WISD
A self-distillation method that requires no external teacher or privileged information; it constructs a weakened copy of the original model through low-bit quantization or magnitude pruning and uses the behavioral differences between the original and weakened models on unlabeled prompts as a training signal to improve the reasoning ability of large reasoning models.
Timeline (last 90 days)
The paper "Weaken to Strengthen: Weakening-Induced On-Policy Self-Distillation for Large Reasoning Models" proposes the WISD method, which actively weakens the model and uses the contrast before and after weakening as a training signal, without requiring an external teacher or privileged information.
WISD constructs weakened models through two approaches, low-bit quantization and magnitude pruning; in the method these two operations are used to generate training signals rather than for deployment optimization.
On math and code reasoning benchmarks, WISD improves average accuracy by 4.0 percentage points with quantization-based weakening and by 5.5 percentage points with pruning-based weakening.
In WISD, the average improvement from pruning is higher than that from quantization, and is especially pronounced on small models.
WISD adopts on-policy self-distillation, where the training signal comes from comparing parallel trajectories generated in real time by the original model and the weakened model on the same unlabeled prompt, with the comparison occurring in policy space rather than answer space.
WISD can operate relying only on unlabeled prompts, without requiring an external teacher or privileged information.
WISD's weakened model shares the same origin as the original model, sharing pretrained knowledge, tokenization, and reasoning style, thereby eliminating the teacher-student distribution mismatch problem.
WISD has so far been validated only on math and code reasoning tasks, and whether it applies to broader natural language reasoning tasks remains unclear.
The choice of weakening strength in WISD lacks theoretical grounding and may require tuning for different models and tasks.
All Facts (9)
The paper "Weaken to Strengthen: Weakening-Induced On-Policy Self-Distillation for Large Reasoning Models" proposes the WISD method, which actively weakens the model and uses the contrast before and after weakening as a training signal, without requiring an external teacher or privileged information.
50%UnverifiedWISD constructs weakened models through two approaches, low-bit quantization and magnitude pruning; in the method these two operations are used to generate training signals rather than for deployment optimization.
50%UnverifiedOn math and code reasoning benchmarks, WISD improves average accuracy by 4.0 percentage points with quantization-based weakening and by 5.5 percentage points with pruning-based weakening.
50%UnverifiedIn WISD, the average improvement from pruning is higher than that from quantization, and is especially pronounced on small models.
50%UnverifiedWISD adopts on-policy self-distillation, where the training signal comes from comparing parallel trajectories generated in real time by the original model and the weakened model on the same unlabeled prompt, with the comparison occurring in policy space rather than answer space.
50%UnverifiedWISD can operate relying only on unlabeled prompts, without requiring an external teacher or privileged information.
50%UnverifiedWISD's weakened model shares the same origin as the original model, sharing pretrained knowledge, tokenization, and reasoning style, thereby eliminating the teacher-student distribution mismatch problem.
50%UnverifiedWISD has so far been validated only on math and code reasoning tasks, and whether it applies to broader natural language reasoning tasks remains unclear.
50%UnverifiedThe choice of weakening strength in WISD lacks theoretical grounding and may require tuning for different models and tasks.
50%