Protein language models (pLMs) are powerful—but they mostly "read" sequences, not structures. That’s a problem for structural biology, where shape = function. Enter SaESM2 and SaAMPLIFY—a new framework that teaches pLMs the language of protein structure using contrastive and
@jiqizhixin
-
R1-Searcher++: Dynamic Knowledge Acquisition for LLMs
By
–
R1-Searcher++: Incentivizing the Dynamic Knowledge Acquisition of LLMs via Reinforcement Learning
Paper: https://
arxiv.org/pdf/2505.17005
.pdf
…
Code: https://
github.com/RUCAIBox/R1-Se
archer-plus
… -
R1-Searcher++ Advances Retrieval-Augmented Generation for LLMs
By
–
LLMs are smart—but static. They hallucinate when their internal knowledge runs dry. Retrieval-Augmented Generation (RAG) is one fix, but today’s RAGs are often expensive, hard to generalize, neglectful of the model’s own knowledge. R1-Searcher++ changes that. It’s a retrieval +
-
Soft Thinking: Enhancing LLM Reasoning in Continuous Concept Space
By
–
Soft Thinking: Unlocking the Reasoning Potential of LLMs in Continuous Concept Space
Paper: https://
arxiv.org/pdf/2505.15778
.pdf
…
Code: https://
github.com/eric-ai-lab/So
ft-Thinking
…
Page: https://
soft-thinking.github.io -
Soft Thinking: LLMs Reasoning Like Humans Through Soft Tokens
By
–
LLMs reason in discrete tokens, but human thought is fluid, continuous, and abstract. This paper from UCSB and UCSC asks: What if models could reason more like us? Their proposed Soft Thinking is a training-free method that generates soft concept tokens (i.e., weighted mixtures
-
Reward Reasoning Models: Microsoft’s Chain-of-Thought Innovation
By
–
How do we make reward models smarter without extra training data? Meet Reward Reasoning Models (RRMs) from Microsoft— they bring deliberate chain-of-thought reasoning into reward modeling. Instead of outputting a score instantly, RRMs think first, then score. The innovations
-
AdaptThink: Reasoning Models Learn When to Think Efficiently
By
–
Sure, reasoning models can learn when to think, but how? A team from Tsinghua University find that NoThinking is a better choice for relatively simple tasks in terms of both performance and efficiency. Nothing surprise here. but they propose AdaptThink too! AdaptThink is a
-
Fractured Chain-of-Thought Reasoning Advances AI Reasoning
By
–
Fractured Chain-of-Thought Reasoning
Paper: https://
arxiv.org/pdf/2505.12992 -
Fractured Sampling Boosts Reasoning Inference Without Retraining
By
–
Full and Long CoT boost reasoning by expanding intermediate steps—but at a high token cost. Not ideal for latency- or cost-sensitive apps. This paper introduces Fractured Sampling, a practical inference-time technique that turbocharges reasoning without retraining, using far
-
Moonshot’s Kimi Chatbot Expands Globally From China
By
–
Big news: Moonshot, creator of Kimi (one of China’s top chatbots), is going global!