LLM
-
Tiny-vLLM Rebuilds the Inference Engine in C++ and CUDA So You Can Read Every Kernel
Tiny-vLLM is a fully functional LLM inference engine in C++ and CUDA — written deliberately small, with every kernel and every line of math walked through in a free companion course. It’s the “younger and smaller sibling” of vLLM, built for understanding rather than deployment. ## What’s actually in there The implementation is not a… Continue reading
-
LongTraceRL Mines Tiered Distractors From Search-Agent Traces for Long-Context RL
LongTraceRL trains long-context reasoning the way an agent actually experiences a long context — by reusing what real search agents do — and grades the model with entity-level rubric rewards instead of a single yes/no on the final answer. ## Tiered distractors from agent traces Long-context RL has been bottlenecked by sparse rewards and easy… Continue reading
-
ACC Turns Agent Trajectories Into Long-Context Training Data
ACC (Agent Context Compilation) tackles a clean idea: when agents solve problems, they produce massive trajectories — tool calls, environment observations, dozens of turns. That data is usually thrown away after the task. ACC converts it into long-context training material instead. ## How it works When an agent works a search, software-engineering, or database-querying task,… Continue reading
-
Alibaba’s Qwen 3.7-Max Runs 35 Hours Straight Across 1,158 Tool Calls
Alibaba announced Qwen 3.7-Max at the Apsara Cloud Summit, with API rollout already started. It’s positioned as Alibaba’s most capable agent model so far — built for long-running, multi-step workflows rather than single-prompt question answering — and ships with a 1M-token context window and a native extended-thinking mode. ## The 35-hour demo The standout number… Continue reading
-
A Power Law for LoRA Memory Pinpoints When Models Memorize Verbatim
A team from Zhejiang University and Alibaba turned LoRA into something it usually isn’t — a measurement instrument — and used it to pin down how much a model can actually memorize through low-rank fine-tuning. The headline is a Parametric Memory Law: a power-law relationship between loss reduction, effective parameters, and sequence length. ## A… Continue reading
-
LaRA Detects RL Post-Training Contamination by Watching Layer-Wise Representations
LaRA targets a fast-growing eval-integrity problem: as models go through reinforcement learning post-training, benchmark questions can quietly leak into the training data — and the model passes the test by remembering rather than reasoning. Layer-wise Representation Analysis is the proposed detector. ## Looking inside, not outside Most contamination detection compares answers to suspected sources from… Continue reading
-
StepFun Ships Step 3.7 Flash, a 198B Vision-Language MoE With Tunable Reasoning
StepFun released Step 3.7 Flash, a 198B-parameter sparse Mixture-of-Experts vision-language model aimed squarely at agentic workflows that mix perception, search, and reasoning. About 11B parameters activate per token, which is the lever that keeps a model this large at Flash-speed throughput. ## Tunable reasoning, native vision The architecture pairs a 196B language backbone with a… Continue reading
-
DenoiseRL Trains Reasoning Models to Recover From Their Own Wrong Starts
DenoiseRL is a reinforcement learning method that trains reasoning models on something most pipelines throw away: wrong answers. Instead of leaning on a stronger teacher model for supervision, it learns directly from the failures of weak models, turning bad reasoning traces into training signal. ## Conditioning on mistakes The trick is to start the model… Continue reading
-
MemTrace Pinpoints Why an AI Agent Memory Failed, Then Fixes It
MemTrace is a framework for a debugging problem that gets worse as agents get longer-lived: when an agent’s memory gives a wrong answer, why did it fail? Was it stored wrong, retrieved wrong, or lost along the way? MemTrace turns a memory pipeline into an executable “memory evolution graph” so you can trace the information… Continue reading
-
Bidirectional Evolutionary Search Lets Language Models Improve Themselves
Bidirectional Evolutionary Search (BES) is a new framework for self-improving language models that attacks a quiet weakness in how models search for answers: expanding from a single starting point keeps producing similar candidates. BES pairs a forward search that breeds variety with a backward search that creates feedback. ## Forward evolution, backward decomposition In the… Continue reading
Top AI Product 