The paper proposes a unified Dirichlet framework for spatial-temporal risk assessment, proving that a single Dirichlet posterior per cell with an additive evidence-update rule is the unique update–predictor pair satisfying four axioms and is limit-equivalent to seven classical methods (AHP, Dempster–Shafer, Hawkes, kernel density estimation, etc.). The framework simultaneously yields a severity score and threat characterization from the posterior. On a large-scale benchmark of 41 regions × 10,000 cells × 365 days, it achieves an one-vs-rest AUROC of 0.666 and severity AUROC of 0.725, statistically significantly outperforming 15 structured baselines (Holm-corrected p < 10⁻²⁶), while delivering threat characterization accuracy of 79.1%—compared to only 0–26% for competitors with comparable AUROC. Real-world transfer to 1.69M London and 119K Chicago crime events preserves the dual-output advantage, and a pre-registered specialization experiment confirms the operational configuration beats the matched specialist. The method requires 3.6× less memory than seven independent models (128 vs. 464 bytes/cell) at 41K signals/sec throughput.
This paper identifies that object hallucinations in large vision-language models (LVLMs) originate from visual encoders, uncovering three core issues: statistical bias, inherent bias, and vulnerability. To address these, SHIELD is introduced as a training-free framework that applies three strategies: re-weighting visual tokens to reduce statistical bias, injecting noise-derived tokens to counteract inherent bias, and employing adversarial attacks with contrastive decoding to mitigate vulnerability. Experiments across multiple benchmarks and LVLM families demonstrate SHIELD effectively reduces object hallucinations while maintaining strong general performance, and the code is publicly available.
AI skin cancer triage systems generate about 106 excess unnecessary referrals per 1,000 darker-skin patients due to over-prediction, not missed cancers. This over-prediction stems from melanin concentration reducing lesion-background optical contrast, causing class overlap. The authors formalize this as a signal-to-noise ratio (SNR) framework, predicting a 5.2× SNR reduction from lighter to darker skin tones. Experiments on the HAM10000 dataset with a high-confidence ITA subset show dark skin achieves slightly higher sensitivity (0.848 vs. 0.821) but substantially lower specificity (0.720 vs. 0.831, Δ=−11.1pp). An ablation study compares ITA-based tone conditioning (feature calibration) and dark-skin augmentation (decision boundary adjustment), revealing their distinct effects. Zero-shot transfer to the DDI dataset (n=656) confirms the AUC gap. Code and trained weights are publicly released.
IncidentMind is a token-budget multi-agent system for autonomous root cause analysis of production AI failures. It pre-syncs Slack, Confluence, and Jira into a HydraDB temporal knowledge graph via MCP, converting all agent queries into a single graph traversal. A tri-tier inference strategy uses minilm-l6 for sync-time tasks, quantized Llama-3-14B for agent reasoning, and GPT-4o-mini only when confidence falls below 85%, reducing per-incident cost from $1.50 to $0.003. Structured token budgeting compresses 50,000 raw log tokens to 1,050 tokens (98% reduction). Across 847 production incidents, IncidentMind achieved 91% fix accuracy and reduced mean time to detect from 4.2 hours to 3 minutes.
The paper introduces Self-Aligned Reward (SAR), a fine-grained RL signal that complements verifiable rewards to improve both accuracy and efficiency of LLM reasoning. SAR is defined as the relative perplexity difference between a query-conditioned answer and the standalone answer, thereby favoring concise, query-specific responses and penalizing redundancy. Quantitative analysis confirms that SAR reliably ranks answer quality, assigning higher scores to concise correct answers than to verbose ones. Integrating SAR with PPO or GRPO reduces average answer length by 30% while boosting accuracy by 4% across four model families and seven benchmarks, with strong out-of-domain generalization. The approach achieves a Pareto-optimal frontier between correctness and efficiency, shortening unnecessary elaboration without hurting advanced reasoning behaviors. Code and data are publicly released.
Autoregressive language model inference is not fully determined by fixed weights; instability phenomena like drift and hallucination arise from structural trajectory dynamics. Causal isolation experiments using gradient scrambling demonstrate that trajectory geometry constitutes a control field, and state-dependent feedback (e.g., switching between two frozen models without parameter updates) is both necessary and sufficient for stability. Fixed-setpoint control fails due to control friction, while the proposed boundary-aware Dynamic Operator Mixing (Band DOM) achieves stability with approximately 79% of inference steps requiring zero control input. A fundamental limit is identified: dynamic stability and semantic consistency are decoupled; stabilized trajectories exhibit mode-switching in over 85% of trials while maintaining geometric smoothness, revealing a kinetic/potential decomposition of inference dynamics.