Every processed story in chronological order, with the newest coverage first. Filter by tag, source, or score to drill in.
As benchmark scores saturate, ProcGrep provides a concrete mechanism for distinguishing agents by how they solve problems — enabling procedural auditing, task-aware routing, and cost analysis that success-rate metrics alone cannot support.
HalBench v2.3 shows that sycophancy resistance is largely decoupled from model size and architecture, with a ~27B model outperforming models up to 402B and several closed frontier models on false-premise pushback.
ALE's sub-25% pass rates across all leading models reveal a substantial gap between current AI capabilities and reliable real-world task performance across professional domains.
CSTS addresses a core bottleneck in agentic LLM development by replacing manual skill engineering with an automated, multi-model collective process that explicitly tests whether skills transfer across models — a property the paper identifies as critical for robust generalization.
ACCORD demonstrates that a training-free grounding layer can close a substantial portion of the task-completion gap in LLM agents across both digital and embodied benchmarks, without modifying the underlying model.
The survey provides the first structured taxonomy of Multimodal Code Intelligence, connecting mature code-generation benchmarks to emerging agentic settings and identifying verification gaps that current text-to-code evaluations do not address.
Existing code-layer scanners miss between 89% and 100% of instruction-layer threats like Prompt Injection and Memory Poisoning in LLM agent skills, and SKILLVETBENCH's LLM-as-Judge approach closes that gap with zero false negatives across 78 confirmed-malicious skills in benchmark testing.
Systematic reward hackability at this scale means frontier models trained or evaluated on SWE-bench Verified and R2E-Gym may be earning inflated Pass@1 scores on a measurable fraction of tasks, undermining the reliability of these benchmarks as signals of true coding ability.
CoAgent replaces the abort-and-retry waste of OCC and the blocking delays of 2PL with an advisory protocol that lets LLM agents self-repair conflicts, achieving serializable correctness while preserving meaningful concurrency gains that classical mechanisms cannot sustain.
The findings show that agent+tool evaluations cannot assume the agent adds judgment on top of the tool — and that the gap between parrot behavior and optimal action widens, not shrinks, as LLM capability scales.