RiskNet: A large-scale dataset of AI risk incidents from news with alignment and multi-dimensional annotations

Paper 1 offers a foundational, large-scale dataset addressing the critical and timely issue of AI safety and governance. While Paper 2 provides a highly valuable technical optimization for LLM inference, datasets like RiskNet typically generate broader, long-lasting cross-disciplinary impact by establishing new benchmarks and enabling extensive downstream research across AI development, policy-making, and societal studies.

RiskNet addresses a critical infrastructure gap in AI governance by providing a large-scale, structured empirical dataset for tracking real-world AI risk incidents. Its breadth of impact spans AI safety, governance, policy, and multiple research communities. The resource nature of the contribution means it enables numerous downstream studies. Paper 2, while technically sound with a novel framework for federated graph learning with missing modalities, addresses a narrower technical problem. RiskNet's timeliness amid growing AI deployment concerns and its potential to inform policy decisions give it broader and higher estimated impact.

Paper 1 exposes a critical gap between theoretical differential privacy guarantees and empirical vulnerabilities in LLM adaptation, specifically regarding distribution shifts between pretraining and fine-tuning data. This provides highly actionable, technical insights for deploying LLMs in sensitive domains. While Paper 2 offers a valuable dataset for AI governance, Paper 1 directly impacts core machine learning methodologies and security practices, likely driving immediate changes in how practitioners and researchers approach privacy-preserving LLM fine-tuning.

Paper 1 introduces a novel architectural contribution (GraMO) that combines state-space models with graph-based learning in a principled way, addressing fundamental challenges in modeling interacting dynamical systems. It demonstrates strong empirical results across multiple benchmarks. While Paper 2 (RiskNet) provides a valuable dataset for AI governance research, dataset papers typically have narrower methodological impact. Paper 1's technical innovation in coupling spatial-temporal dynamics within a single recurrence has broader applicability across physics simulation, robotics, and scientific computing, and advances the methodological frontier more significantly.

Paper 1 introduces a novel, broadly applicable training framework that addresses a key limitation of RL with verifiable rewards (zero group-level advantage) via trace tournaments and efficient Bradley–Terry ranking, yielding demonstrated gains in reasoning benchmarks and compute savings—likely to influence future LLM training methods across domains. Paper 2 provides a valuable dataset for AI governance and risk analysis, but its scientific impact hinges more on adoption/maintenance and may be narrower methodologically. Overall, Paper 1 is more technically innovative, timely for LLM training, and likely to propagate across multiple research areas.

Paper 1 likely has higher scientific impact due to broader cross-field relevance (AI safety, governance, policy, NLP/IR, incident analysis), strong timeliness, and high leverage as a large-scale, reusable dataset/platform that can enable many downstream studies and benchmarks. Its real-world applicability is immediate for monitoring and evaluating AI harms. Paper 2 is innovative and rigorous for aerial manipulation and sim-to-real meta-RL, but its impact is narrower (robotics/UAVs) and depends more on adoption and reproducibility in specific hardware settings.

Paper 2 offers higher scientific impact due to its novel contribution to mechanistic interpretability of LLMs, specifically causally localizing temporal preference representations—a previously unexplored area. It combines multiple rigorous methods (gradient attribution, activation patching, steering vectors) to provide actionable insights for AI alignment and control. The finding that LLMs discount the future differently than humans has broad implications for AI safety and deployment in decision-making. Paper 1, while useful as a dataset resource, is primarily an infrastructure contribution with incremental novelty over existing incident databases, and its impact depends on downstream adoption.

Paper 1 introduces a fundamental algorithmic breakthrough in reinforcement learning, unifying model-free and model-based methods. Its rigorous theoretical proofs, error bounds, and extensive empirical validation across 80 diverse environments demonstrate exceptional methodological rigor. While Paper 2 provides a timely and valuable dataset for AI governance, Paper 1's potential to advance core AI capabilities and inspire follow-up algorithmic research gives it a higher estimated scientific impact.

Paper 2 has higher potential scientific impact due to a novel methodological contribution (amortizing entropy-search acquisition via PFNs/in-context learning) that directly improves a widely used core tool (Bayesian optimization) with large practical gains (50x speedups) and open-source code. Its approach is broadly applicable across ML, AutoML, experimental design, robotics, and hyperparameter tuning, making cross-field uptake likely. Paper 1 provides a valuable dataset/platform for AI risk analysis, but its impact depends on sustained curation, coverage, and adoption, and is more domain-specific and sensitive to data/annotation biases.

GraphDETR introduces a novel deep learning framework that reformulates a fundamental NP-complete graph theory problem (subgraph isomorphism) as a set prediction task, drawing an innovative analogy to object detection (DETR). It offers both exact and approximate matching capabilities with strong empirical results on molecular functional group detection. This methodological innovation has broad applicability across scientific domains (chemistry, biology, network analysis). While RiskNet is a valuable dataset contribution for AI governance, it is primarily a curation/annotation effort with narrower scope; GraphDETR's algorithmic novelty and cross-domain applicability suggest higher long-term scientific impact.