probabilities from a direct forward pass—with a slower 'System 2' that reads the same probabilities after the model has generated a 'thought' sequence. Both systems operate over the same set of options, enabling calibrated decisions without retraining the underlying model.
The approach allows a single hybrid-thinking model to answer under any 'thinking budget,' as the thought tokens can be read after any number of generated tokens. This means one sampled thought yields answers for multiple computational budgets, reducing redundant computation when escalating from System 1 to System 2.
On a test set of 8,359 decisions, Qwen3.5-27B using ThinkJev answered 24.5% of questions via System 1 alone while maintaining a realized error rate of 3.9% against a 5% target. For the remaining questions, the system stopped thinking as soon as it was confident, achieving 81.6% accuracy with 72% fewer thought tokens compared to always thinking. Qwen3.5-4B and Qwen3.5-9B achieved 6.5% and 0.0% error rates, respectively, under the same target.
On the public subset of JevBench, an independent benchmark for typed-decision services, Qwen3.5-27B used System 2 for 29.4% of decisions and matched the accuracy of two proprietary thinking APIs. The authors note that relying solely on System 1’s confidence to decide when to think is less effective, as the decisions where System 1 is least confident are not always the ones improved by thinking.
- Nov 11, 2026 · The Verge — AI
Google DeepMind unveils AlphaGenome 2 to predict genetic variant effects on cell types
Trust79 - Nov 11, 2026 · Google DeepMind — Blog
Google DeepMind unveils AlphaGenome 2 to predict DNA variant effects in specific cell types
Trust84 - Nov 11, 2026 · arXiv cs.AI
Researchers demonstrate auditable human-agent teams for autonomous synchrotron X-ray experiments
Trust79