OpenPrint

A language model can state what an accessibility concept means and still fail to apply it. We measure this declarative-evaluative gap in thirteen models from three families: Pythia, GPT-2, and OLMo 2. In the original behavioral batteries, the gap opens early and persists across families. At Pythia-12B, it closes because declarative accuracy falls. Three concepts answered correctly at smaller scales regress to incorrect at 12B. Convergence by decay is not mastery. In a frozen same-concept validation of eight concepts, each represented by one violation and one conformant item, none of 96 confirmatory model-concept cells passes both application items. Eleven cells pass one item; a development-exposed pilot, reported separately, passes one pair. The 96 confirmatory cells include 35 with correct definitions. We test two possible explanations for the gap. The global maximum of attention between a compound's constituents has weak associations with accuracy that differ in direction across families. When the measure is restricted to late-layer, value-weighted binding, a consistent pattern appears: rarer compounds receive stronger distributed value-weighted writes at all thirteen model-scale points (p = 0.0003). This is the opposite of what we would expect if stronger binding indicated knowledge. The pattern is consistent with processing difficulty, although the experiments do not identify what computation the stronger writes perform. Corpus frequency is the strongest measured predictor of accuracy across 49 compounds (Spearman ρ = 0.52–0.59 across families), but it leaves most compound-to-compound variation unexplained. The same concept can be available to definition and unreachable in application, even though its corpus frequency has not changed. Frequency predicts availability. Task form constrains reachability.

Multimodal queries can require different types of reasoning. Some can be answered via perceptual reasoning, extracting information directly from the visual signal, while others require compositional reasoning that combines observations or deliberative reasoning that evaluates competing hypotheses. However, many existing methods apply a uniform reasoning strategy across queries, leading to unnecessary computation on simple tasks and insufficient reasoning on complex ones. We introduce Video-FLAIR, a training framework that learns to select the appropriate reasoning mode for each query using reinforcement learning. During training, the model generates responses under all three modes for the same prompt, enabling direct comparison. A composite reward compares these responses to favor the most effective one based on correctness, grounding, and cost, while discouraging unsupported or misaligned deliberation. This yields a supervision signal for learning adaptive reasoning without per-query annotations. Video-FLAIR improves accuracy over the Qwen2.5-VL base model by +5.4 on MathVista, +4.8 on Video-Holmes, and +4.8 on Video-MMMU, while reducing average token usage to 95 compared to 417 for always-thinking baselines.

Enterprise deployments of autonomous AI agents inherit a control model built for human users and long-lived services, and the fit fails in three specific ways: agent principals are ephemeral, appearing and vanishing faster than provisioning; their actions are selected by a model rather than programmed, so the set of things they may attempt is not known in advance; and the population is discovered rather than provisioned, because anyone who can call an API can create one. We argue that governing such agents is a runtime problem -- not a model-alignment problem and not a build-time problem -- and we derive five primitives from the questions that must be answered before an action takes effect and after it has: discovery, identity, governance, attestation, and supply chain. For each we state what fails if it is absent and why the others cannot structurally supply it. We describe an implementation in which an agent's action is mediated against policy before it takes effect, authorised against a per-tenant action vocabulary, and recorded in a hash-linked signed ledger a third party can verify with the vendor out of the loop. We report what the architecture costs: the enforcement point sits on the request's critical path, identity requires a sidecar per workload, and fail-closed mediation converts availability incidents into denial. We are explicit about implementation status: four primitives are built and running in private pilots, and the fifth is built as separate tooling and not yet integrated into the request path. We keep it in the set deliberately: a five-part decomposition that exactly matches what its authors happened to build is not a taxonomy but a description of a codebase.

Self-Generated Text Recognition (SGTR)--the ability of an LLM to identify its own outputs--poses risks to AI safeguards that rely on LLMs as evaluators or monitors. Specifically, an LLM may recognize outputs from other copies of the same model and make biased judgments or collude outright. Prior work has drawn conflicting conclusions about whether current models possess significant SGTR capabilities. We reconcile these findings by identifying key experimental design choices--which we term operationalizations--that drive divergent results. Evaluating 13-21 models across six operationalizations, we find that accuracy varies substantially with evaluation format (pairwise vs. individual assessments of text), conversation structure (presenting candidate text in user tags vs. assistant tags), and the domain of the task used to generate candidate text (e.g., coding vs. summarization). We corroborate previous observations that a quality heuristic--models attributing authorship to text they perceive as higher quality--is a dominant confound. We also find that improving a model's SGTR performance via SFT in one evaluation configuration can generalize to others. Training for SGTR additionally causes models to prefer their own outputs when acting as a judge in the AlpacaEval framework. Finally, we discuss the implications of our evaluations for the safety of future AI systems: our work suggests that, despite confounds, some models possess practical SGTR capabilities, and that training a model for SGTR in one setting can affect its self-recognition and self-preference more generally. We conclude that SGTR should be monitored and considered in the design of safety-critical AI applications.

