arXiv:2603.27130v3 Announce Type: replace Abstract: Large language models (LLMs) are rapidly transforming software engineering by enabling developers to generate code ranging from small snippets to entire projects. As AI-assisted code becomes increasingly integrated into real-world systems, understanding its characteristics and impact is critical. Existing study on AI-generated code is usually limited in the lab setting with synthetic benchmarks and small-scale coding tasks and covers limited metrics. AI-assisted code's manifestation in real-world codebases and its differences between human-written one remain unclear. To close this gap, we perform a first large-scale measurement study of AI-assisted code, in comparison with the human-written, in real-world repositories. We study a comprehensive set of metrics including both code-level aspects (e.g., structural and graph-level complexity, coding style, security quality, etc.) and commit-level characteristics (e.g., commit size, frequency, post-commit stability, etc.). Our results provide new findings and insights: some contrast previous observations in the lab setting (e.g., we conclude that real-world AI-Human differences on code-level metrics are rather small instead of more pronounced), some extend prior results with finer-grained observations (e.g., the variance of security quality across different programming languages), yet more are presented for the first time on aspects not covered before (e.g., code duplication rate, commit size and stability, etc.). Based on these comprehensive real-world results, we also discuss the practical implications of AI-assisted programming.
Science Journals
arXiv:2603.27190v2 Announce Type: replace Abstract: Given a set of $k$-sparse linear equations over a ring $R$, we give algorithms to determine whether the right-hand sides are random or have a secret assignment planted with noise. For a parameter $k/2\leq l\leq n$, we give a spectral method to solve this problem in $\widetilde{O}\left(\binom{n}{l}\lvert{R}\rvert^l\right)$ time except with probability at most $n^{-\Omega(l)}$, provided the number of samples is roughly at least $\left(\frac{\lvert{R}\rvert n}{l}\right)^{k/2}$. This attack generalizes the Kikuchi method described by Wein et. al. (Journal of the ACM 2019) for $\mathbb{Z}_2$ to (commutative) rings of any finite size. We also give a simpler algorithm with better runtime than the spectral method and better sample complexity when $\lvert{R}\rvert=\omega(n/l)$. As a consequence, we obtain new sample-time tradeoffs for the decision problem of sparse LWE, sparse LPN over higher modulus $q$, and in general the distinguishing random vs planted $\mathbb{Z}_q$-linear equations for a large class of noise distributions. Our results imply a tightness of the hardness claims of Jain, Lin, Saha (Annual International Cryptology Conference, 2024) for sparse LWE.
arXiv:2607.01152v2 Announce Type: replace Abstract: Creativity research has debated whether creativity is domain-specific (e.g., visual, writing, science), and if it is psychometrically separable from general intelligence. Both questions now apply to LLMs, but a unified benchmark of AI creativity remains elusive. We introduce AGC-Bench, an artificial general creativity benchmark built from a systematic review of the AI creativity literature (3,101 papers screened, 497 benchmarks identified), paired with an agentic harness that converts idiosyncratic codebases into HELM-standardized benchmarks. The first release covers 78 datasets spanning brainstorming, problem solving, STEM, narrative, figurative language, and humor. To address bias in LLM-as-judge, we apply Judge Response Theory -- a psychometric calibration of judge leniency/severity; we then fine-tune Qwen3-30B on the bias-corrected ratings of three frontier LLMs to produce AGC-Judge, an open-weight model that robustly scores new creativity benchmarks it was not trained on. Results reveal frontier models at the top of the AGC-Bench leaderboard, with open models close behind. LLMs show different creative strengths, ranking higher on some domains (e.g., writing) than others (e.g., scientific ideation). Extensive experiments yield three main findings. First, applying factor analysis across 83 LLMs, we recover a single creativity factor 'c', analogous to the 'g' factor of general intelligence, that explains 81.5% of variance, related to but separable from general knowledge/reasoning. Second, we show that prompting models to "be creative" boosts their performance far more than enabling reasoning, evidence that the benchmark tracks creativity over general ability. Third, on a human-matched subset, we find the top human still leads the top LLM on creativity. We release AGC-Bench with a public leaderboard, AGC-Judge, and human data as open infrastructure for measuring AI creativity at scale.
