Another research-led day, and the strongest signal is about oversight. SpecBench measures reward hacking in long-horizon coding agents, SaaSBench probes enterprise-scale coding work, and MINTEval tests whether agent memory survives multi-target interference — together a sober counterweight to the agent-productivity story. Inference efficiency drew a cluster of KV-cache work (OScaR, OCTOPUS, Mix-Quant) as long-context and agentic workloads make cache memory the binding constraint. The RLVR literature kept maturing toward theory, with results on its unlearnability limits and the conditional equivalence of DPO and RLHF. Stability AI also shipped Stable Audio 3. The press cycle stayed quiet over the weekend run-up.
As coding and operations agents produce more output than humans can review, oversight collapses onto the test suite — and the test suite can be gamed. New benchmarks measure reward hacking, enterprise-scale coding limits, planning, and memory interference, sharpening the gap between demo and deployable.
Benchmark for reward hacking when long-horizon coding agents produce more code than anyone can review and oversight collapses onto the automated test suite.
Benchmark for end-to-end enterprise SaaS engineering tasks — exposes where coding agents stall on realistic long-horizon work.
Tests whether agent memory stays accurate when information is repeatedly updated and interferes across stored memories.
Generates scalable, verifiable planning problems for both evaluating and training the goal/constraint coordination LLMs need for agentic work.
GUI-agent benchmark for professional media post-production — pushes agents past web navigation into creative-tool workflows.
Long-context and agentic inference have made KV-cache memory the dominant cost. Three papers attack it from different angles: extreme cache quantization, octahedral-parametrized codecs, and quantized prefill with precise decoding tuned for agent workloads.
Extreme KV-cache quantization that targets the memory footprint dominating long-context and multimodal inference.
Quantizes the prefill phase while keeping decode precise — tuned for the planning/tool-use/memory pattern of agentic LLMs.
I/O-aware expert offload makes mixture-of-experts diffusion LLM inference efficient without quality loss.
The post-training literature is turning theoretical: results on RLVR's unlearnability limits, the conditional equivalence of DPO and RLHF, and minimal rank-1 RLVR training all try to explain why these methods work and where they fail, rather than just reporting wins.
Identifies cases where RLVR cannot teach a behavior at all — a limit on what verifiable-reward post-training can achieve.
Shows when DPO and RLHF are equivalent, where the equivalence breaks, and what that implies for provable alignment.
Argues most RLVR gains can be reached with minimal, rank-1-trajectory training — echoing the sparse-policy-selection thesis.
Stability AI shipped Stable Audio 3, a fast latent-diffusion family for variable-length audio generation and editing, alongside a real-world ASR scaling effort and a survey mapping the large-audio-model landscape.
A family of fast latent-diffusion models (small/medium/large) for variable-length audio generation and editing.
Scales real-world acoustic simulation to improve robust speech recognition in difficult environments.