When EOS Tokens Disagree: The Response-Length Blowup Problem in On-Policy Distillation
Mismatched EOS tokens inflate response length in on-policy distillation; unifying equivalent EOS as one stop curbs it.
DeepSeek-V4.1-Flash: A 1M-Token MoE Model That Cuts KV Cache to a Quarter
DeepSeek has released a 552B multimodal MoE model that shrinks KV cache footprint to a quarter of its predecessor—using CSA2, FP4 KV caching, and a CED architecture—while actually pushing performance higher.
NavHarness: A Lifelong Navigation Agent That Carries Experience Across Broken Conversations
NavHarness folds memory handling directly into the navigation loop, with no additional training, so a robot keeps reusing its map and exploration history even as sessions turn over. It sets new state-of-the-art results on GOAT-Bench and IR2R-CE.
The Hidden Flaw in PPO's Critic — "Value Flattening" — and SP³O, Which Supervises Only a Handful of Points
A paper that diagnoses "value flattening," where the PPO critic predicts sharply changing state values as nearly flat, and proposes SP³O, which mitigates it by supervising the critic at only a few points per response.
When Users Change Their Minds: Measuring and Repairing Intent Drift in LLM Agents
A study that quantifies "intent drift" — where already-abandoned pieces of a shifting user request contaminate the final output — with a measurement benchmark (IntentFlux), and partially recovers the lost performance using StateForge, which explicitly keeps the currently valid requirements front and center.
The Last AI Humans Will Build — A 5-Level Roadmap for Recursive Self-Improvement (RSI)
An honest walkthrough of a survey paper that sorts recursive self-improvement — where an AI fixes itself and eventually rewrites how it fixes itself — into autonomy levels B0 through L5, with every figure traced back to its source.
Thinking While Talking: A Deep Dive into StepAudio 3 Realtime, a Real-Time Voice AI
StepAudio 3 Realtime is an audio-language foundation model built to reason privately while it speaks, chasing both the snappy response times and the depth of thought that real-time voice conversation demands.
When AIs Trade Internal Representations Instead of Words, Team Safety Starts to Buckle
A CISPA study shows that in "latent communication," where multi-agent systems exchange internal representations instead of text, you can compromise a team's safety by tampering only with the connecting link, leaving the agents themselves untouched.
How to Remember Your First Lesson After 100 Rounds of Learning: Composing Continual Learning Mechanisms
When a language model learns 100 tasks one after another, catastrophic forgetting sets in. This is a plain-language look at research that combined three kinds of anchors with merged LoRA to push final retention from 1.2% to 34.9%, a 28x gain.
Reading the Atria Dawn Paper: Agentic Superintelligence and What's Left for Humans
A close read of the Atria Dawn paper (744B MoE agentic model, 16 benchmarks) and what it implies for AI PM products.
3-4x Faster AI Research Agents with World Models: A WMRL Walkthrough
WMRL replaces real experiment execution with a "world model" to kill the RL training bottleneck. Its 4B and 9B agents outperform 48B and 120B ones.
AuK: An Open-Source Speech Foundation Model That Both Creates and Repairs Audio
AuK is an open-source foundation model that unifies speech generation, editing, and restoration behind a single natural-language instruction. Here is a plain-English walkthrough of how its hybrid rectified-flow architecture and distillation get inference down to just four steps.