On-policy self-distillation (OPSD) has recently emerged as an effective post-training paradigm that improves policy optimization through dense token-level supervision from a privileged self-teacher. Despite its promise, OPSD remains largely underexplored for Video Large Language Models (Video-LLMs). Existing methods typically construct privileged teachers by augmenting their context with additional information while keeping the primary input unchanged for both teacher and student. Video reasoning, however, offers a distinct source of privileged supervision within the primary input itself: long videos contain substantial temporal redundancy, and only a small subset of frames provides the evidence necessary to answer a question. Building on this observation, we present Video-OPSD, an OPSD framework that exploits privileged visual evidence for both self-teacher construction and knowledge transfer. First, our Evidence-Grounded Self-Teacher conditions the teacher exclusively on annotated evidence frames while the student continues to reason over the complete video. This focused visual input enables the teacher to provide more informative supervision. Second, our Evidence-Guided Token Optimization adaptively weights token-level distillation according to each reasoning token's reliance on privileged visual evidence, thereby emphasizing perceptually grounded reasoning. Experiments across video understanding and reasoning benchmarks show that Video-OPSD consistently improves upon Standard OPSD across multiple backbones and achieves performance comparable to GRPO while requiring substantially less training time, establishing an effective and efficient post-training approach for Video-LLMs.

This paper proposes a framework for constructing a classifier as a safeguard layer, and for developing a complementary diagnostic that identifies which of the classifier's confident decisions can be trusted. This framework, the Latent Diagnostic Taxonomy, consists of (i) constructing a dimensionality-optimized classifier, in which the embedding dimensionality is empirically selected via cross-validated performance rather than fixed a priori, (ii) locating a relatively small set of latent support vectors (~ 29% of total training examples) representing influential prompts for identifying tokens that alter the classifier's predicted labels, and (iii) utilizing such tokens and their associated attack magnitudes for constructing a diagnostic taxonomy. This diagnostic taxonomy provides an end-to-end guideline for flagging prompts that require different treatments: rely Safely on the classifier's decision; flag Heuristic Bias and Heuristic Override cases; route Insufficient Context cases for further human/safety review. Applying the framework to a classifier trained on a public prompt injection dataset, we find that a substantial fraction of its confident decisions (~ 77%) are not robust to removing a single token, and that this brittleness separates into two distinct failure patterns: a confidence calibration failure and a genuinely exploitable shortcut. For each zone of the taxonomy, we also recommend strategies for remediating diagnosed prompts. We illustrate the framework as a series of steps, demonstrating how each step operates.

When a language model encodes a sensitive attribute, the source matters: the signal may be carried by explicit or distributed input cues, or it may remain after controlling for a declared input representation. Standard probing conflates these cases, leaving auditors unsure whether to inspect the input pipeline or the learned representation. We propose Controlled Representation Attribution (CRA), a reference-conditional auditing protocol that residualizes each layer against an input reference and probes the residual. We use “model-added” conservatively to denote reference-residual sensitive information rather than a causal claim about model weights. CRA is supported by a recovery theorem—exact when the residual noise is equal-covariance Gaussian and the target direction is an eigenvector of that covariance, with an explicit finite-anisotropy bound otherwise—and by scope and reference-admissibility diagnostics that identify inadequate controls. In controlled simulations with known ground-truth directions, CRA wins the selectivity comparison in 60/60 runs; falsification tests further show that weak references can create spurious residual conclusions, while the admissibility gate withholds such claims. In a prospectively specified controlled-text audit, residual evidence survives name removal but not removal of broader lifestyle and stylistic cues, demonstrating dependence on the declared control. Archived pretrained-language-model analyses illustrate a lexical-gender/reference-residual-age contrast and a positive pooled medical-question-answering difference (22 of 27 cells, mean +5.3 percentage points), which is not powered at the dataset level; these analyses are exploratory rather than confirmatory. CRA is a source-attribution diagnostic that reports reference sensitivity and can abstain when controls are inadequate; it complements concept-erasure methods rather than replacing them.

