Research Papers Topics
Adapting pretrained models to new tasks cheaply, including parameter-efficient methods like LoRA.
AI Agents · Shanghai Jiao Tong University
A hypernetwork compiles a textual skill into a LoRA adapter in one forward pass. On ALFWorld, LatentSkill lifts success by 21.4 points (seen) with 64.1% fewer prefill tokens.
Reinforcement Learning · Tianjin University
When you RL-tune an LLM across math, code, QA, and writing in sequence, math drops from 66.49 to 57.66 even though gradients look orthogonal. A short math refresh pulls it back to 66.04 without wrecking the other three.
LLM Reasoning · Shanghai AI Laboratory
ThoughtFold trims the redundant reasoning of DeepSeek-R1-Distill-Qwen-7B by about 56% of tokens while keeping accuracy on AIME, MATH-500, and GPQA-Diamond intact, using a masked preference objective.
Fine-Tuning & Adaptation · The Hong Kong Polytechnic University
Teachability-Aware OPD supervises only ~5% of tokens, those where the teacher's correction lands inside the student's top-K support, matching or beating full-token distillation (44.89 vs 42.37 on Qwen3-4B to 1.7B).
LLM Reasoning · Samsung Research
TrOPD masks on-policy distillation to the tokens where the teacher is actually trustworthy, adding +3.06 to +3.52 average points over standard OPD on math, code, and STEM benchmarks with 1.5B-1.7B students.
LLM Reasoning · Shanghai AI Laboratory
SU-01, a 30B-A3B open model from Shanghai AI Lab, hits 35 points on IMO 2025 and clears gold lines at IPhO 2024/2025 using only ~338K short SFT trajectories plus a 200-step two-stage RL pipeline.
AntiSD inverts self-distillation — it rewards tokens where a privileged context disagrees with the base model — reaching GRPO's accuracy in 2 to 10x fewer steps and ending up to 11.5 points higher on 4B-30B models.
Code Generation · University of Waterloo
Code2LoRA trains a hypernetwork to emit a repo-specific LoRA adapter for a code model with no inference-time token cost — 66.2% in-repo and 63.8% cross-repo exact match, plus an Evo variant that tracks diffs with a GRU.
LLM Reasoning · Renmin University of China
DelTA reweights RLVR updates so credit lands on tokens that actually separate right answers from wrong ones, lifting Qwen3-8B-Base by 3.26 and Qwen3-14B-Base by 2.62 average points over the strongest baselines.
LLM Reasoning · Alibaba Qwen Team
DVAO weights each reward by its in-group variance instead of fixed coefficients, lifting Qwen3-4B-Base from 38.99% to 42.19% average accuracy and length compliance to 99.91% in math-plus-tool-use RL.
Text-to-Image · University of Science and Technology of China
Flow-OPD trains one specialist teacher per reward, then distills them on-policy into one SD 3.5 student — lifting GenEval 0.63 to 0.92 and OCR 0.59 to 0.94 without the aesthetic collapse of multi-reward GRPO.
Fine-Tuning & Adaptation · HKUST
On-policy distillation does not sit between SFT and RLVR — it carves its own geometry. Its updates touch fewer weights, avoid principal directions, and lock into a narrow low-dimensional subspace early in training.
Efficient AI · Microsoft Research
LoRA freezes a pretrained model and trains tiny low-rank matrices per layer instead — cutting trainable parameters up to 10,000x and GPU memory 3x versus full GPT-3 175B fine-tuning, with no extra latency.
Speech Recognition · Shanghai AI Laboratory
Mega-ASR fights ASR's noise-robustness gap by synthesizing 2.4M clips across 54 compound acoustic scenarios, then training Qwen3-ASR-1.7B in two stages — cutting WER to 45.69% vs 54.01% on VOiCES R4-B-F.
Fine-Tuning & Adaptation · Mind Lab
MinT keeps one frontier base model resident and swaps only LoRA adapters, cutting the model-handoff step by 18.3x on a 4B dense model and 2.85x on a 30B MoE, while addressing million-scale adapter catalogs.
Fine-Tuning & Adaptation · Mind Lab
A position paper reframing LoRA adapters as persistent personal state, not a cheap full-finetune substitute, across three axes: scale up the base, scale down the adapter, scale out to millions, plus a serving stack MinT.
AI Agents · Zhejiang University
SDAR adds a gated, token-level self-distillation signal from a skill-augmented teacher on top of GRPO, lifting multi-turn agents by up to +10.2 points on WebShop and +9.4 on ALFWorld for small Qwen models.
AI Agents · University of Science and Technology of China
Skill1 trains a single Qwen2.5-7B policy to retrieve, apply, and create reusable skills under one task-outcome reward — reaching 97.5% on ALFWorld, 6.5 points over the strongest RL-only baseline.
Fine-Tuning & Adaptation · T-Tech
On-policy distillation wastes teacher supervision on a student's weak early rollouts. TRB blends teacher-like behavior inside a KL trust region during warmup, then anneals it to zero — best average on two math settings.