DASH: dynamically adaptive self-distillation horizon¶

简体中文

Fidelity: core mechanism. Implements clipped local divergence, a stop-gradient gate, and backward horizon aggregation; candidate positions stand in for token positions.

Paper information¶

Field	Details
Paper	DASH (arXiv 2608.06243)
Organization	Nanjing University
First public release	2026-08-06 (arXiv v1)
Authors' code	DBtxy/DASH-OPSD is public.
Adapter / method	`dash`
Local reproduction code	`src/auto_research/post_training/`

Original paper¶

Motivation and method¶

Ordinary OPSD matches the privileged teacher separately at each token, making it hard to propagate credit from reliable later reasoning to earlier decisions. DASH uses local teacher/student divergence to form a stop-gradient gate and recursively aggregates weights backward. It obtains an adaptive distillation horizon without extra teacher forward passes.

flowchart RL
 D["clipped local divergence"] --> G["stop-gradient gate"]
 G --> A["backward horizon aggregation"]
 A --> L["weighted OPSD objective"]

Key figure from the paper¶

Figure 2: Overview of the paper's method and key components. From the original paper; copyright remains with the authors.

Core equations¶

\[ \lambda_t=\operatorname{sg}\sigma[-\kappa(d_t-\bar d)],\qquad A_t=d_t+\lambda_t A_{t+1},\qquad \mathcal L=\frac1T\sum_t A_t. \]

Paper results¶

Qwen3-1.7B averages 45.07 against 41.87 for OPSD; 4B scores 65.00 versus 63.60, and 8B 66.40 versus 64.80. No production A/B is reported.

Local reproduction¶

Arithmetic candidate suite, 120 steps, seed 42: accuracy 0.2344 → 0.5781 (+146.67%); final mean gate 0.4822, mean horizon coefficient 1.5617, and zero extra teacher forwards.

auto-research post-train --algorithm dash --dataset arithmetic-smoke --maximum-examples 256 --steps 120 --seed 42
auto-research evolve --model post-training --dataset arithmetic-smoke --direction "比较 OPSD 与 DASH 自适应 horizon" --generations 2 --population 4

See the fixed metrics.

Reproduction boundaries¶

Candidate positions are not complete token trajectories. Qwen3, mathematics reasoning data, and multi-GPU training were not reproduced.