4/5 The confusion: Scheduler saw "1 token allocated" and assumed decode. But 1 token can also mean "new request that hit token budget limits." Now fixed: check whether it's the request's first-ever token, not the allocation size. In vLLM v0.14.0 – update if running Mamba.
LLMS
-
GPU Memory Utilization Bug: Token Classification Mismatch
By
–
2/5 Used gpu_memory_utilization=0.2 to reproduce quickly, but this happens naturally when the scheduler runs out of token budget and GPU blocks get recycled. New request gets 1 token → misclassified as "decode" But num_computed_tokens=0 → should be "prefill".
-
SSM Recurrence Security: Mamba State Propagation vs Attention Safety
By
–
3/5 SSMs are recurrent, each token depends on previous state. Mamba decode reads state → updates it. Misclassified request reads garbage → propagates recursively through all tokens. Attention writes K/V first, then attends. Safe even if misclassified.
-
Debugging vLLM: Silent Corruption Bug in Jamba RL Training
By
–
1/5 Debugging vLLM: The silent corruption bug
1/1000 Jamba generations collapsed into confident gibberish during RL training. No crashes, no errors, just wrong outputs with high logprobs. The culprit? A scheduler edge case that only triggers under memory pressure. -
Reasoning Models: The Next Leap In AI With K2-Think
By
–
Why Reasoning Models Are The Next Leap In AI: Inside MBZUAI's K2-Think https://t.co/8u9SvKBQju
— Bernard Marr (@BernardMarr) 29 janvier 2026Why Reasoning Models Are The Next Leap In AI: Inside MBZUAI's K2-Think
-
Prompt ensemble boosts startup investment evaluation accuracy by 20%
By
–
Real example from Stanford's research: Task: Evaluate startup investment opportunity Single prompt accuracy: 67%
Prompt ensemble (5 variations): 87% accuracy Why? Different prompts caught different red flags. Synthesis found the pattern. 20% accuracy jump = millions in -
Prompt Ensembling Beats Single Prompts for High-Impact Decisions
By
–
When prompt ensembling beats single prompts: High-stakes business decisions ($100K+ impact) Medical/legal advice (where errors cost lives/money) Creative work (logo design, brand names, campaign ideas) Strategic planning (5-year roadmaps, market entry) Code
-
Ask your question 5 ways and combine answers for better results
By
–
The concept is simple: Instead of asking your question once and hoping for the best, you ask it 5 different ways and combine the answers. Think of it like getting second opinions from 5 doctors instead of trusting one diagnosis. Stanford tested this on GPT-5.2, Claude 4.5, and
-
Stanford prompt ensembling: merging 5 variations improves LLMs
By
–
Stanford researchers just published a prompting technique that makes today’s LLMs behave like better versions of themselves. It’s called “prompt ensembling” and it runs 5 variations of the same prompt, then merges the outputs. Here’s how it works
-
Style-conditioned AI models with prompt engineering capabilities
By
–
Oh yes. I liked the original Ai2 project for that reason, it's contained to a single codebase! So it can learn your style… I bet it'd be relatively easy to make it style-conditioned model as well, like a promptable feel/patterns/quality!