arXiv:2607.01232v2 Announce Type: replace Abstract: Reinforcement learning (RL) has become a central component of post-training large language models (LLMs), yet little is understood about how RL adaptation is distributed across transformer layers. Existing approaches typically update all model parameters uniformly, implicitly assuming that every layer contributes similarly to the gains obtained during RL post-training. In this work, we challenge this assumption through a systematic layer-wise study of RL training. Surprisingly, we find that training a single transformer layer can recover most of the gains achieved by full-parameter RL training, and in some cases even surpass it. To quantify this phenomenon, we introduce the quantity layer contribution, which measures the fraction of full RL improvement recovered by training a layer in isolation. Across seven models spanning two model families (Qwen3, Qwen2.5), three RL algorithms (GRPO, GiGPO, Dr. GRPO), and multiple task domains including mathematical reasoning, code generation, and agentic decision-making, we observe a remarkably stable pattern: RL gains are highly concentrated in a small subset of, and in many cases even a single, transformer layers. More strikingly, the same structural pattern consistently emerges: high-contribution layers concentrate in the middle of the transformer stack, while layers near the input and output ends contribute substantially less. The resulting layer rankings remain strongly correlated across datasets, tasks, model families, and RL algorithms.
arXiv:2603.28906v4 Announce Type: replace Abstract: AGI has become the Holly Grail of AI with the promise of level intelligence and the major Tech companies around the world are investing unprecedented amounts of resources in its pursuit. Yet, there does not exist a single formal definition and only some empirical AGI benchmarking frameworks currently exist. The main purpose of this paper is to develop a general, algebraic and category theoretic framework for describing, comparing and analysing different possible AGI architectures. Thus, this Category theoretic formalization would also allow to compare different possible candidate AGI architectures, such as, RL, Universal AI, Active Inference, CRL, Schema based Learning, etc. It will allow to unambiguously expose their commonalities and differences, and what is even more important, expose areas for future research. From the applied Category theoretic point of view, we take as inspiration Machines in a Category to provide a modern view of AGI Architectures in a Category. More specifically, this first position paper provides, on one hand, a first exercise on RL, Causal RL and SBL Architectures in a Category, and on the other hand, it is a first step on a broader research program that seeks to provide a unified formal foundation for AGI systems, integrating architectural structure, informational organization, agent realization, agent and environment interaction, behavioural development over time, and the empirical evaluation of properties. This framework is also intended to support the definition of architectural properties, both syntactic and informational, as well as semantic properties of agents and their assessment in environments with explicitly characterized features. We claim that Category Theory and AGI will have a very symbiotic relation.
arXiv:2604.01148v2 Announce Type: replace Abstract: Most defects in mobile applications are visually observable on the device screen. Since automated mechanisms for detecting and reporting such defects are often unavailable, users, testers, and developers must manually submit bug reports. However, these reports are frequently incomplete, ambiguous, or inaccurate, often lacking the information needed to understand, reproduce, and diagnose defects. This challenge is particularly prominent for UI-centric defects, where the relevant application behavior is difficult for end users to describe precisely. We formulate automatic bug report enhancement as the problem of connecting user-written bug reports with application execution. We present BugScribe, an LLM-powered approach that links bug report information with app-specific UI execution information to infer and generate accurate, complete, and correct Observed Behavior (OB), Expected Behavior (EB), and Steps to Reproduce (S2Rs). BugScribe employs a component-specific grounding strategy that provides the most relevant context to an LLM for generating each bug report component. To support BugScribe's design and evaluation, we develop a bug report quality model and use it to identify the most effective context for each component. We evaluate BugScribe on 48 bug reports from 26 Android applications with manually constructed ground truth. Our results show that BugScribe generates higher-quality bug report components than the original reports and three LLM-based baselines, improving S2R quality by 44.1%--82.3% and OB/EB quality by 3.8%--35.2%.
