Generally speaking, diffusion LLMs (dLLMs) are faster than autoregressive LLMs (AR LLMs), but dLLMs can be made even faster. Nvidia researchers have proposed a Fast-dLLM. By enabling KV caching and parallel decoding, Fast-dLLM achieves training-free acceleration. Two Key
@jiqizhixin
-
BARL: Bayesian Adaptive RL for Reflective LLM Exploration
By
–
This Google paper proposes BARL — Bayesian Adaptive RL for Reflective Exploration. What it can do: – Encourages reflective behaviors to emerge naturally during training.
– Guides LLMs to explore when needed, rather than relying on static policies.
– Results in fewer tokens used -
AgentDNS: Root Domain Naming System for LLM Agents
By
–
AgentDNS: A Root Domain Naming System for LLM Agents
Paper: https://
arxiv.org/pdf/2505.22368
.pdf
…
Code: https://
github.com/agentdns (coming soon?) -
DeepSeek-R1-0528 Now Available on Hugging Face
By
–
DeepSeek-R1-0528 is now on Hugging Face！ https://
huggingface.co/deepseek-ai/De
epSeek-R1-0528/tree/main
… -
Large Reasoning Models Self-Training Capabilities Study
By
–
Can Large Reasoning Models Self-Train?
Paper: https://
arxiv.org/pdf/2505.21444
v1.pdf
… -
Self-Consistency Training for LLM Scaling Without Human Supervision
By
–
Scaling LLMs through Reinforcement Learning (RL) usually needs human-crafted verifiers or gold answers, which limits scalability. Can models train themselves, without external supervision? This paper propose: using the model’s own self-consistency (i.e., agreement across
-
DeepSeek-R1 Model Upgraded with Enhanced Deep Think Features
By
–
DeepSeek-R1 upgraded？ DeepSeek just posted a notice in their WeChat group： [Notice] The DeepSeek R1 model has completed a minor version trial upgrade. You are welcome to go to the official website, APP, and mini program for testing (open Deep Think). The API interface and
-
RRO: LLM Agent Optimization Through Rising Reward Trajectories
By
–
RRO: LLM Agent Optimization Through Rising Reward Trajectories
Paper: https://
arxiv.org/pdf/2505.20737
.pdf
… -
LLM agents struggle with multi-step reasoning and scaling
By
–
Large language model agents still struggle with multi-step reasoning, where one misstep can collapse the whole plan. While Process Reward Models (PRMs) aim to correct reasoning step-by-step using RL, they don’t scale well due to the high cost of evaluating tons of action
-
Distillation Outperforms Zero-RL Through Flexible Reasoning
By
–
Why Distillation can Outperform Zero-RL: The Role of Flexible Reasoning
Paper: https://
arxiv.org/pdf/2505.21067
.pdf
…