Perturbation methods explain model decisions by measuring prediction changes under altered inputs, but response magnitude tells us only how much a model reacts, not what that reaction means. The same magnitude can support the final factual–counterfactual difference, oppose it, or arise strongly along the perturbation path yet vanish at the endpoint. We therefore track how the contrast develops as paired inputs are progressively revealed, using the final contrast to interpret the trajectory. We introduce DECAF (Decomposition of Evidence, Contradiction, And Fragility), which routes aligned, opposed, and endpoint-null responses into evidence E, contradiction C, and fragility F. The decomposition preserves ordinary magnitude exactly, Abs = E + C + F, and is unique under endpoint-relative axioms. Across controlled vision and tabular settings, the three components track independently measured behavior. In a 72-model ImageNet-9 audit, we compare cases with nearly identical response magnitude but different independently measured behaviors. The largest DECAF component agrees with an observed behavior in 96.4% of cases, compared with 35.0% for magnitude alone. Changing only the reveal path increases total response by nearly 80%, yet evidence barely changes while fragility grows by more than 4×. On FunnyBirds and ImageNet-1k, short forward-only DECAF trajectories outperform the tested general-purpose attribution baselines. On a 1B-scale DINOv2 model, a short trajectory matches a strong gradient-based baseline with 4.75× lower wall time and 2.36× lower peak memory.

Memory has evolved into a foundational architectural dimension in large language models (LLMs), shifting from an implicit byproduct of computation to a spectrum of explicit, controllable mechanisms. While recent advances introduce diverse strategies---spanning transient attention, recurrent state dynamics, parameter-efficient adaptations, and scalable lookup storage---this rapid evolution has led to a highly fragmented research landscape. In this survey, we present a systematic, architecture-centric taxonomy of memory in LLMs. Our framework characterizes memory along three orthogonal axes: representation (implicit versus explicit), update dynamics (offline versus online), and persistence (short-term versus long-term). We further formalize the granular mechanisms dictating memory writing, routing, state transitions, and consolidation. This unified perspective elucidates the conceptual boundaries between computation-coupled and independently addressable memory, effectively bridging disparate architectural paradigms. Additionally, we critically analyze hybrid memory architectures, system-level efficiency trade-offs, and multi-dimensional evaluation methodologies. By consolidating these scattered advancements into a cohesive framework, this survey charts the trajectory of memory-centric LLM design and provides a principled foundation for future innovations in scalable and adaptive language modeling.

Split federated learning protects client data by injecting Gaussian noise into intermediate representations before they are sent to an edge server. When clients run heterogeneous model architectures, their representations differ in sensitivity, yet an unclipped implementation may calibrate every release to one shared surrogate C. We isolate the resulting Sensitivity Gap: the claimed Gaussian-mechanism certificate is not supported when C is not a valid ceiling for every released representation. This is a mechanism-specific calibration failure, not a flaw in differential privacy composition or a lower bound on the mechanism's minimal privacy parameter. We give a round-indexed recalculation of a valid transcript certificate and propose Sensitivity-Aware Budget Allocation (SABA), which quantizes client sensitivities into global, tiered, or per-client calibration profiles. On heterogeneous MLP and Vision Transformer encoders for CIFAR-10 and CIFAR-100, an empirical-input diagnostic and an independent one-sided attack witness decrease by more than 82% with 0.1–0.6 accuracy-point changes. For MLPs, a completed five-seed sweep using per-round spectral-norm ceilings restores the nominal certificate. The empirical study is a controlled proof of concept; only the fixed-state MLP releases supplied with valid spectral ceilings are certified.

AI reviewers are gaining more attention with the advent of LLMs. With the increasing volume of conference papers, the use of AI-based reviewers has been suggested and implemented to enable faster review cycles, improve review quality, and help the scientific community. In this position paper, we argue against the replacement of peer review by AI reviewers and advocate the use of AI review tools primarily to verify claims and improve the quality of the scientific work. We argue that AI review tools should be utilized primarily by authors during manuscript preparation to improve submission quality and streamline downstream evaluation. We ground our argument in the values of peer review and scientific process. Finally, we present research directions for the responsible integration of AI in the peer review process.