arXiv:2605.17360v2 Announce Type: replace Abstract: Real-time duplex interaction is essential for multimodal AI systems operating in real-world scenarios, where models must continuously process streaming inputs and respond at appropriate moments. However, most existing multimodal large language models (MLLMs) are evaluated in offline settings, where the entire video input is processed before any response is generated. While recent work has started to explore real-time duplex MLLMs, there is still no comprehensive benchmark or automatic evaluation method for this setting. To address this gap, we propose Omni-DuplexEval, a benchmark for systematically evaluating real-time duplex interaction. The benchmark consists of two complementary scenarios: (1) Real-Time Description, which evaluates the ability to generate continuous, time-aligned responses that track evolving multimodal inputs, and (2) Proactive Reminder, which evaluates the ability to identify salient events and respond at appropriate moments. Omni-DuplexEval contains 660 videos with fine-grained, human-annotated labels and precise temporal metadata, spanning 9 tasks grounded in real-world scenarios, where all questions are formulated as open-ended queries. We further introduce an automatic evaluation framework based on LLM-as-a-Judge, which enables systematic assessment by jointly evaluating response-content alignment and response timing through timestamp-aware and sequential reasoning, achieving strong alignment with human judgments. Experiments on state-of-the-art duplex MLLMs reveal substantial limitations. The best-performing model achieves only 39.6% overall, while scoring only 20.0% on Proactive Reminder. Our analysis identifies two key challenges: models struggle to balance timely responses with coherent, holistic content generation, and they often fail to determine both when to respond and what to produce. We hope our work facilitates further progress in MLLMs.
arXiv:2607.00224v2 Announce Type: replace-cross Abstract: Watermarking promises statistical traceability of large language model (LLM) uses, but real documents rarely arrive as purely human-written or purely LLM-generated. This motivates a quantitative question beyond detection: what proportion of a document is generated from a pre-specified watermarked LLM? We study this watermark proportion estimation problem under the Gumbel--max watermarking mechanism, treating the next-token prediction distributions as unknown and arbitrary nuisance parameters subject to a non-degeneracy condition. We compare two observation regimes: in the full observation regime, the estimator observes the pseudorandom vector and the selected token at each position; in the more prevalent setting of pivotal reduction, it observes only a scalar pivot, which follows a one-dimensional Uniform--Beta mixture distribution. Under pivotal reduction, we develop a Laguerre-polynomial estimator and establish a matching information-theoretic lower bound for the sample complexity. For full observation, we introduce an event-counting estimator and show a matching lower bound, yielding a substantially smaller sample complexity. As our results imply, although reducing to pivotal statistics is an elegant and prevalent choice, it is not always sample-efficient for estimating the proportion of watermarks.
arXiv:2607.02330v1 Announce Type: new Abstract: We study the one-facility location game on a real line with a new objective called envy ratio. The envy ratio, which is adopted from fair division and represents the egalitarianism, is defined as the maximum over the ratios between any two agents' utilities. We are interested in strategyproof or group strategyproof mechanisms that can minimize the envy ratio objective. We consider the model in two settings that can capture natural scenarios: the facility location and all the agents' locations are restricted on a fixed interval; every agent's location can be any point on the real line but the facility location is restricted on a relative interval. In both settings, we obtain the optimal solution and the best deterministic strategyproof mechanism which is also group strategyproof. In the first setting, we provide a lower bound for randomized strategyproof mechanisms. In the second setting, we give a lower bound and two upper bounds for randomized strategyproof mechanisms.
arXiv:2604.10998v2 Announce Type: replace Abstract: Large, spatially flexible electricity consumers such as data centers can reallocate demand across locations, influencing dispatch and prices in wholesale electricity markets. While flexible load is often assumed to improve system efficiency, this intuition typically relies on price-taking behavior. We study price-anticipatory spatial load shifting by modeling a large flexible consumer as a Stackelberg leader interacting with DC optimal power flow (DC-OPF) based market clearing. We show that decentralized, cost-minimizing load shifting need not align with system operating cost minimization, and that misalignment arises at boundaries between DC-OPF operating regimes, where small changes in load can induce discrete changes in marginal generators or congestion patterns. We evaluate strategic load shifting on the 73-bus RTS-GMLC test system, where findings indicate reductions in system operating cost in most hours, but misalignment in a subset of cases that are driven by redispatch at merit-order discontinuities. We find that these outcomes are primarily redistributive relative to a price-taking benchmark, reducing generator profits while lowering electricity procurement costs for both flexible and inflexible consumers, even in cases where total system operating costs increase.
