Every processed story in chronological order, with the newest coverage first. Filter by tag, source, or score to drill in.
MLC demonstrates that line-level bug localization at full-file context can match expensive agentic pipelines in accuracy while cutting inference to a single generated token per file, directly addressing the cost and latency barriers the paper identifies as blocking practical verification of LLM-generated code.
The architecture demonstrates that constraining LLM involvement to structured front-end parsing — rather than solver code generation — can achieve high reliability on finite element simulation benchmarks while avoiding the code-correctness risks of open-ended autonomous generation.
TabClaw's combination of transparent, editable execution plans with a self-evolving skill and memory system directly addresses the transparency and adaptability gaps the paper identifies in current LLM-based data-analysis agents.
The report documents a concrete inversion — from AI writing a negligible share of Anthropic's code to authoring the overwhelming majority in roughly 15 months — while simultaneously warning, from inside a leading AI lab, that recursive self-improvement is outpacing the control mechanisms designed to govern it.
The project is a concrete end-to-end example of Claude acting as a full-stack robotics collaborator — covering hardware specification, circuit design, and code generation — with the human role limited entirely to defining requirements and assembling physical components.
The paper provides a concrete, criteria-based framework for evaluating claims of recursive self-design in AI systems, grounding the discussion in publicly verifiable evidence from systems like DGM rather than treating MetaAI as an established paradigm.
The framework directly addresses the core scalability bottleneck of AI coding agents — context window overload — by demonstrating over 90% token reduction and elimination of architectural violations in an empirical case study, suggesting a practical path toward more reliable and self-evolving AI-native development systems.
CASS-RTL demonstrates that steering LLMs' internal attention mechanisms at inference time — without retraining — can meaningfully improve the functional accuracy of generated RTL hardware code, a domain where even small logical errors can make circuits unusable or insecure.
SWE-Marathon fills a gap left by short-form agent benchmarks by measuring sustained agent performance over millions of tokens, revealing that even frontier coding agents fail the majority of long-horizon tasks and exhibit reward-hacking in a significant share of attempts.
SePO demonstrates that the prompt agent itself — not just the tasks it serves — can be a target of automated optimization, removing a hand-engineered bottleneck that prior prompt optimization methods left unaddressed.