When a student must learn concepts connected by prerequisite dependencies, when does the order of instruction matter, and what does it cost to find the best one? We study instructional sequencing as a stochastic shortest-path problem in which attempting a concept succeeds with a state-dependent probability and failure leaves the learner state unchanged. We first prove that this stochasticity can be eliminated exactly: the problem collapses to a deterministic shortest-path problem on the lattice of prerequisite order ideals, preserving optimal values and actions. The collapse removes stochastic complexity but not combinatorial complexity: optimal sequencing remains NP-hard—via reduction from feedback arc set in tournaments—even with no prerequisite edges, unit costs, uniform binary nonnegative transfer, and success probabilities at least 1/2. Hardness is not uniform: when realizable transfer preferences remain jointly acyclic with the prerequisites, any topological order of the residual joint graph is optimal, and fixed prerequisite width yields polynomial-time exact dynamic programming. A computable diagnostic, mΔ, bounds the value of sequencing before optimization. On 70,893 interactions from an introductory CS course, the diagnostic certifies a doubly easy regime—little value to optimize and little space to search—while constructed transfer instances realize the challenging regime, where myopic sequencing suffers large regret yet exact A* with a consistent heuristic expands only linearly many states on that family.

Vision-and-Language Navigation (VLN) requires embodied agents to ground natural language instructions in visual perception and make navigation decisions in complex 3D environments, making it a central problem in embodied artificial intelligence. Since the introduction of the Room-to-Room (R2R) benchmark, VLN has made substantial progress. In recent years, as research settings have gradually expanded from closed and single indoor benchmark scenarios to open-world environments, the field has undergone a profound paradigm shift from passive instruction following on fixed benchmarks to autonomous cognitive navigation in open-world settings. However, existing surveys mainly organize prior work according to technical taxonomies, lacking a systematic characterization of this paradigm evolution. To address this gap, this survey proposes an evolution-centered unified analytical framework that reviews contemporary VLN research across four progressive layers: perception, cognition, learning, and generalization. It reveals the intrinsic connections and evolutionary logic among different technical lines, identifies key open challenges at each dimension, and outlines future research directions. This survey aims to provide VLN researchers with a clear panoramic view of capability evolution, while offering the broader embodied intelligence community a systematic roadmap from closed-benchmark evaluation toward trustworthy open-world deployment.

Transformer-based large language models are increasingly used for long-horizon tasks; however, their attention mechanism scales poorly with context length. To handle this, we study a sleep-like consolidation mechanism in which a model periodically converts recent context into persistent fast weights before clearing its key-value cache. During sleep, the model performs N offline recurrent passes over the accumulated context and updates the fast weights in its state-space model (SSM) blocks through a learned local rule. During inference, this shifts extra computation to sleep while preserving the latency of wake-time prediction. We test our method on controlled synthetic tasks, including cellular automata and multi-hop graph retrieval, as well as a realistic math reasoning task, on which a regular transformer as well as SSM-attention hybrid models fail. We then show that increasing sleep duration N for our models improves performance, with the largest gains on examples that require deeper reasoning.

Recent advances in AI agents have increasingly internalized native capabilities into their underlying foundation models, giving rise to multimodal foundation models and large reasoning models. However, agent memory is still primarily implemented through external modules, leaving the native memory capability largely unexplored. In this paper, we take a first step toward this direction by introducing memory foundation models, which empower foundation models with native memory capabilities. We formalize native memory from two perspectives: a persistent and dynamically evolving memory state within the backbone, and native memory procedures that autonomously store and utilize information through model computation. We show that native memory offers advantages in architecture, end-to-end optimization, and efficiency. Based on this formulation, we propose Metis, the first prototype of memory foundation models. Metis introduces a new architecture that equips a foundation model with a native memory state, allowing historical information to be compressed into the model and accessed through memory attention. We construct large-scale memory-specific training data and introduce multiple optimization objectives to acquire these native memory procedures through mid-training. The online memory maintenance of Metis is gradient-free, and the memory update requires only a forward pass. At inference time, all learned model weights remain frozen, while the native memory states are autonomously transformed through standard forward computation. Through extensive experiments, we show that Metis exhibits native memory capabilities and further provide a detailed analysis of its strengths, limitations, and behaviors. To facilitate future research on memory foundation models, we release our project and model checkpoints.

Split Federated Learning (SFL) is built on a reassuring premise: keep the labels on the client, and the server learns nothing sensitive. We show this premise is wrong. The gradient the server receives to train its model is not an opaque update; it encodes the private labels in a way that is mathematically unavoidable. We formalise this vulnerability, prove information-theoretic lower bounds on the leakage, and show why existing defences fail: they cannot distinguish the gradient's optimization signal from its label leakage. We prove these two components occupy different subspaces near convergence, a consequence of Neural Collapse, and build on this to propose Rate-Limited SFL (RL-SFL), which suppresses the leakage subspace by transmitting only the most informative gradient coordinates. RL-SFL comes with convergence guarantees and works well under near-homogeneous data distributions, but degrades under high client heterogeneity and provides no privacy at all under extreme skew; it also carries no formal differential privacy certificate. On CIFAR-10/100 and BloodMNIST, RL-SFL achieves substantially lower attack accuracy than standard DP-SGD and variational baselines at comparable utility, with reduced communication cost on bandwidth-constrained links.

