Build, Don't Play: How LLM Research Agents Created a Superhuman Game AI
AutoTavern shifts AI agents from playing games directly to autonomously building simulators and RL policies, reaching a 15,000 MMR and crushing top human players.
The day's biggest AI story, explained at length with its sources. Page 2 of 12.
AutoTavern shifts AI agents from playing games directly to autonomously building simulators and RL policies, reaching a 15,000 MMR and crushing top human players.
Stanford and SambaNova introduce ACE, a tripartite framework that fixes context collapse and brevity bias to deliver self-improving LLM agents via delta-updated playbook…
By embedding TARFlow into the causal stream, UPenn and Meta researchers give continuous reasoning exact likelihoods, native KV-cache decoding, and direct RL optimization.
ByteDance and HKU researchers discover why causal attention undertrains distant tokens by 128,000x—and unveil a training-free positional shift that beats GPT-4 on RULER.
How Alibaba's dynamic GPU scheduler and prefix-deduplicating trajectory processor trained a 2.4T MoE across 700K-token rollouts, slashing end-to-end training time by 44%…
Cognizant AI Lab and UT Austin unveil MAKER, smashing the long-horizon reasoning cliff to complete 1,048,575 consecutive LLM actions without a single failure.
MIT HAN Lab shows that a large fraction of transformer attention heads (e.g., 50% to 75% depending on architecture) can be treated as streaming heads, enabling 3.3M cont…
As static MCP schemas consume up to 72% of agent context windows, engineers are pitting Unix CLI pipelines against JSON-RPC protocols to redefine autonomous tooling.
Researchers eliminate the sequential bottleneck of speculative decoding by running asynchronous pre-speculations on parallel GPUs to deliver up to 5x inference speedups.
MemTensor debuts open-weight 288B/17B and 36B/3B models using Gated DeltaNet-2 and Mixture-of-Storage, trading bloated KV caches for fixed-size persistent neural memory.
Google Cloud AI finds 34% of consensus claims in frontier agents are wrong, introducing an agentic framework that stress-tests outputs against ground-truth environment e…
NVIDIA and Georgia Tech introduce ThinKV, decomposing Chain-of-Thought traces into distinct thought classes to achieve near-lossless reasoning with 5.8x inference throug…