AI Safety
Overview

This report introduces a retrieval-feedback-driven distillation and preference alignment framework to transfer retrieval-friendly query expansion behavior from a large teacher LLM (DeepSeek-685B) to a compact student model (Qwen3-4B), reducing inference cost while preserving 97% of the teacher's nDCG@10 performance. The approach leverages teacher-generated expansions under zero-shot and few-shot conditions as supervision signals and candidate pools, then applies Direct Preference Optimization to align generation with retrieval objectives. The method demonstrates strong practicality across English and Chinese retrieval benchmarks (TREC DL19/20/21 and MIRACL-zh).

The report introduces Eye2Eye, a framework that enables cognitive alignment between humans and AI in collaborative tasks by sharing a first-person perspective through augmented reality (AR). It addresses inefficiencies in current vision-based assistants by reducing communication and understanding gulfs, demonstrating significant improvements in task completion time, interaction load, and trust in user studies.

This report introduces a risk-aware evaluation framework for assessing LLM security failures in the Banking, Financial Services, and Insurance (BFSI) sector, addressing the gap in domain-specific red-teaming benchmarks. It proposes the Risk-Adjusted Harm Score (RAHS) to quantify operational severity of disclosures and demonstrates how prolonged adversarial interactions escalate harm severity, challenging current single-turn security evaluations.

This report empirically evaluates whether alignment tasks like moral reasoning require diversity-seeking algorithms by comparing distribution-matching and reward-maximizing approaches using RLVR on the MoReBench dataset. The study finds that standard reward-maximizing methods perform equally or better than diversity-preserving techniques, challenging the assumption that alignment inherently demands diverse responses.

This report introduces a cross-view contrastive framework for joint imaging-ROI representation learning in brain disorder classification, addressing the gap between global volumetric imaging and localized ROI-based graph modeling. The method enables systematic evaluation of representation contributions and demonstrates consistent performance improvements over single-branch approaches on ADHD-200 and ABIDE datasets.

This report examines the evaluation of tabular foundation models (e.g., TabPFN, TabICL) for regression tasks, highlighting a critical mismatch between current benchmarks (e.g., TabArena, TALENT) and the models' probabilistic outputs. It argues for adopting proper scoring rules (CRPS, CRLS, Interval Score) to assess full predictive distributions, demonstrating that these rules reveal model limitations obscured by point-estimate metrics and can guide fine-tuning for task-specific objectives.

The Future of Life Institute (FLI) has published the AI Safety Index, a ranking system evaluating frontier AI labs based on their publicly verifiable safety commitments. This report matters because it introduces a transparent, third-party assessment of corporate safety practices, enabling stakeholders to compare and incentivize responsible AI development at scale.