This article presents a structured framework of behavioral indicators that may signal progression toward potentially catastrophic threats from artificial intelligence systems. We adopt a pragmatic approach, inspired by established methodologies in cybersecurity and national security. By establishing clear metrics, indicators, and thresholds across multiple dimensions of AI capability and behavior, this framework enables researchers and policymakers to implement evidence-based monitoring protocols.
T. Bauer, W. P. Kegelmeyer, E. Begoli et al.· 0 citations
Libraries and archives manage large collections with limited staff and computing budgets, yet common benchmarks do not systematically test their bibliographic work. They need to know which methods work for their tasks and what those methods require to run. SHELF, the Synthetic Harness for Evaluating LLM Fitness, addresses this gap. It is a Python system that turns labelled taxonomies, writing specifications, and a generation budget into controlled benchmark data and evaluation tasks. This first release contains 62,899 model-written documents based on Library of Congress vocabularies, with tasks for classification, clustering, retrieval, pair classification, and instruction retrieval. We compare TF, TF-IDF, BM25, popular encoders, and, on subject classification only, zero-shot decoders; each method appears only on tasks that support it. Subject classification reaches 0.8887, while genre-form classification reaches only 0.2605, and several pair and clustering tasks remain near chance. Sparse methods remain competitive on classification, while TF-IDF is the fastest measured arm in the subject timing experiment. SHELF also varies bibliographic facets independently and can generate new, verifiably unseen documents after a model's training cutoff. Comparisons with LCSHBench and Project Gutenberg show that model rankings transfer more reliably than absolute scores, but SHELF scores do not estimate accuracy on production catalogue data. We release all source code and data under permissive licenses on GitHub and Hugging Face.
Mechanistic interpretability is increasingly used to guide interventions such as activation steering, circuit removal, and safety monitoring. Yet an internal estimate that is accurate on average can still choose a poor action.
We present ObserverBench, a benchmark framework for testing whether an internal estimator---an observer---is adequate for the intervention, control, or safety task it directs. Each task fixes the model, information boundary, allowed actions, decision rule, held-out cases, and loss. The benchmark reports estimation accuracy separately from the loss caused by the chosen action.
Theory and experiments show why both are needed. In closed-loop control, observer errors matter at the starting point and along directions the allowed intervention can reach. On circuit-intervention tasks in GPT-2-small and Qwen2.5-7B, pairwise observers predict unseen effects more accurately without always choosing better actions; observers trained on action loss choose lower-loss actions. In safety triage, a score that perfectly separates violations can allocate a fixed intervention budget poorly when violations have different costs. Across Qwen2.5-7B, Gemma-2-9B-it, and prospectively frozen Qwen3.5-9B APPS tasks, AUROC can rank monitors differently from deployment loss, and the best information source changes across models. Sparse SAE readouts also trail their layer-matched dense controls on the reported Qwen panels, under disclosed activation-density or checkpoint mismatches.
ObserverBench provides fixed task contracts, runnable baselines, and table-based submissions for evaluating interpretability methods through the actions they enable.
On-policy distillation (OPD) accelerates post-training by providing dense token-level supervision from a frozen teacher on the student's own rollouts. Vanilla OPD applies this supervision uniformly across prompts, without checking whether the teacher is reliable for each prompt. Because reverse KL is mode-seeking, a confidently wrong teacher can induce a strong yet misleading update. Distributional proxies, such as entropy or teacher-student likelihood agreement, measure uncertainty or agreement but do not directly verify outcome correctness. We introduce Teacher-Gated On-Policy Distillation (TGOPD), built on the principle that teacher reliability should be verified at the prompt level before dense supervision is admitted. TGOPD estimates reliability from a small set of verifier-scored teacher probes and routes each prompt exclusively to dense OPD when the reliability check passes or to verifier-grounded GRPO otherwise. Across 4B and 35B students in mathematics, code, and instruction following, TGOPD outperforms Vanilla OPD in all six single-domain settings and achieves higher seven-benchmark averages at both scales under multi-domain training. By using otherwise-idle teacher capacity for reliability estimation, TGOPD also reduces teacher-side compute waste in asynchronous OPD, increasing teacher-node GPU utilization from 9.8% to 78.9% in the measured 4B single-domain run.
