SynthSelf enables continual post-training of small models with +22pp acquisition and only -1pp forgetting
LG-AI-Research · hf · 2026-10-12
KAIST AI and LG AI Research introduce SynthSelf, tackling the dual challenge of acquisition and retention in continual post-training. Existing SDFT relies on strong in-context learning and fails on small models — precisely those needed for low-latency, on-device use.
- Root cause: poor acquisition and forgetting both correlate linearly with the post-trained model's distribution shift, showing in-distribution learning is essential.
- Method: SynthSelf mimics SDFT's in-distribution rewrite by outsourcing the task to a minimally larger, same-family aligner model, coupled with off-policy supervision and rejection sampling, delivering high-quality supervision without breaking the model's original distribution.
- Results: Across 4 acquisition datasets and 14 retention benchmarks at 3 small-model scales, SynthSelf consistently outperforms 6 methods including off- and on-policy distillation — gaining roughly +22pp acquisition while losing only 1pp retention versus the base model.
More from Research
- SoftwareGym synthesizes 182 app environments, lifting CUA OSWorld Pass@3 to 56.97% — Tian Jin · 2026-10-12
- Qualcomm AI uses entropy and CKA to pick layers, beating EAGLE3 defaults in speculative decoding — Qualcomm-AI-Research · 2026-10-12
- Evolved e-puck robots show communication emerges with information value — uv-mex · 2026-10-12
- PreSTO Speeds Up LLM Power Sampling Up to 2.33x via Predictive Subtree Prefetching — du_yilun · 2026-10-12
- AWS+HSBC paper co-evolves agent graders from just 10 labeled examples, matching ground truth — rohanpaul_ai · 2026-10-12
- Nature MI paper: LLMs discover complementary heuristics for combinatorial optimization — lulzxdxdxd · 2026-10-12