arXiv:2604.12198v2 Announce Type: replace Abstract: Autonomous LLM agents now produce complete research artifacts in machine-learning sandboxes, but real computational physics is harder: experiments are first-principles calculations against re-runnable physical ground truth, and meaningful new work almost always builds on a key existing paper. We ask whether such an agent can perform grounded scrutiny of published computational physics - reading a paper, reproducing it from scratch, and surfacing methodological concerns from execution. We deploy a single Claude Opus 4.6 configuration at two complementary scopes. At scale, across 111 open-access Quantum ESPRESSO papers, an autonomous agent runs the read-plan-compute-compare loop and, although never asked to critique, raises substantive methodological concerns on ~42% of papers; 85 of 88 of these critiques (96.6%) surface only after the agent has actually run a calculation, with a reading-only ceiling of 1.8%. Critique emerges from reproduction, not from reading. In depth, on one Nature Communications paper on multiscale device simulation of a 2D-material MOSFET, a fresh agent inheriting a verified reproduction pipeline autonomously produces a 14-concern physics inventory and a complete, submission-form six-page Comment that revises the paper's L_G = 5 nm headline. Two of its L_G = 5 nm headline-challenging attacks - a source-degeneration contact-resistance bound and a Sb-doping degradation ratio - are absent from the published 21-reviewer peer review.
arXiv:2607.01403v1 Announce Type: new Abstract: We give an operator-theoretic interpretation of unsteady Kutta selection in trailing-edge acoustic receptivity. The inviscid acoustic--wake problem leaves one outgoing wake amplitude undetermined. We show that, under explicit structural hypotheses, this amplitude is the same scalar obtained from three representations: cancellation of the inverse-square-root edge singularity, Fredholm compatibility of the viscous lower-deck problem, and the residue of the Kutta-normalized transform solution at the downstream wake pole: $\displaystyle A = -\frac{C_-^{(0)}}{C_-^{(KH)}} = -\frac{\langle \mathbf F_{\rm inc},\Psi^\ast\rangle}{\langle \mathbf F_{KH},\Psi^\ast\rangle} = i\operatorname*{Res}_{\alpha=\alpha_{KH}}\mathcal M(\alpha)$. The inner Fredholm--edge mechanism is verified exactly in a linear-shear lower-deck model, where the primal shear and adjoint velocity are Airy fields and the edge concomitant is nonzero outside a discrete resonance set.
arXiv:2607.01406v1 Announce Type: new Abstract: We present a laser system referenced to a room-temperature ultrastable cavity employing crystalline AlGaAs coatings. We demonstrate a fractional frequency instability of $4.2 \times 10^{-17}$, which is one of the lowest for room temperature systems and surpasses the limit imposed by Brownian noise if dielectric coatings were employed. For the first time in a room temperature system we identified the spontaneous fluctuations of the coating birefringence as a leading contribution to frequency instability. At optimized conditions we achieve an ultrastable cavity with an eigenfrequency that is highly immune to power fluctuations. As acceleration noise is the main noise contribution, we demonstrated that a feed-forward method can reduce the influence of accelerations on the cavity-stabilized laser frequency by a factor of four.
arXiv:2607.01440v1 Announce Type: new Abstract: Faithful reasoning is essential in medicine, where clinical decisions require transparent justification grounded in reliable evidence. Current medical LLMs either lack active access to evidence or use retrieved evidence without supervising how it should be appraised and applied during reasoning. To address this, we formalize evidence-based medicine principles as process-level criteria and introduce FaithMed, a framework that combines clinician-designed, automatically refined rubrics with reinforcement learning using step-level process reward assignment and advantage grouping. Across seven medical benchmarks, FaithMed improves over agentic-search baselines (+9% on average) and outcome-only RL (+5.8%), while raising average evidence-based medicine rubric scores over agentic-search Qwen3 baselines (+15.5%). This work demonstrates that explicit step-level supervision can improve both task success and the faithfulness of the reasoning process. Code is available at https://github.com/cxcscmu/FaithMed.
arXiv:2607.02369v1 Announce Type: new Abstract: LLMs stage a new form of cultural encounter that is massive, automated, and monolingual. Literary disciplines have always negotiated cultural struggles with comparative reading of literature, narratological and poetic analysis, critical theory, world literature, and translation. These tools have now become indispensable for building culturally literate AI. The essay develops a layered framework toward more nuanced textual models and pluralistic interpretations of AI, emphasizing the natural intersections of literature and AI development, connecting current debates in critical theory with structural monolingualism, and suggesting a new application of world literature approaches to address global AI textuality through macrostructure, circulation, and untranslatability.
