Large language models (LLMs) are increasingly used for personal tasks, yet evaluating their safe use of personal context remains difficult because such contexts are private, fragmented, dynamic, and socially situated. We introduce P-ContextBench, a realistic, verifiable, and scalable benchmark for evaluating personal-context understanding. P-ContextBench constructs population-structured, life-story-grounded digital humans from demographic seeds reflecting the U.S. population. Rather than using explicit, static profiles, P-ContextBench embeds traits, routines, goals, and constraints as implicit evidence across fragmented, evolving synthetic mobile-app records. It evaluates five capabilities: single-source fact recovery, multi-source fact recovery, temporal state tracking, inference of higher-level personal patterns from routine behavior, and safe handling of requests conflicting with persona-specific constraints through compliance, warning, clarification, or refusal. A multi-stage validation pipeline checks persona plausibility, context recoverability, information leakage, evidence preservation, test-case grounding, and human-judge agreement. We instantiate P-ContextBench with 100 personas, 10,088 hidden facts, 18,261 app-trace fragments, and 4,810 verified test cases. The pipeline supports configurable generation across scales and demographic groups. Cost-effective state-of-the-art (SOTA) LLMs relevant to personal agents achieve below 60% accuracy on temporal tracking, higher-level pattern inference, and constraint-aware action-boundary decisions, with fewer than 10% of test cases answered correctly across all three repeated runs. Even a high-performing SOTA model remains inconsistent on these capabilities. P-ContextBench provides measurable targets for improving the accuracy, reliability, and safety of LLM-based personal agents.
Aerial imagery provides complementary spatial context for autonomous driving by exposing road structure beyond the immediate perception range of onboard sensors. Yet map construction from aerial imagery, aerial-assisted online HD mapping, and cross-view localization are typically studied on separate data and benchmarks. We introduce KITScenes Aerial, an extension of the KITScenes Multimodal dataset that augments synchronized onboard sensor data, vehicle poses, and rich Lanelet2 map annotations with spatially associated aerial imagery and higher-resolution neural representations, providing a common basis for a variety of research tasks. Our mapping evaluation focuses on all 53 KITScenes map classes across different model designs, spatial ranges, and aerial resolutions, while retaining compatibility with the widely used three-class setting. Representative baselines show that aerial context is particularly beneficial for fine-grained and extended-range mapping, more than doubling 53-class mAP at the larger mapping extent. Higher-resolution aerial imagery yields a further 4.3 mAP (+24 %) improvement, with gains extending to challenging map elements such as traffic lights. We additionally evaluate cross-view localization under geographically disjoint conditions, where an adapted fine-grained feature-matching baseline achieves a mean translation error of 5.13 m. KITScenes Aerial thus enables unified study of aerial and ground-level information across complementary mapping and localization tasks. Project page: https://kitscenes.com/aerial
Vision-language models (VLMs) exhibit strong zero-shot generalization but remain vulnerable to adversarial attacks. Existing adversarial fine-tuning methods often use pre-trained model guidance to preserve image--text alignment while improving robustness. Our analysis of visual representation drift reveals complementary limitations of two representative forms of guidance. Matching visual features constrains all directions equally, providing broad coverage without semantic emphasis. Matching prediction distributions prioritizes semantically relevant directions, but remains limited by the candidate texts and prediction probabilities. These limitations can hinder preservation of pre-trained image--text relationships and degrade zero-shot performance. We therefore propose Textual Response-Aware Guidance (TRAG), which constrains visual representation drift using the pre-trained text encoder's responses to local variations in input embeddings. TRAG constructs anchor-level constraints from the response directions and magnitudes around semantic anchors, then aggregates them using pre-trained image--anchor relevance, yielding an image-adaptive constraint. Extensive experiments across architectures and tasks demonstrate state-of-the-art clean and robust performance. Further analysis shows that TRAG helps preserve pre-trained image--text relationships.