Zhiwei Zhang, Zechen Sun, Fei Zhao et al.· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
Graph neural networks (GNNs) are a class of neural networks suitable for learning on graph-structured data. Their application to spatial data is a natural extension, however its relatively unclear which message-passing operations, architectural configurations, and graph representation is best suited for classifying changes to objects in electronic navigational charts (ENCs)--geospatial vector datasets used for marine navigation. Maintaining these datasets is a challenge, and categorizing changes to objects in the ENC based on their significance to navigational safety is of particular importance. Here, we propose to represent these vector navigation datasets as a graph structure where the spatial objects serve as nodes and their spatial and semantic relationships form edges. We encode both the old ENC dataset and new ENC dataset into a pair of graphs and frame the task as a graph-pair classification problem. Building on this representation, we investigate the use of GNN architectures to classify whether the encoded graphs constitutes a critical or non-critical risk to navigational safety. We train and evaluate several GNN architectures and model configurations on ENC changes reviewed by maritime experts. Our results demonstrate that graph-based representations improve the classification of ENC updates, providing a scalable approach for automating or improving ENC maintenance workflows.
To scale up collective decision-making, participatory democracy platforms such as Polis and Remesh enable online deliberation among thousands of participants. However, at this scale, participants cannot review every opinion submitted by others, producing highly sparse voting data that misrepresent patterns of consensus, conflict, and minority support. Platforms therefore increasingly rely on Preference Inference (PI) models to predict missing votes. Yet this automation is not neutral: inferred preferences can artificially amplify, suppress, or reorder existing patterns of support, ultimately reshaping how the outcomes of a deliberation are interpreted. More generally, we lack a systematic understanding of how existing PI methods affect the collective preference landscape. To address this gap, we benchmark several existing PI approaches in this context. Moving beyond conventional user-centric evaluations centered on the accuracy of individual predictions, we introduce a collective-centric evaluation framework that measures whether inferred votes preserve salient properties of the broader preference landscape. We further contribute the largest multilingual dataset of its kind: four consultations spanning over 90k participants, 1M votes, and 22 languages. Our experiments show that models with comparable predictive accuracy can differ substantially in the degree to which they preserve the collective structure. These results demonstrate that accuracy alone is insufficient for evaluating \emph{PI} in democratic settings. By contributing a novel comprehensive and collective-centric evaluation benchmark for the task of PI, this work aims to support the development of AI systems that scale deliberation without compromising the integrity of its democratic outcomes.
Pierre-Antoine Lequeu, Salim Hafid, Paul Lerner et al.· 0 citations
Topology-guided safeguards for LLM-based multi-agent systems (MAS) train a GNN over the inter-agent communication graph to localize risky agents and intervene on the topology---but they assume one operator can pool all labeled traces. Across organizations that assumption breaks: episodes contain private prompts, tool outputs, and proprietary workflows, and no silo alone sees the full attack distribution. We cast privacy-preserving MAS safeguarding as graph federated learning and instantiate FGLGuard: each operator fits an edge-featured graph attention detector on its own judge-labeled episode graphs and shares only model updates. The method couples a proximal local objective for non-IID clients, domain-balanced aggregation, over-refusal-constrained threshold calibration, corroborated upstream scoring, and a guarded rewrite for blocked answers. Federation is not optional: off-the-shelf transfer collapses under distribution shift (AUROC 0.51 to 0.70 only after in-domain retraining), so a deployable guard must adapt on each site's private traces. On Agent-SafetyBench, R-Judge, and AgentDojo, federated FGLGuard exceeds the in-domain centralized ceiling on all three benchmarks without pooling any data---where unsupervised anomaly guards and local-only training fail. One guard federated across four different-domain operators comes within 0.03 AUROC of multi-domain centralization, while any single-domain guard collapses on the others. Live FGLGuard cuts AgentDojo's ground-truth attack-success rate by 43% at near-unguarded utility, zero API cost, and negligible capability loss.
Jinxi Yu, Eric Hanchen Jiang, Levina Li et al.· 0 citations
This paper focuses on defending generative search engines against malicious Generative Engine Optimization (GEO), which rewrites web documents to match engines' citation preferences and thereby manipulates generated answers. Recent GEO methods have advanced from hand-crafted rewriting to automated and agentic optimization, substantially increasing the visibility of target documents in generated answers. However, defending against such manipulation poses two major challenges: attack documents remain factually consistent with their originals, rendering fact verification and perplexity filtering ineffective, and the features they amplify equally characterize high-quality benign content. To address these limitations, we propose GEO Defender, a two-stage defense aligned with the attack chain that requires no fine-tuning of the target LLM. GEO Defender consists of Shield Reranker and Training-Free Shield Generation (TFSG). Specifically, Shield Reranker learns a preference-based defensive residual over a frozen base reranker, demoting GEO-rewritten documents while preserving relevance judgments, and TFSG distills defense outcomes into a natural-language experience library that guides the target LLM's source use at inference. Experiments on two state-of-the-art closed-source LLMs and three open-source LLMs across seven GEO attacks demonstrate that GEO Defender reduces the average attack success rate from 50.32% to 6.20%, retains 94.12% of benign-evidence use, preserves answer quality, and generalizes to unseen attacks from construction instances.