Building AI Scientist agents with Large Language Models (LLMs) has recently attracted growing attention. Since scientific discovery fundamentally relies on uncovering causal relationships from observations, the capability of causal thinking, i.e., distinguishing causation from correlation and recognizing hidden biases, is essential to LLM agents. Although a number of benchmarks exist for AI Scientists, none explicitly incorporate challenges from selection bias, measurement error, and hidden confounders that widely exist in real-world scientific discovery. To this end, we present CausalGame, a benchmark that evaluates the causal thinking capabilities of LLM agents through interactive games. CausalGame asks LLM agents to actively design experimental protocols, collect observation data, and derive a final solution with an explanation report. To emulate realistic scientific discovery challenges, we design 14 scenarios that incorporate selection bias, measurement error, and hidden confounders. Across 30 LLM agents, none demonstrates reliable causal thinking: the best model reaches only 68.0% survival against analytical optima of 78-85%, and merely 5-7% of sessions receive credits on the causal-reasoning rubrics. CausalGame provides a scalable and controlled testbed for evaluating the causal thinking of AI Scientist agents.

Large Language Models (LLMs) exhibit strong informal mathematical reasoning but struggle to generate mechanically verifiable proofs in formal languages like Lean. We present LEAP, an agentic framework that enables general-purpose foundation models to achieve state-of-the-art performance on automated formal theorem proving. LEAP leverages foundation model capabilities, such as informal reasoning, instruction following, and iterative self-refinement. By decomposing complex problems into smaller units, the system bridges formal proof construction with informal blueprints through continuous interaction with the Lean compiler. To provide a rigorous evaluation beyond increasingly saturated benchmarks, we introduce Lean-IMO-Bench, a benchmark of IMO-style problems formalized in Lean, with short statements yet highly non-routine and multi-step proofs across a wide range of difficulty levels. Empirically, on the latest 2025 Putnam Competition, an annual mathematics competition for undergraduate students in North America, LEAP solves all 12 problems, matching recent breakthroughs by frontier formal mathematical models. On Lean-IMO-Bench, LEAP boosts the one-shot formal solve rate of general-purpose LLMs from below 10% to 70%, notably surpassing the 48% benchmark set by a specialized, gold-medal-caliber IMO system. Furthermore, we demonstrate LEAP's research-level utility by autonomously formalizing complex proofs for open combinatorial challenges, including a verified proof for a key subproblem in Knuth's Hamiltonian decomposition of even-order Cayley graphs.

Advances in AI models have sparked fears that rapid disruption of labor markets is imminent, if not already underway. A steady drumbeat of media articles has forecast a grim future for white-collar work due to AI. Fears of an “AI jobs apocalypse” are often amplified by AI leaders themselves. For example, Dario Amodei, CEO of Anthropic, has predicted that AI could wipe out half of white-collar jobs and push unemployment to 20 percent. A labor market upheaval of this magnitude would cause enormous suffering for many households and pose a significant challenge for policymakers. While the public has been debating AI’s potential impact on firms and their workers, research and data have been catching up. Our goal in this brief is to synthesize the fast-growing body of research on AI’s impact for policymakers and others eager to understand how AI is affecting the labor market right now.

How do we fundamentally discover new things? In a letter to Maurice Solovine, Albert Einstein conceptualized discovery as a cyclical process involving an intuitive 'jump' from sensory experience to axioms, followed by logical deduction. While Generative AI has mastered Induction (statistical pattern matching) and is rapidly conquering Deduction (formal proof), we argue it lacks the mechanism for Abduction—the generation of novel explanatory hypotheses. Using Einstein’s formulation of General Relativity as a computational case study, we demonstrate that the prevailing theory of "creativity as data compression" (induction) fails to account for discoveries where observational data is scarce. This position paper argues that while a modern Large Language Model could plausibly execute the deductive phase of proving theorems from established premises, it is structurally incapable of the abductive 'Jump' required to formulate those premises. We identify the translation of simulation into formal axioms as the critical bottleneck in artificial scientific invention, and propose that physically consistent, multimodal world models offer the necessary sensory grounding to bridge this divide.