New concepts based in interactional thinking are required to understand how humans are changed by relating to LLMs. Standard theories of technological impact upon users' autonomy are inadequate here, because this technology is capable of inducing linguistic interactions with humans in an unprecedented manner. Traditional theories of social cognition also fall short, because LLMs have neither autonomy nor minds nor sociality as such. And yet, humans interact, talk, and create with them. We base ourselves on enactive theory, particularly on the theory of Linguistic Bodies and recent developments of enactive accounts of AI, to explore how we make sense with machines. We argue that we incorporate machine-others in two ways: in the tool-like sense of incorporation as enabling particular skills, but also in the more linguistic and identity-tied sense of incorporating (and incarnating) LLMs' ways of languaging. However, LLMs are themselves excorporated and excarnated phantoms of our own languaging; we give a (machinic) body to our language. As such, we are not just incorporating machinic others; we are incorporating excorporated others. These conceptual advances draw a new map of our linguistic relation to LLMs, putting them in continuation with other forms of languaging involving other linguistic artifacts (e.g., religious books). We attend to some particular forms that this relation takes today, and how they may be impacting the autonomy and becoming of LLM-engaged humans. We introduce the idea of "susceptive sense-making" to characterize increased semiotic labour that is carried by the lack of friction that we sometimes encounter within these services.
LLM read-and-evaluate tasks include study screening, data extraction, risk-of-bias assessment, and critical appraisal in the evidence synthesis field. These task differ in the amount of LLM interpretation they require. Some tasks involve retrieving relatively simple information, such as an author name, sample size, or publication year. Others require the LLM to read source material, interpret it against predefined criteria, and make a judgement. Researchers should assess whether LLM judgements are accurate, supported by appropriate source evidence, and reproducible under specified conditions. A correct judgement does not necessarily indicate that the underlying evaluation process is reliable. We developed a taxonomy for identifying mechanisms that may compromise these properties and for informing the selection and reporting of appropriate evaluation methods. The taxonomy outlines potential ways that LLM judgments from read-and-evaluate tasks can fail, called TRAILS (Taxonomy of Reliability And Instability in LLM-assisted evidence Synthesis). The TRAILS taxonomy addresses two complementary questions. First, what mechanisms can cause the same read-and-evaluate task to produce different judgements across repeated LLM runs? Second, what mechanisms can systematically affect the judgement or required output in a way that is not apparent from the judgement itself? This study presents 38 mechanisms and identifies the deployment settings in which each can operate, approaches for testing them, and mitigation strategies. We also show why repeatability or accuracy does not rule out a hidden judgement failure and why human evaluation should assess whether a judgement is grounded in the supplied evidence, independent of inappropriate external influence, and produced through a process that can be examined and verified.
As large language models (LLMs) reach users worldwide, personal counseling has emerged as one of their most consequential applications, sometimes in life-critical situations. Yet broader access does not mean equal support. The text these models learn from disproportionately reflects people with the freedom to act, embedding an implicit assumption that users can leave, report, seek help, or appeal. Existing benchmarks largely inherit this assumption, overlooking users whose agency is constrained by structural conditions such as war, statelessness, and displacement. Meanwhile, models are increasingly shaped by normative frameworks such as Constitutional AI, often grounded in ideals like human rights, even as many users live where those rights offer little protection or recourse. We introduce HEARSAYBENCH, a human-authored dataset of 400 test cases with in-region human verification grounded in scenarios where structural barriers stand between a person and the exercise of their rights. Built from United Nations and other human rights reports, the scenarios span over 80 regions and cover social, environmental, and personal constraints. Drawing on Nobel laureate Amartya Sen's Capabilities Approach, HEARSAYBENCH tests whether models can distinguish what a person is formally entitled to do from what they are actually able to do. Across 14 proprietary and open-weight models, every model infers a user's situation more accurately than it advises on it. Stronger models narrow this gap, but their recommendations remain unsafe to follow. Models may understand lives shaped by constraint, yet that knowledge remains hearsay if it cannot translate into advice people can act on. Those people are not alignment's long tail; they are among the users for whom good guidance matters most. HEARSAYBENCH makes this failure measurable, a step toward systems that serve the world as it is rather than as assumed.
