Razvan Dumitru

@rjsabouhi @maenstru56 @Vikas_NLP_UA @PanLiangming Exactly. ConciseRL formalizes that intuition: encode “be concise yet correct” in the reward, let PPO follow the gradient, and the attractors are short-but-sufficient chains. Appreciate the gradient-field lens!

@RazvanDuu @maenstru56 @Vikas_NLP_UA @PanLiangming You already knew the attractors were encoded in the field. This makes it formal. “Gradient fields” learn the subtask structure hidden in the reward. Latent constraint flow as behavior. Symbolic recursion in motion.

Thrilled to share that my first author paper ConciseRL is accepted to EMNLP 2025 (Findings). We train models to be right and succinct with a simple LLM-judged conciseness reward. Explanations below 👇 Paper: arxiv.org/abs/2505.17250 Code: github.com/RazvanDu/Conci… (1/6)

+2.2 accuracy points using ~12.5× fewer tokens on TheoremQA. Length adapts to difficulty; traces read cleaner; drops smoothly into existing RL pipelines. (5/6)

On MATH: up to 31× fewer tokens on easier problems with ~+7 accuracy points; on the hardest tier: ~3.6× fewer tokens with +7.5 accuracy points. (4/6)

Reward semantic conciseness (not just shortness). An LLM judge scores whether a chain is dense and sufficient; combine with correctness so we trim fat without losing substance. (3/6)

Reasoning models overthink—long chains past the answer. That burns tokens, slows inference, and muddies evaluation. (2/6)

Co-authors: @Yminglai @Vikas_NLP_UA @msurd — thank you! See you in Suzhou, Nov 4–9. 🙏 #EMNLP2025 (6/6)

Big news: My first author paper CopySpec is accepted to EMNLP 2025 (Main). If the next tokens just repeat the context, we stop re-writing and copy (then verify). I’ll break it down in the replies 👇 Paper: arxiv.org/abs/2502.08923 Code: github.com/RazvanDu/CopyS… (1/6)

Up to 3.08× faster (second-turn MT-Redundant) and +49% on top of speculative decoding; benefits grow with longer context. Details in the paper. (5/6)

Drop-in on top of speculative decoding: rolling hashes find matches → a copy-and-verify path runs alongside your drafter → target model accepts/rejects. No special memory tricks. (4/6)

Real workflows repeat: multi-turn chat, RAG follow-ups, summaries, code edits. That redundancy is free speed if you can safely reuse instead of regenerate. (3/6)

CopySpec watches the running text, spots a chunk that already exists in the context/previous turn, proposes a “copy block,” and lets the target model check it. Same output, less generation. (2/6)

🎉 Our work “Variable Layerwise Quantization: A Simple and Effective Approach to Quantize LLMs” is accepted at #ACLFindings2025 📎 arxiv.org/abs/2406.17415 — Keep key layers high-precision, push others lower → compact LLMs w/ ~no accuracy loss — Simple LIM & ZD scores rank layers

🧵Excited to attend #ICML2024 and present our paper titled "Enhancing Transformer RNNs with Multiple Temporal Perspectives" at the "Next Generation of Sequence Modeling Architectures" workshop! #AI