I trained TIME: short context-triggered thinking on Qwen model instead of overthinking
Summary
A personal project led to an ACL 2026 paper introducing TIME, a method training Qwen3 models to engage in short, context-triggered thinking rather than excessive reasoning. The work uses QLoRA and a four-phase curriculum, with all data and code released open-source.
Similar Articles
LessThink-Qwen3-4B: the same model, with far less thinking [P]
The author post-trained Qwen3-4B to reduce token usage by 44% in reasoning while preserving knowledge and style, with training completed on a single GPU.
TEMPO: Scaling Test-time Training for Large Reasoning Models
TEMPO introduces a test-time training framework that alternates policy refinement with critic recalibration to prevent diversity collapse and sustain performance gains in large reasoning models, boosting AIME 2024 scores for Qwen3-14B from 42.3% to 65.8%.
Teaching Thinking Models to Reason with Tools: A Full-Pipeline Recipe for Tool-Integrated Reasoning
This paper presents a full-pipeline recipe for teaching thinking models to reason with tools, achieving state-of-the-art performance on benchmarks like AIME 2025 when applied to Qwen3 models.
Qwen 3.8 27B Overthinking, It has to be done, it has to be overthinking to punch Opus 4.6
The article discusses Qwen 3.8 27B, a 27B parameter model that uses extensive reasoning tokens to compete with larger models, emphasizing trade-offs in token usage and benefits for local deployment.
I asked 3 Qwen 3.8 models to solve recent AtCoder problems at 4 thinking efforts
A community member benchmarked three Qwen 3 quantized models (Flash, Swift, 27B) at four thinking effort levels on 14 recent AtCoder problems, finding that higher thinking effort rarely improves results and mainly increases token-capped failures, with Flash IQ3_S emerging as the best speed-accuracy compromise.