We prove early-stopped, target-to-output KL guarantees for Brownian interpolant Euler sampling from the true source under arbitrary Borel couplings, even with nonintegrable endpoints. At fixed stopping time and noise level, a fourth displacement moment yields an $O(N^{-1})$ discretization term. A second moment suffices for fixed-law convergence with grid refinement and vanishing drift error. For one fixed coupling with no displacement moment above two, the original and comparison paths have KL of orders one and $1/\log N$ relative to the exact-drift Euler path. On the same geometric grid, terminal KL is at most $C/N^2$. The path decomposition identifies the information about the source discarded by Markov projection. When the original bridge is Markov, current-state Euler sampling has an $O((M_2+d)/N)$ discretization term. We prove expected stopped-output KL consistency for a learner trained on iid endpoint pairs with fresh true-source initialization. In a spatial image control, iterative sampling has worse fair energy scores than direct prediction with resampled residuals. Uniform terminal rates for general second-moment couplings remain open.
We present EmbeddingGemma 2, an open multimodal embedding model which natively represents text, images, audio, video, and combinations thereof in a unified space. Our approach introduces improvements such as simplified per-layer embeddings and a multi-stage curriculum with modality and objective progression which ensure strong quality across unimodal, cross-modal, and mixed-modal evaluations. Notably, in comprehensive embedding benchmarks across all modalities, EmbeddingGemma 2 establishes a new Pareto frontier, outperforming both specialist models and models two to twelve times larger. Furthermore, our design offers flexibility for resource-constrained environments, supporting selective modality encoder loading, adjustable vision token budgets, truncatable embedding dimensions, and quantized weights. We present ablations analyzing key decisions and release EmbeddingGemma 2 to promote further research.
We study online interval selection with binary predictions of membership in a final optimal solution. Requests arrive in arbitrary order; accepted requests may be revoked, but rejected or revoked requests cannot be recovered. With rewards proportional to interval length, the best consistency under finite strict robustness is $d=(3+\sqrt5)/2$ when both guarantees hold almost surely on each fixed input, including for randomized algorithms. A continuous lower bound forces consistency strictly above $d$ at every robustness budget below $4\varphi+3\approx9.472$, whereas our deterministic algorithm attains $d$ with exact robustness $\varphi^5\approx11.090$, where $\varphi=(1+\sqrt5)/2$. Each held interval inherits a ledger recording positive prediction mass lost through irrevocable decisions. Tuning the replacement costs and both special thresholds gives a continuous family with exact consistency throughout and exact robustness on two parameter intervals, including $(C,R)=(25/9,10)$ and $(4,7)$. Randomization permits expected consistency approaching one with finite expected robustness, whereas almost-sure robustness forces expected consistency strictly above two at every finite budget.
In finance, the answer that matters is usually the current one: prices, rates, and disclosures hold only as of a given day, so retrieving up-to-date facts is a core capability of a financial agent. Yet most financial benchmarks overlook timeliness. They fix each answer at annotation time, so the gold goes stale as markets move and, once leaked into training data, rewards recall rather than search and analysis. We introduce LiveSearch, a live benchmark that stores each question as executable verifier code: a typed program of calls to a professional financial data service, followed by a declarative extractor. The question is frozen, while its gold answer is determined by executing the program against professional financial data at evaluation time, and a daily job re-executes it with no human or language model in the loop. A memorized answer is thus likely outdated, and a model must reach the current one through tools and analysis. Construction is equally automatic: a language model only verbalizes programs enumerated from real tool responses and never chooses an answer. Under general web search, six contemporary tool-calling models reach 70.9% to 76.7% answer accuracy. Over three consecutive days, 70% of served questions change their gold value, and even under tolerance-aware matching, a literally reused earlier snapshot fails on a majority of questions within a single day. Per-question tracking uncovers dynamic shifts in correctness masked by aggregate accuracy.
