LeafSkill: Efficient Leaf-Level Action Grounding for Recursive LLM Agents
LeafSkill optimizes recursive LLM agents by caching leaf-level action grounding templates, reducing model re-invocations across tasks.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
LeafSkill optimizes recursive LLM agents by caching leaf-level action grounding templates, reducing model re-invocations across tasks.
Discord's latest update released today has a big improvement for Markdown previews: they should now render correctly instead of with "a bunch of symbols," according to the company. LLMs and AI agents often rely on Markdown files for instructions and context, so these improved…
Evaluates cost-aware abstention policies for LLM-based IT ticket classification on 47,837-ticket dataset with asymmetric error costs.
Introduces EpiSEM-Bench, executable benchmark with 300 task pairs measuring unsafe LLM agent execution when safety-critical facts are missing.
LLM agents can audit training corpora for synthetic contamination; found 5.32% machine-generated documents in Slovak FineWeb2 for $2-3 per run.
Simulated Control Evals (SCEs) expand LLM safety testing by using model-generated tool responses to probe behavior beyond fixed test environments.
VERA dashboard operationalizes EU AI Act compliance for LLM deployment, mapping 18 technical requirements to measurable benchmarks for non-technical stakeholders.
Evaluation of four open LLMs (8–27B parameters) on Spanish MIR medical exam assesses uncertainty calibration, reasoning, and multimodal grounding beyond raw accuracy.
Empirical study shows users defer to confident but incorrect LLMs, revealing trust calibration collapse independent of model sycophancy.
TTR-Compressor uses test-time regression for LLM sequence compression, learning end-to-end via backprop with closed-form coefficient computation at inference.
Social-HUMINT quantifies social-media exposure via LLM-inferred user profiles; measures fidelity and proximity to social engineering post-API landscape shift.
Diffusion language models generate natural adversarial prompts for black-box LLM attacks, improving prompt injection techniques.
Framework audits LLM-generated metadata governance in enterprises by tracking human stewardship decisions to prevent error propagation.
GPU applications are processing ever-growing datasets with increasingly powerful accelerators. According to the 2025 Stanford AI Index, compute for notable AI... GPU applications are processing ever-growing datasets with increasingly powerful accelerators. According to the 2025…
Empirical benchmark of recent LLMs vs. fine-tuned models on knowledge-graph-to-text generation, narrowing performance gaps.
Mechanistic interpretability study identifies causal subspaces controlling metaphor representation in LLMs.
Production case study: structured retrieval context outperforms prompt-based rules for LLM pharmaceutical claims compliance.
Small open-weight LLM trained via RL to select tools from 10k+ corpora, matching proprietary model performance on ToolBench.
Empirical study of LLM inference optimizations (KV-cache pruning, speculative decoding, quantization) on Apple Silicon vs NVIDIA GPUs.
Framework translates structured clinical data into patient-readable LLM health narratives while preserving clinical semantics.
DIAL enables fine-grained numerical control over instruction-adherence in LLMs via distillation, scaling efficiently across diverse instructions.
Study reveals field-level grounding challenges in LLM-generated museum descriptions when metadata composition varies.
Benchmark of 9 prompting strategies across 4 LLM judges in 27 languages shows LLM-as-Judge robustness varies significantly by model and language.
Network topology significantly shapes consensus formation speed and stability in multi-agent LLM populations on naming game tasks.
Study reveals prompt sensitivity in tool-augmented LLM comparisons; proposes matched protocol to isolate reasoning, tool-calling, and readout.
Forward-only training method for continual n-gram memory updates in LLMs without backprop or weight modification.
LLMs augment constraint reasoning to generate contrastive explanations for MIP optimization decisions in decision-support systems.
121k ground-reason dataset from French Court of Cassation (2016–2025) for legal domain LLM analysis and benchmarking.
Multi-agent LLM study shows larger models absorb opinions of smaller models in consensus tasks regardless of task structure.
SCOPE compiles natural-language policies into neurosymbolic programs for stable LLM agent governance over multi-turn trajectories.