Live intelligence

Latest AI updates

8 updates

Research

Harness Design Alters Fixed LLM Compliance With Banking Conduct Rules

A new arXiv study holds four open-weight models fixed while running each through five agent harnesses on 156 pre-registered simulated retail-banking cases, scoring 3,120 two-turn episodes from executed tool traces. Pooled correct-case rates ranged only from 87.3% to 92.1%, but within one model the best-to-worst harness spread was 11.5 percentage points, enough to change its rank from first to last. A same-model compliance reviewer cut violations from 9.6% to 6.4% without increasing over-refusals, while rules requiring date arithmetic failed in every harness. In an exploratory analysis, 26% of violations included a customer message denying the action the agent had just executed, which transcript-only review would miss.

arxiv11h agoScore 74

AI governanceLLM agentsfinancial services compliance

Research

The Split Protocol Decides the Result: Re-Evaluating Encrypted Traffic Classification

This arXiv paper shows that encrypted traffic classifiers can score macro-F1 0.992 under a standard evaluation protocol but only 0.567 when flows from one packet capture are kept on one side of the split. Across five split protocols on backbone QUIC traffic, server-disjoint splitting costs 0.192 macro-F1, while random, temporal and client-disjoint splits are statistically indistinguishable from session-disjoint splitting. The authors propose single-identifier probes as a pre-training diagnostic for identifying the relevant split axis, and report that several pre-registered architectural claims fail under the corrected evaluation.

arxiv11h agoScore 76

encrypted traffic classificationevaluation methodologydata leakage

Research Editor selected

Trick-or-Treat: Adversarial Multimodal Medical Benchmark Exposes Hallucinations in Frontier Models

Researchers introduced Trick-or-Treat, an adversarial multimodal medical benchmark pairing real chest radiographs with synthetic clinical photographs to test cross-modal reasoning. In mismatched pairs, frontier models hallucinated clinically significant findings in 45-57.5% of cases, up from 25-37.5% when photographs were shown alone. The study demonstrates a concerning error pattern and validates synthetic images for adversarial evaluation.

Why it matters: This paper provides concrete evidence of hallucination in medical AI and introduces a novel evaluation method with synthetic images, offering immediate value for AI safety and medical reasoning research.

arxiv11h agoScore 85

AI safetymedical AImultimodal reasoning

Research Editor selected

When Ground Truth Is Not Ground Truth: Automatic Labeling for Hallucination Detection

This paper investigates the reliability of automatic labeling methods for hallucination detection in LLMs, showing that using similarity metrics like BLEURT as proxies for ground truth can be highly unreliable. Through experiments with 15 judges, 7 detectors, 2 generators, 2 QA datasets, and 5 languages, they find that detector performance varies significantly with judge choice. Human validation shows QGen-9B achieves the lowest misclassification rate (12.86%) while BLEURT and BLEURT-n exceed 47%, highlighting the need for more factuality-aware judging methods.

Why it matters: Provides concrete evidence and benchmarks on the pitfalls of common automatic labeling practices, offering actionable insights for improving hallucination detection evaluation.

arxiv1d agoScore 82

LLM hallucinationevaluation methodologyautomatic supervision

Research Editor selected

False Reassurance in LLM Severe-Weather Guidance: A Replayable Evaluation Against Real NWS Alerts

This paper introduces 'false reassurance'—an LLM assistant telling users conditions are calm while a severe-weather alert is active—and measures it by replaying 117 frozen cases built from real NWS alerts through a Claude-based pipeline. The production pipeline achieved 0% false reassurance, while blind prompts caused 37-45% false reassurance, and real alerts significantly reduced it. The study also highlights failures like ingesting NWS test products as real alerts and token budget issues, and releases dataset, labels, and code.

Why it matters: Provides concrete, reproducible evaluation of a critical safety failure mode in LLM weather assistants, with actionable insights for system design.

arxiv1d agoScore 82

LLM safetyweather guidancealert integration

Research Editor selected

Grading the Unannotated: WorkRB Skill Extraction Challenge at RecSys-HR 2026

The WorkRB Challenge at RecSys-HR 2026 replaces binary skill extraction evaluation with graded relevance (0-4) using nDCG@100, addressing the limitations of binary labels at ESCO scale. The paper details a cascade of four binary questions answered by a smaller LLM judge, validated via blind human and frontier-model re-annotation, and presents five benchmark datasets. Submissions converged on multi-stage retrieve-and-rerank pipelines, with gains from filtering nonsense rather than finding more positives, and graded evaluation reveals larger gaps between top systems and baselines than binary evaluation.

Why it matters: Provides concrete methodology and benchmark insights for graded skill extraction, highlighting a shift in evaluation that exposes model limitations and open challenges.

arxiv2d agoScore 85

skill extractionevaluation methodologyLLM-based annotation

Research

Implicit Assumptions About Tasks in Continual Learning

This paper clarifies that continual learning measurements depend on factors like data similarity, change rate, and order, which can be defined at different levels: underlying data, partition, or learner. The authors show these levels often don't coincide and are usually left implicit, leading to ambiguous results. Through teacher-student experiments, they demonstrate that changing granularity, order, or similarity structure alters measured outcomes. They also review benchmarks to see which properties are stated, noting that for foundation models with undisclosed data, properties are inherited by the model's initial state.

arxiv3d agoScore 72

continual learningevaluation methodologytask design

Research

ReuScope: Auditing Data Reuse and Evaluation Sensitivity in Capsule Vision 2024

ReuScope is a protocol for auditing data reuse and evaluation sensitivity, applied to the Capsule Vision 2024 challenge. It reveals byte-identical files crossing the public training-validation boundary and shared source recordings between partitions. Removing Kvasir-Capsule training rows drops public-validation balanced accuracy from 82.5% to 61.2%, but the effect reverses on other sources, indicating sensitivity rather than leakage inflation. The protocol distinguishes observed reuse from conditional performance sensitivity with matched controls and uncertainty intervals.

arxiv3d agoScore 72

data reuseevaluation methodologymedical imaging