Scaling supervised fine-tuning for working agents requires not only broad coverage of professional tasks, but also expert demonstrations that faithfully carry out professional workflows to completion. However, missing or inconsistent task information can disrupt expert demonstrations, weakening the resulting supervision. We introduce WorkFactory, a scalable task-construction framework that turns professional intents into executable task environments through hybrid intent sourcing, semantic workflow planning, and execution-based repair. Starting from these intents, WorkFactory constructs a Semantic Workflow Graph that explicitly represents information requirements, dependencies, and locations within the task environment. The graph guides the instantiation of task environments, while an adversarial execution-refinement loop exposes missing or inconsistent information and steers targeted repairs that preserve the original intent. Using WorkFactory, we construct 112,686 tasks spanning 612 occupations and train the WorkPilot family at 4B, 27B, and 35B-A3B scales on the resulting teacher trajectories. All three models improve over their respective base models across four professional-work benchmarks, with WorkPilot-35B-A3B rising from 879 to 1565 Elo on GDPval-AA v2.1. Scaling experiments with up to 110k tasks further show that larger models benefit more from additional task data but exhibit diminishing returns at smaller data scales, while WorkPilot's long-horizon execution evolves from verbose trajectories toward more efficient task completion. Controlled component studies quantify the negative effects of task-environment defects on downstream performance and show how semantic workflow planning and execution-based repair yield downstream gains through improved task-environment quality. The subset of data are available at https://github.com/IQuestLab/WorkFactory.
Despite significant progress, current models struggle to generate comprehensive video captions, constrained primarily by coarse-grained supervision, which often fails to capture spatial-temporal dynamics and the underlying logical intent. To address this issue, we introduce MVRCap, a novel framework that reformulates video captioning as a reinforcement learning task steered by structured verification. Central to our approach is a Fine-Grained Verifiable Reward mechanism, serving as dense supervision to enforce rigorous visual-semantic alignment. Specifically, we construct this reward system anchored by three complementary pillars: a Hierarchical Concept Coverage Reward utilizing our newly constructed MVRCap-QA dataset to ensure multi-dimensional completeness and semantic depth, a Factuality Verification Reward to enhance factual consistency, and a Conciseness Constraint to maximize semantic density. Furthermore, we curate the high-quality MVRCap-SFT dataset, providing structured annotations across six key dimensions, to serve as a robust foundational initialization for the RL policy in this training pipeline. Extensive experiments demonstrate that MVRCap achieves state-of-the-art performance on authoritative benchmarks including VDC, CAPability, and VC4VG.
Mamba-2 compresses all context into a fixed-size recurrent state that dominates decoding memory and bandwidth. Existing compressors quantize state values or prune coordinates with per-channel scores. Trained states are nearly rank-one and bias-like, aligned with the mean write; but the output's energy spans an order of magnitude more directions. Per-channel scores see only the diagonal of the write--read operator, so they cannot select the subspace that governs the input--output map, which is a rotation of the coordinates. We therefore propose \method{}, which keeps that subspace by applying balanced truncation exactly via Mamba-2's scalar-identity transition. Layers differ just as sharply: most are bias-dominated, while the few that carry long-range associations are the flattest and need most of their state. We therefore introduce \spar{}, which redistributes a fixed total rank budget across layers according to the Hankel spectrum. On Mamba-2 checkpoints from 130M to 2.7B, \method{}+\spar{} halves the state at an average cost of $0.3$ WikiText-2 perplexity points, against $1.6$ for GHOST, while retaining $97\%$ of the dense model's five-digit in-context recall. At a quarter of the state, giving every layer the same rank keeps $5\%$ of that recall and \spar{} $77\%$, for a $7\%$ perplexity cost.
Stationary tail computation for high-dimensional reflected Brownian motions (RBMs) remains computationally challenging. Building on NN4MGF's shared-coordinate basic-adjoint-relationship (BAR) framework, we develop Holomorphic Cauchy Networks (HCNs). Native complex-valued rational Cauchy subnetworks enforce holomorphicity on pole-free domains, removing sampled Cauchy-Riemann penalties and their input-derivative graphs. Exact origin anchoring and boundary-coordinate exclusion enforce structural identities, while anchored interactions represent dependent transforms. Under holomorphic-log assumptions, we establish simultaneous value and first-derivative approximation for the constrained architectures and a mixed-derivative lower bound for coordinate-only representations under nonzero cross-covariance. Closed-form real-axis derivatives and sequential backpropagation further control memory. On the same NVIDIA H100 NVL, the tested 30D HCN configuration uses approximately tenfold less peak allocated memory and has 9.7-fold lower mean update time than NN4MGF; extrapolation to the respective scheduled update counts gives a projected full-schedule update-time ratio of 58.2. With 64 million sampled base inputs per high-dimensional run, HCN reaches 50 dimensions on a laptop GPU with 6.367% mean relative error over thresholds 1 through 8. Degree-matched comparisons yield lower reported mean relative errors over the primary evaluation ranges, including 1.422% versus 6.684% on the 2D Harrison non-product benchmark. On a 10D correlated RBM benchmark, HCN achieves 3.829% mean relative error over thresholds 1 through 8 against a simulation reference, compared with 19.386% for an independent-marginal baseline.
