CommunityMar 11, 20262026-03-114 minVerified
Self-Amplifying R&D Loops And Alignment-Faking Risk Signals
As AI-driven R&D loops accelerate, alignment-faking signals (12%) raise operational risk. Lock in TEVV, independent review, and monitoring.
SourceMar 11, 20262026-03-114 minVerified
Testing Intersectional SDoH Bias in Clinical LLM Recommendations
Clinical LLM recommendations can shift with intersecting SDoH (gender, insurance, housing). Test cross-profiles and measure over-refusal before deployment.
SourceMar 10, 20262026-03-104 minVerified
Bridging Pathology AI Benchmarks and Real-World Clinical Deployment
Why pathology AI lags after strong benchmarks: external validation, drift/OOD monitoring, workflow fit, and auditable logging.
CommunityMar 10, 20262026-03-104 minVerified
Distinguishing Logprobs From Self-Reported Confidence in Prompts
Explains why token logprobs differ from natural-language confidence, and how to test multi-candidate prompts with seeds and evals.
SourceMar 10, 20262026-03-104 minVerified
Grounding Self-Driving Explanations With Retrieval-Augmented Demonstrations
RAG-Driver grounds driving explanations with retrieved expert demonstrations via RA-ICL, but evaluation still relies on BLEU, METEOR, and CIDEr.
CommunityMar 10, 20262026-03-104 minVerified
Designing Reproducible Rubrics for LLM Code Integration Evaluations
Move beyond context/output limits: evaluate LLM code integration with task decomposition, tool parity, and reproducible build/test rubrics.
SourceMar 10, 20262026-03-104 minVerified
RM-R1 Reward Models That Reason Before Scoring
RM-R1 proposes reward models that reason before scoring, reporting up to 4.9% gains on public RM benchmarks and highlighting safety evaluation gaps.
CommunityMar 10, 20262026-03-105 minVerified
Three Policy Checks Before Choosing an AI API
How auth (OAuth/OIDC vs API keys), rate/spend limits, and tiered model access policies shape SaaS cost, security, and reliability.
TrustedMar 10, 20262026-03-104 minVerified
Ulysses Sequence Parallelism for Long Context Training Efficiency
Ulysses splits sequences across GPUs and exchanges K/V via all-to-all to reduce long-context attention bottlenecks and track throughput.
CommunityMar 9, 20262026-03-094 minVerified
Copilot Cowork Shifts AI From Prompts To Workflows
Microsoft introduces Copilot Cowork as a research preview, focusing on long-running, multi-step work and human-in-the-loop execution.
SourceMar 9, 20262026-03-094 minVerified
Predictive Coding Networks: Fixed Points, Inference, Backprop Links
Overview of PCN: iterative inference, fixed-point convergence (dv≈0), links to backprop equivalence/approximation, and compute bottlenecks.
SourceMar 9, 20262026-03-093 minVerified
Prompt Group Training For Robust Text Guided Segmentation
Summarizes prompt group-aware training that aligns predictions across equivalent prompts, reducing variance and improving average zero-shot Dice.
CommunityMar 8, 20262026-03-084 minVerified
Beyond Benchmark Scores: Reproducible, Multi-Metric Model Evaluation
Why tiny benchmark gaps mislead: evaluation settings, reproducible logs, and multi-metric, roadmap-driven model selection.
CommunityMar 8, 20262026-03-085 minVerified
Combining LLM Agents With Science Models for Reliable Loops
A practical pattern: LLMs handle planning and interpretation, while science models provide constraint-based scoring and stopping gates.
CommunityMar 8, 20262026-03-084 minVerified
When 4-Bit Quantization Beats FP16 Perplexity
Explain why 4-bit quantized models can show lower PPL than FP16, and outline a reproducible evaluation protocol.
CommunityMar 8, 20262026-03-084 minVerified
Why AI Talk Runs Long While Drinking
How acute alcohol use can weaken response inhibition and make AI talk too long, plus simple rules to keep rapport in social settings.
CommunityMar 8, 20262026-03-085 minVerified
Why Custom Instructions Personas Drift Under Hierarchy
Model Spec’s chain of command can override custom instructions, causing persona and reasoning drift. Design priorities, exceptions, and fallbacks to improve reproducibility.
CommunityMar 7, 20262026-03-074 minVerified
CAPTCHA Security Friction Tradeoffs In Real User Flows
Real-user data shows CAPTCHA time varies by context, while ML and relay attacks raise friction without guaranteed security gains.
SourceMar 7, 20262026-03-073 minVerified
Evaluating Zero-Shot MLLMs for Reliable Video Anomaly Alerts
Assesses zero-shot MLLMs for video anomaly detection, focusing on false alarms/misses, prompt specificity, 1–3s clips, and PR/F1 evaluation.
SourceMar 7, 20262026-03-074 minVerified
Gating Robot Autonomy Using Deep Perception Uncertainty Signals
SPIRIT uses deep perception uncertainty to gate shared autonomy, switching between semi-autonomous manipulation and haptic teleoperation.
CommunityMar 7, 20262026-03-074 minVerified
Designing Language for LLM Expectations and Verification
How to reduce anthropomorphism, overconfidence, and hallucinations by structuring work as claim-evidence-verification checklists.
SourceMar 7, 20262026-03-074 minVerified
LegalBench And Auditable Argumentation For Legal LLMs
How LegalBench evaluates legal LLM reasoning beyond accuracy, emphasizing justification and auditability through structured argumentation and governance.
SourceMar 7, 20262026-03-074 minVerified
Logi-PAR Adds Differentiable Rules to Clinical Activity Recognition
Logi-PAR (arXiv:2603.05184v1) integrates neural-guided differentiable rules into clinical PAR, enabling rule traces and counterfactual interventions.
SourceMar 7, 20262026-03-074 minVerified
Memory Admission Control for Reliable LLM Agents
A practical look at memory admission control for LLM agents, reducing long-term memory pollution while improving auditability and metrics.