Training a single model to answer a yes/no in under 50 tokens and also work through a 48-step AIME proof sounds straightforward. A new paper from Universiteit van Amsterdam — "Fusion Training for Mathematical Generalization in Large Language Models" — shows it is not. Researchers Congfeng Cao, Pengyu Zhang, and Jelke Bloem swept Thinking Mode Fusion (TMF) training schedules and thinking-to-non-thinking data ratios. Their central finding: adding non-thinking supervision degrades thinking-mode accuracy. The two modes conflict, and the optimal schedule shifts with the data ratio.

TMF is already in production. Qwen3's four-stage pipeline uses thinking mode fusion as its third stage, mixing chain-of-thought data with instruction-tuning data into one checkpoint. The benefit is obvious: a single model, a control token in the prompt (<think> or \no_think), and no need to serve two separate weights. The Fusion Bench paper is the first systematic study of what the training recipe costs across schedules and ratios.

Recommended two-phase Thinking Mode Fusion training schedule: think-only pre-training followed by fused think+no-think stage, yielding a single model switchable via control token.
FIG. 02 Recommended two-phase Thinking Mode Fusion training schedule: think-only pre-training followed by fused think+no-think stage, yielding a single model switchable via control token. — Demystifying Hybrid Thinking, arXiv 2510.12680; Fusion Bench paper (UvA)

The asymmetry is the core result. Raising the non-thinking data proportion does not add a new capability — it erodes the existing one. The authors quantify a negative correlation: the more you reinforce short direct answers, the more you weaken chain-of-thought reasoning. Three training schedules tested against multiple thinking-to-non-thinking ratios showed no universal recipe. The optimal schedule varies with the ratio.

Think-mode accuracy vs. no-think accuracy on Qwen3-8B: a 9.9-point gap on MATH500 widens to 39 points on AIME24, quantifying the cost of non-thinking supervision.
FIG. 03 Think-mode accuracy vs. no-think accuracy on Qwen3-8B: a 9.9-point gap on MATH500 widens to 39 points on AIME24, quantifying the cost of non-thinking supervision. — Demystifying Hybrid Thinking, arXiv 2510.12680

Case Western and Meta AI published concurrent confirmation ("Demystifying Hybrid Thinking," arXiv 2510.12680). Testing Qwen3-8B on MATH500, think mode hits 92.82% at 4,384 tokens; no-think mode drops to 82.90% at 958 tokens. On AIME24 the gap widens to 39 points: 63.33% think versus 24.00% no-think, with no-think still consuming 4,062 tokens — more than twice the pure-instruct baseline of 1,729. Reasoning bleeds through: even with \no_think set, Qwen3-8B emits 646 "wait" tokens on MATH500, versus zero from Qwen2.5-7B-Instruct.

BenchmarkMode / ModelAccuracy (%)Avg Output Tokens"Wait" Tokens (MATH500)
MATH500Think — Qwen3-8B92.824,384—
MATH500No-Think — Qwen3-8B82.90958646
MATH500Instruct baseline — Qwen2.5-7B—1,7290
AIME24Think — Qwen3-8B63.33——
AIME24No-Think — Qwen3-8B24.004,062—
FIG. 04 Qwen3-8B think vs. no-think mode: accuracy and average output token counts per benchmark (CWRU / Meta AI evaluation) — Demystifying Hybrid Thinking, arXiv 2510.12680

That optimized training recipe cuts no-think output from 1,085 to 585 tokens on MATH500 and "wait" occurrences from 5,917 to 522. The recipe requires four elements: 140,000+ training samples for stable mode switching, thinking and non-thinking answers from different questions (paired data damages controllability), a moderate tilt toward no-think data, and a two-phase schedule — think-only first, then fuse. Two-phase scheduling versus random mixing at 20k samples: no-think output shrinks from 2,214 to 870 on MATH500 and from 5,654 to 1,847 on AIME24.

ComparisonMetricBeforeAfter
Optimized vs. Naive RecipeMATH500 no-think output (tokens)1,085585
Optimized vs. Naive RecipeMATH500 "wait" occurrences5,917522
Two-Phase vs. Random Mixing (20k samples)MATH500 no-think output (tokens)2,214870
Two-Phase vs. Random Mixing (20k samples)AIME24 no-think output (tokens)5,6541,847
FIG. 05 Impact of training recipe optimizations on no-think output length and reasoning bleed-through ("wait" token occurrences) — Demystifying Hybrid Thinking, arXiv 2510.12680

The fundamental tension remains unresolved. Fusion Bench quantifies the trade-off; the CWRU/Meta work minimizes it but cannot eliminate it. Hybrid models cannot achieve full mode separation. Routing simple queries through no-think mode after training with aggressive thinking ratios leaves latency and cost gains on the table. The non-thinking mode has not fully shed its reasoning character.

For teams choosing between unified checkpoints and separate think/no-think models: training-time decisions on data ratio and schedule directly determine your no-think cost floor. A unified model trained at naive 1:1 ratio with no phased schedule will emit reasoning tokens on trivial queries. If you did not control the training recipe, treat no-think output lengths as measured values, not assumptions.