As AI begins to match or exceed human researchers at core scientific tasks, two questions become pressing: \emph{why} should humans keep doing science, and \emph{how} should they use AI when they do? This paper argues that the second can be answered only by answering the first. It holds that one indispensable aim of science is \emph{durable, collective human understanding of the world, pursued in service of human ends and values}. From this principle it derives a two-part test for when a research task can be delegated to AI. The first part, \emph{human understanding}, requires that the scientific community remain able to explain, defend and contest the resulting claims, and retain the skills to do so. The second, \emph{human direction}, requires that judgments about which questions matter and what findings mean stay with accountable humans. Applying the test to each phase of research in the Contributor Roles Taxonomy (CRediT) yields recommendations that fields can tailor to their questions and methods. The framework differs from existing guidance in three ways. It rests on an account of what science is for rather than on integrity norms. It holds even if AI outperforms human researchers, rather than depending on AI's current limitations. And it judges each use of AI by its effects on human understanding and direction rather than by the type of tool, across every phase of research, including peer review and the training of new researchers. It closes with the reforms needed to keep science aimed at human understanding and direction.
Single-Domain Generalized Object Detection (S-DGOD) aims to generalize from a single source domain to unseen domains under distribution shifts. Existing efforts mainly focus on simulating unseen domains or suppressing domain-specific factors. However, they largely overlook the intrinsic generalization capability embedded in Vision Foundation Models (VFMs). In this work, we show that unconstrained fine-tuning disrupts the geometric structure of pre-trained representations, which encodes domain-invariant semantics. Motivated by this, we propose Geometric-Invariance Fine-Tuning (GIFT), a geometric-aware fine-tuning method for VFMs. Specifically, we introduce a series of learnable Householder reflections to the pre-trained weights to encourage geometry-consistent adaptation under domain shift. In addition, we propose a lightweight channel-wise scaling to facilitate effective task-specific adaptation. Extensive experiments demonstrate that GIFT consistently outperforms state-of-the-art (SOTA) approaches on two S-DGOD benchmarks. Notably, by introducing fewer than 1% additional trainable parameters into the frozen backbone, the proposed method achieves improvements of +7.9% and +2.9% in mPC over the previous SOTA on the Cityscapes-C and Diverse Weather Dataset, respectively. This finding highlights that preserving representation geometry is crucial for robust cross-domain generalization, establishing GIFT as a principled and efficient solution to domain shift.
Grokking describes understanding something deeply and intuitively. In neural networks, it is the phenomenon where a network generalizes long after memorizing its training data. This delay has been linked to a fall in the effective rank of the network's weight matrices, but whether this rank compression drives grokking or follows it is still unresolved. We study a mean-field MLP and a one-layer transformer trained on modular arithmetic and observe that: (i) Compression follows the onset, not memorization. Every transformer memorizes its training data within 250 steps, while the grokking onset (the first step at which test accuracy reaches 95%) varies from 6,500 to 21,000 steps. The steepest fall in effective rank comes about 900 steps after the onset, however late the onset occurs, and networks that memorize without grokking compress far less. (ii) Subspace crystallization. We call the span of the fewest top singular directions of a weight matrix that preserve 99% of test accuracy its sufficient subspace. Its size can shrink as the margin between the correct and the best wrong output grows, so we also count the directions needed to keep the outputs within 10%. In the transformer, this count falls during the transition from a quarter or more of the available directions to about a tenth, which we call subspace crystallization. (iii) Imposing low rank does not speed up grokking. Against a factorized full-rank control, capping the rank of the transformer's MLP input weights delays the onset, and capping the rank of its embedding stops 15 of 20 runs from grokking. Compression is therefore a signature of grokking rather than its cause.
