🔥Telegram İfşa

Cool new paper from @aiatmeta and brilliant idea. Proposes DARLING, a training method that makes LLM answers both higher quality and more varied. Unlike traditional reinforcement learning, which only rewards a model for producing the single highest quality response, DARLING

3T tokens, ~1800 languages, 2 models - we’re releasing mmBERT, a modern multilingual encoder model!

Today Thinking Machines Lab is launching our research blog, Connectionism. Our first blog post is “Defeating Nondeterminism in LLM Inference” We believe that science is better when shared. Connectionism will cover topics as varied as our research is: from kernel numerics to

I was curious about the voice agent, particularly its ability for tool-use and reasoning, while listening to a user. So we train voice agents by building a sandbox (based on tau-bench) that uses GPT-4.1 for user simulation and SEED-TTS for speech synthesis. #Agents #ToolUse

Thanks, Wenhao, for sharing our work—and congrats on yours as well. For clarity: our classifier is a fine-tuned embedding model tailored to specific definitions of diversity. A natural next step is to use an LM judge to assess diversity directly in online RL.

(🧵) Today, we release Meta Code World Model (CWM), a 32-billion-parameter dense LLM that enables novel research on improving code generation through agentic reasoning and planning with world models. ai.meta.com/research/publi…

LoRA makes fine-tuning more accessible, but it's unclear how it compares to full fine-tuning. We find that the performance often matches closely---more often than you might expect. In our latest Connectionism post, we share our experimental results and recommendations for LoRA.

🛑 What's the Flaw of Averages? 📄: arxiv.org/abs/2509.25671 We’re in an evaluation crisis. Benchmarks are saturating, creating a false sense that tasks are solved. As training/eval chase these sets, plateaued averages hide shortcutting and distributional skew. 🧵1/7

🌀New Self-Driven RL Method: RESTRAIN 🌀 📝: arxiv.org/abs/2510.02172 - RESTRAIN turns spurious votes → self-Improving signals. No labels needed - Does this through self-penalizing unreliable reasoning paths: ✔️ Uses all rollouts, not just the majority, ✔️ Offsets

🚨Prompt Curriculum Learning (PCL) - Efficient LLM RL training algo! - We investigate factors that affect convergence: bsz, # prompt, # gen, prompt selection - We propose PCL: lightweight algo that *dynamically selects intermediate-difficulty prompts* using a learned value model

Collecting large human preference data is expensive—the biggest bottleneck in reward modeling. In our #NeurIPS2025 paper, we introduce latent-space synthesis for preference data, which is 18× faster and uses a network that’s 16,000× smaller (0.5M vs 8B parameters) than

Wenhao Yu and I are recruiting 2026 Spring/Summer Research Interns at Tencent AI Lab 🚀 Topics include self-evolving, Agent Systems, Complex Reasoning, etc. We are also hiring full-time researchers with PhD degrees, fully publication-driven. Please DM or email.

🤖➡️📉 Post-training made LLMs better at chat and reasoning—but worse at distributional alignment, diversity, and sometimes even steering(!) We measure this with our new resource (Spectrum Suite) and introduce Spectrum Tuning (method) to bring them back into our models! 🌈 1/🧵

Hybrid Reinforcement (HERO): When Reward Is Sparse, It’s Better to Be Dense 🦸‍♂️ 💪 📝: arxiv.org/abs/2510.07242 - HERO bridges 0–1 verifiable rewards and dense reward models into one 'hybrid' RL method - Tackles the brittleness of binary signals and the noise of pure reward

💃New Multi-Agent RL Method: WaltzRL💃 📝: arxiv.org/abs/2510.08240 - Makes LLM safety a positive-sum game between a conversation & feedback agent - At inference feedback is adaptive, used when needed -> Improves safety & reduces overrefusals without degrading capabilities! 🧵1/5

it’s tokenization again! 🤯 did you know tokenize(detokenize(token_ids)) ≠ token_ids? RL researchers from Agent Lightning coined the term Retokenization Drift — a subtle mismatch between what your model generated and what your trainer thinks it generated. why? because most

How does training data shape model behavior? Well, it’s complicated… 1/10

publish.obsidian.md/ueaj/Machine+L…

Thanks for sharing our work Nate! In our work we found that explicitly optimizing for diversity allows the model to beat baseline GRPO in both pass@1 & pass@k. The code is now open-sourced at github.com/facebookresear….