Four threads dominate today's research: agent loop control and rubric-level reward design, long-horizon video memory, agent-security guardrails, and low-cost open pretraining. LoopArena and Rubric-to-Code Credit Assignment both push toward finer-grained control of coding agents, with the latter's RCCA-trained model edging past Claude Opus 4.5 and GPT-5 on two app-building benchmarks. Security scrutiny of agent automation intensified from both directions: a public write-up broke Claude Code's Opus 5 auto mode, while StepGuard and LMSM propose pre-execution and kernel-style guardrails to catch exactly this class of exploit. Ring Forcing and LayerRecall separately tackle the same video-diffusion weakness, letting generated video retain object identity across minutes rather than seconds. Puro-2B's under-$7,000 pretraining recipe, released with full data and code, lowers the bar for reproducible LLM research outside frontier labs. Outside research, the U.S.-China drone and robotics standoff, Texas's Flock camera funding freeze, and Meta's Pocket game-creation tool all point to the same theme: AI infrastructure and tooling decisions are increasingly being made in courtrooms and statehouses, not just labs.
A public write-up broke Claude Code's Opus 5 auto mode, the exact scenario step-level and kernel-style guardrail research is racing to address. StepGuard checks tool calls before execution rather than after, and LMSM adapts the Linux Security Modules pattern to separate interpretability signals from enforcement in LLM serving, together pointing at pre-execution mediation as the next defense layer for agentic automation.
A security researcher demonstrates techniques for breaking Claude Code's Opus 5 auto mode, drawing significant Hacker News attention.
Two long-form technical explainers on diffusion-based language modeling landed on Hacker News the same day and both drew triple-digit point totals, an unusually strong signal of renewed community interest in diffusion as an alternative to autoregressive LLMs.
A widely-read blog post walks through building a diffusion language model from scratch, drawing 109 Hacker News points.
A companion explainer on continuous diffusion language models draws similar traction (113 points), reflecting renewed interest in diffusion as an LLM alternative.
New benchmarks and training methods target the layer above raw coding-agent capability: directing a fixed agent through a long task, and routing reward signal to the code spans that earned it. LoopArena finds even the best controllers succeed on full tasks under 25% of the time; Rubric-to-Code Credit Assignment edges past Claude Opus 4.5 and GPT-5 on two app-building benchmarks by localizing reward.
LoopArena benchmarks how well one model can act as a 'Controller' directing a fixed coding-agent 'Worker' through long tasks, finding the best full-task strict success rate is still under 25%.
RCCA converts rubric-level functional feedback into token-level RL rewards for code generation; the resulting model edges out Claude Opus 4.5 and GPT-5 on two app-building benchmarks.
DART-SD retrains multi-turn tool-calling agents by localizing correction to specific recovery steps in a trajectory graph, instead of penalizing entire successful-but-different paths.
A survey of 259 works defines 'agentic artifact creation' as stateful, verification-guided construction of deliverables, and distills principles for keeping repair targeted and state revalidated after change.
Two independent papers address the same weakness in autoregressive video diffusion: recency-based context windows drop object identity and scene state once minutes of video have passed. Ring Forcing and LayerRecall both add structured memory retrieval to fix it, while separate work speeds up spatio-temporal grounding and repurposes video generators for geometry estimation.
Ring Forcing adds ring-structured training and KV compression to autoregressive video diffusion, letting generated video hold object identity and scene state over minutes-long spans.
LayerRecall routes retrieved long-range memory into only the video-diffusion layers that are memory-sensitive, improving recall of reappearing subjects with negligible inference overhead.
Parallel Tube Decoding reformulates video grounding to decode spatial boxes simultaneously instead of frame-by-frame, cutting tube completion latency 79x while improving accuracy.
GeoNeXt repurposes pretrained video generative models for zero-shot depth and surface-normal estimation, rivaling discriminative models trained on over 100x more data.
A public write-up broke Claude Code's Opus 5 auto mode, the exact scenario step-level and kernel-style guardrail research is racing to address. StepGuard checks tool calls before execution rather than after, and LMSM adapts the Linux Security Modules pattern to separate interpretability signals from enforcement in LLM serving, together pointing at pre-execution mediation as the next defense layer for agentic automation.
StepGuard checks agent tool calls before execution rather than after the fact, cutting attack success rate 77.3% on AgentDojo/AgentDyn with only a 2.8-point utility cost.
LMSM borrows the Linux Security Modules pattern to separate interpretability-derived signals from policy and enforcement in LLM serving, cutting HarmBench attack success from 39.2% to 3.32% on Qwen3-4B.
Puro-2B releases a full pretraining recipe, data, code, and weights for a model approaching Qwen2-1.5B performance, trained on consumer RTX 5090 GPUs for under $7,000. The accompanying cost-scaling law and end-to-end data-curriculum study give smaller labs a reproducible reference point that most open-weight releases don't provide.
Puro-2B trains a Qwen2-1.5B-competitive model from scratch on consumer RTX 5090 GPUs for under $7,000, releasing the full data, code, and weights under Apache 2.0.