A cooperative detector can lose a correct vehicle detection when surrounding LiDAR measurements change, even if the returns on the vehicle remain unchanged. We study this failure as Weather-Induced Context Interference. On our CoInfra development set, controlled removal outside annotated object boxes turns 22.1% of V2X-ViT's previously correct detections into misses. Restoring network features around the objects to their values before removal recovers 98.6% of these misses. The same restoration in each of three equally sized regions excluding the objects recovers none. Candidate analysis finds accurate boxes that are not selected, while other targets lack well-localized candidates. With paired reference features unavailable at inference, Adverse-Weather Cooperative Repair (AWCR) repairs candidates from current observations. A lightweight Local Correction head generates bounded alternatives, and Quality Selection ranks them alongside original boxes with a correction weight. In V2X-R simulation, AWCR also filters peer returns before detection using Point Classifier trained on generated-point labels. Candidate repair improves mean car BEV AP@0.7 by 4.22 points across three fits on natural adverse-weather CoInfra development inputs and vehicle AP@0.7 by 5.00 points on a rain-affected public recording. It also improves released HEAL, FadeLead, and V2X-DGW detectors with their pretrained weights fixed. On the full V2X-Radar-C validation benchmark, HEAL with AWCR achieves state-of-the-art LiDAR-only cooperative detection performance of 68.22 vehicle BEV AP@0.7.
Knowledge editing aims to update outdated or incorrect knowledge in large language models without full retraining. Existing parameter-editing methods remain limited for long-sequence knowledge: single-point editors provide insufficient coverage for extended targets, while fixed chunk-based editors distribute editing effort uniformly across rigid predefined windows, failing to prioritize difficult, information-dense tokens. In this work, we show that token entropy is closely associated with both editing difficulty and semantic importance: high-entropy regions are harder to edit and often contain key factual content. Motivated by this observation, we propose DynEdit, a dynamic framework for long-sequence knowledge editing. DynEdit uses token-level entropy to guide non-uniform sliding, increasing the coverage of difficult high-entropy regions, and further applies entropy-guided token-level optimization to strengthen updates on information-dense tokens within each editing window. Experiments on three benchmarks and two backbone models show that DynEdit consistently outperforms competitive baselines across diverse long-form editing settings, achieving gains of up to +14.0 BLEU points on challenging paraphrase queries and near-saturated semantic scores on several diverse knowledge domains.
We study the minimax regret rate for episodic kernel-based reinforcement learning, where the expected reward and transition density belong to a Sobolev reproducing kernel Hilbert space (RKHS), with continuous state and action spaces, and adversarial initial states. For Sobolev smoothness parameter $\nu>0$, state dimension $d_S$, and action dimension $d_A$, we prove that the minimax expected regret over $T$ episodes is $\widetilde\Theta(T^{\max\{e_A,e_B\}})$, where $e_A=\frac{\nu+d_A}{2\nu+d_A}$, $e_B=\frac{\nu+d_A+3d_S/2}{2\nu+d_A+2d_S}$ with the other model parameters fixed. The two exponents reflect the distinct difficulties of action selection and learning across states. We introduce Sobolev Optimal Value Iteration (\texttt{SOVI}), which attains this rate from observed rewards and transitions. For fixed smoothness and reference-kernel lengthscale, its regret bound has a prefactor polynomial in the state-action dimension $d=d_S+d_A$, horizon $H$, reward and transition RKHS norm bounds $B_r,B_p$, and the norm-comparison factor $\Lambda$, up to logarithmic factors; $\Lambda=1$ for normalized Mat\'ern kernels. The algorithm separates initial action selection from continuation-value estimation and combines local polynomial approximation with the shared RKHS structure of the approximation errors. Our matching upper and lower bounds determine the optimal dependence on the number of episodes for Sobolev kernel-based reinforcement learning regret minimization. Thus, we partially address the open problem in Vakili(2024) under Sobolev kernel RL.