Peng Li · page 2
Papers
- G2: Guided Generation for Enhanced Output Diversity in LLMs (2025 · 6 citations)
LLMs often return near-identical answers when asked the same question repeatedly. This hurts creative writing, best-of-N selection, synthetic data creation and test-time scaling for reasoning. Raising the sampling temperature adds variety but lowers quality, and re-prompting with earlier answers can drift away from… - Writing-RL: Advancing Long-form Writing via Adaptive Curriculum Reinforcement Learning (2026 · 4 citations)
Large language models can now produce long articles, but further gains are hard to get. Supervised fine-tuning stalls because it can only imitate teacher or human outputs, and the RL recipes that work for math and code need verifiable answers that open-ended writing lacks. The authors propose Writing-RL, a… - MUSEG: Reinforcing Video Temporal Understanding via Timestamp-Aware Multi-Segment Grounding (2026 · 1 citation)
Multimodal LLMs remain weak at fine-grained temporal reasoning in video, such as locating when events happen. - Scaling External Knowledge Input Beyond Context Windows of LLMs via Multi-Agent Collaboration (2026 · 1 citation)
Large language models can use more retrieved evidence to solve harder tasks, but their context window caps how much they can take in. - Enabling Stroke-Level Structural Analysis of Hieroglyphic Scripts without Language-Specific Priors (2026)
Logographic and hieroglyphic scripts carry meaning in how strokes are arranged inside each character. Current LLMs treat characters as opaque tokens, and MLLMs see them only as pixel grids, so neither captures stroke structure. Earlier structural methods are usually tied to one script and depend on costly annotation… - GIFT: Guided Fine-Tuning and Transfer for Enhancing Instruction-Tuned Language Models (2026)
Adapting instruction-tuned LLMs to a specific domain is risky: fine-tuning them directly can disturb the balance of skills built during post-training. - InstructDiff: Domain-Adaptive Data Selection via Contrastive Entropy for Efficient LLM Fine-Tuning (2026)
Fine-tuning LLMs on entire instruction datasets is expensive and yields diminishing returns. Existing data-selection techniques also tend to work only in the domain they were designed for: methods tuned for general instruction-following transfer poorly to reasoning, and vice versa. InstructDiff addresses this with a…
This page in Markdown → https://md.research.elamur.ai/free/en/authors/peng-li/page/2/