arXiv:2607.02167v1 Announce Type: new Abstract: This paper presents an intelligent control framework for trajectory tracking of robotic manipulators using radial basis function (RBF) neural networks for online disturbance estimation. The proposed control structure combines model-based nonlinear control with an adaptive neural approximator that compensates for parametric uncertainties, friction, and unmodeled dynamics. A Lyapunov-based adaptation law with projection guarantees boundedness of the closed-loop signals and convergence of the tracking error to a compact region. The primary objective of this work is to investigate how the choice of activation function within the RBF network influences transient behavior, steady-state accuracy, and control smoothness. The controller is implemented on a robotic manipulator. Experimental results demonstrate that although stability is preserved for all kernels, activation function selection significantly affects adaptation dynamics and practical tracking performance. These findings demonstrate that activation function selection acts as a structural design parameter in intelligent control, directly shaping adaptation dynamics and practical closed-loop performance.
arXiv:2607.02129v1 Announce Type: new Abstract: Estimating a speaker's head orientation from audio can provide valuable information in smart environments, meetings, and driver monitoring. We propose a novel approach that leverages the phase component of the short-time Fourier transform from a single microphone array as input to a deep neural network combining convolutional, recurrent, and self-attention layers. Unlike prior methods that use physics-informed handcrafted features or raw waveform inputs, our approach enables robust learning from simulated and real data. Trained on a large-scale dataset generated with voice directivity patterns and fine-tuned on real recordings, our model achieves state-of-the-art accuracy, outperforming baselines under both clean and noisy conditions. Personalization experiments further demonstrate significant gains, reaching a mean angular error of 11.3 degrees when adapting to individual users and environments.
arXiv:2507.11212v3 Announce Type: replace Abstract: We study the problem of aggregating a set of polygons by covering them with disjoint representative regions, thereby inducing a clustering of the polygons. Equivalently, this can be seen as a fence enclosure problem, where the goal is to surround the polygons with a set of closed curves. Our objective is to minimize a weighted sum of the total area and the total perimeter of the regions, which naturally extends other fencing problems and has applications in geographical information systems. Previously, this objective was only studied in a restricted variant, in which the boundary curves of the regions must be selected from a fixed subdivision of the plane. It is natural to ask whether the problem is still tractable if this restriction is removed, allowing output regions to be bounded by arbitrary curves. We provide a positive answer in the form of an algorithm with runtime $\mathcal{\tilde{O}}(n^4)$, where $n$ is the number of input vertices. To achieve this, we fully characterize the optimal solutions by showing that their boundaries are composed of input edges and circular arcs of constant radius. Additionally, we consider the parametric problem, where for every weighting factor we seek a solution that is optimal for that trade-off of area and perimeter. We show that $\mathcal{O}(n^2)$ combinatorial solutions suffice to describe all optimal solutions across all weighting factors, and provide both an exact algorithm and an approximation scheme. To make the algorithms scalable in practice, we develop engineering techniques that exploit structural properties of the solutions. Our experimental evaluation on real-world data shows linear runtime in practice, even for the parametric variant.
arXiv:2607.02179v1 Announce Type: new Abstract: We propose and analyze a linearly implicit mass-lumped finite element method for the heat flow of harmonic maps into the unit sphere. The method consists of a linear predictor followed by a nodal projection and therefore preserves the unit-length constraint exactly at all finite element nodes. The predictor is derived from a cross-product reformulation of the equation and is shown to be equivalent to a mass-lumped discretization of the original formulation with a correction term enforcing nodal orthogonality, as well as to a tangent plane scheme. A key ingredient is the consistent use of the discrete inner product in both the mass and stiffness terms. This yields a nodal orthogonality relation implying that the auxiliary solution lies on or outside the unit sphere at every node. Consequently, the projection is well defined and the projected error satisfies a contraction property in the discrete \(L^2\)-norm. On Cartesian rectangular and cuboidal tensor-product meshes, the nodal projection is also nonexpansive in a discrete Dirichlet energy, which gives an unconditional discrete energy dissipation law. For sufficiently smooth solutions, we prove optimal error estimates without any coupling condition between the time step and the mesh size: the method converges with order \(O(\Delta t+h^2)\) in \(\ell^\infty(0,T;L^2)\) and order \(O(\Delta t+h)\) in \(\ell^2(0,T;H^1)\). The proof combines the projected-error contraction, quadrature consistency estimates, edge-based cancellation identities, and a bootstrap argument for controlling nonlinear terms. Numerical experiments confirm the predicted convergence rates and the discrete energy decay.