Haozhang Li, Yangguang Shao, Xinjie Lin et al.· 0 citations
What does a language model predict when it has few clues? The answer lurks in its unembedding geometry: a single direction of the unembedding matrix encodes the unigram distribution of the training corpus, which serves as the Bayesian prior the model falls back on when uncertain. This structure --- which we term the \emph{direction of ignorance} --- appears in all four model families examined (\texttt{Llama}, \texttt{Qwen}, \texttt{Gemma}, and \texttt{Pythia}), ranging from 0.4B to 405B parameters. Projecting the final prediction state onto this direction yields a per-token \emph{prior loading factor} $\lambda$, which, empirically, declines steadily as the context becomes more informative. Formally, the same projection decomposes the prediction state into two orthogonal vectors that correspond exactly to the two factors of a tempered Bayesian update: a unigram prior raised to the exponent $\lambda$ and a context-driven likelihood. This geometric-probabilistic interpretation calibrates $\lambda$, making it meaningfully comparable across model sizes and families, with larger models generally exhibiting lower prior reliance in the high-context limit. Finally, we show that the direction of ignorance is causally active: raising or lowering $\lambda$ at the final prediction state steers the prediction toward or away from the unigram prior in KL divergence.
Toni J. B. Liu, Jiajun Bao, Yizhou Liu et al.· 0 citations
Sensor-intensive environments enable many intelligent services by inferring user applications from heterogeneous data streams. However, not all applications should be exposed: users want some activities to stay private. This creates a tension between inferring applications for useful services and preventing unwanted inference. Existing approaches such as differential privacy and rule-based filtering protect individual streams but cannot address the privacy risk from cross-sensor inference.
We introduce Privatehub, which uses contrastive learning within a diffusion model to generate synthetic multi-sensor streams that keep non-private applications detectable while concealing private ones. Privatehub has two stages: App-Conditioned Pre-training (ACP), which conditions the model on multi-sensor data with application embeddings, and App-Aware Fine-tuning (AAF), which separates private from non-private data via contrastive learning. We also define a threat model for the multi-sensor sharing setting. Experiments on three real-world multi-sensor datasets show Privatehub lowers private-application accuracy by 40 to 50\% without hurting non-private performance, and stays robust when the attacker retrains on the synthetic data.
Jiechao Gao, Yuandong Pan, Jie Wang et al.· 0 citations
Significant challenges remain in AI-driven educational systems in balancing privacy preservation with accurate cognitive diagnosis. To overcome this, we propose a federated inference framework in which several commercial LLM APIs collaborate without requiring access to raw student data or proprietary model internals. Using multiple federated entities, such as LLaMA-3.3-70B, GPT-4o-mini, and Claude-3-Haiku, our framework builds upon a heterogeneous multi-LLM architecture. The predictions generated by these entities are combined with epsilon-local differential privacy by adding Laplace noise locally to each entity's prediction output before aggregation, while residual-based aggregation mitigates model heterogeneity. Our approach is predicated on an honest-but-curious trust paradigm in which API providers are presumed not to abuse submitted queries, and our differential privacy mechanism shields the published diagnostic results from external inference. We conduct rigorous privacy-utility analysis showing strong privacy guarantees with minimal accuracy loss, and extensive real-world evaluations across three educational benchmarks confirm the framework's practical usability and cross-domain generalizability.
Democratizing data access through natural language is a crucial goal for modern enterprises, but the practical adoption of Text-to-SQL is critically hindered by real-world complexities: 1. Obscure and large database schemas, 2. Ineffective retrieval of relevant tables and columns due to structured setting of schemas and vague user query, 3. Generation of syntactically or logically flawed SQL due to a lack of robust validation and correction mechanism. To address these systemic challenges, we introduce Reflect-SQL, a novel framework for Text to SQL, grounded in multi-stage self-reflection approach to develop understanding of obscure schema using a knowledge base, setup a process for effective retrieval and system to generate syntactically/semantically SQL. Instead of a single-pass attempt, our system employs an LLM-as-a-judge driven scoring mechanism within interconnected feedback loops to iteratively refine the results at every stage. A feedback-driven retrieval loop refines the user's natural language query, while a synthesis loop validates and corrects the SQL and finally, an entailment loop optimizes the end-to-end process and continuously enriches the knowledge base. By integrating these layers of reflection, Reflect-SQL bridges the critical gap between user intent and complex data. On the challenging BIRD benchmark, our framework achieves an execution accuracy of 72.03%, significantly outperforming state-of-the-art baselines, demonstrating a major leap in reliability for enterprise applications.