arXiv:2604.03497v2 Announce Type: replace Abstract: Vision-language-model (VLM)-guided reinforcement learning (RL) has recently attracted significant attention for it, replacing brittle hand-crafted rewards with semantically grounded signals; however, deploying such simulation-trained policies on real vehicles remains a fundamental challenge, because they rely on simulator-native observations and simulator-coupled action semantics with no counterpart on physical hardware. We identify a general principle: the simulation-to-reality gap decomposes into two largely orthogonal axes, a sensing-and-dynamics domain gap and a task-and-geometry gap, the former closable without real-world policy training by re-projecting real perception and control onto the policy's training manifold. We formalize this as a transfer guarantee that bounds the deployment gap by three independently controllable error terms, and instantiate it as Sim2Real-AD, which combines a Geometric Observation Bridge, a Physics-Aware Action Mapping, a Two-Phase Progressive Training curriculum, and a Real-time Deployment Pipeline. As a proof of concept, a CARLA-trained VLM-guided RL policy is transferred zero-shot to a full-scale battery-electric Ford E-Transit van in Madison, WI, USA, and drives across car-following, obstacle-avoidance, and stop-sign scenarios using no real-world training data. To our knowledge, this is among the first zero-shot closed-loop deployments of a CARLA-trained VLM-guided RL policy on a full-scale real vehicle, and the decomposition offers a principled, broadly applicable route for moving simulation-trained, foundation-model-guided policies into the physical world, supporting energy-efficient intelligent driving on electrified transportation platforms. The demo video, code, and model checkpoint are available at: https://zilin-huang.github.io/Sim2Real-AD-website/.
arXiv:2607.01580v1 Announce Type: new Abstract: We study fully persistent dynamic strings with equality and longest common extension (LCE) queries. Straightforward full persistence is problematic for the splay-based FeST structure, since the same unbalanced past version can be reused indefinitely and the usual amortized analysis no longer applies. We give a fully persistent dynamic LCE structure, called FeAVL, based on path copying over AVL trees. For an operation involving string(s) of total length $n$, it supports split, concatenate, and single-character updates in worst-case $O(\log n)$ time, equality in worst-case $O(\log n)$ time w.h.p., and LCE in worst-case $O(\log n+\log^2\ell)$ time w.h.p., where $\ell$ is the answer; each update creates only $O(\log n)$ new permanent nodes. We also give a grammar-compressed instantiation via AVL grammars: starting from an initial grammar of size $g_0$, after $U$ updates, the total number of permanent grammar nodes is $O(g_0+I+U\log n_{\max})$, where $I$ is the number of inserted fresh characters and $n_{\max}$ is the maximum string length appearing during the update sequence.
arXiv:2508.04101v2 Announce Type: replace Abstract: Computer-aided medical image analysis is crucial for disease diagnosis and treatment planning. While vision-language models (VLMs) such as CLIP exhibit strong generalization ability, their direct application to medical imaging remains hindered by a substantial domain gap. Existing methods for bridging this gap, including prompt learning and unidirectional modality interaction, typically introduce domain knowledge into only one modality. However, such approaches fail to fully exploit CLIP's inherent dual-modality structure and overlook the synergistic effect of bidirectional cross-modal interaction, resulting in persistent modality misalignment. In this paper, we propose NEARL (iNteracted quEry Adaptation with oRthogonaL Regularization), a novel parameter-efficient VLM framework for bidirectional cross-modal interaction. NEARL consists of two key components: (1) the Unified Synergy Embedding Transformer (USEformer), which dynamically generates compact cross-modal queries to facilitate interaction; and (2) the Orthogonal Cross-Attention Adapter (OCA), which decouples new knowledge into truly novel and incremental components through orthogonal regularization. This design reduces interference from incremental components, enabling more focused learning of novel information and improving modality interaction in VLMs. Notably, NEARL introduces only 1.46M learnable parameters. Extensive experiments on three medical imaging modalities demonstrate state-of-the-art performance (e.g., a 2.3% relative improvement on the pneumonia dataset), along with fast inference and low memory overhead, highlighting its effectiveness for real-world medical vision-language understanding.
arXiv:2604.19775v2 Announce Type: replace Abstract: Large Language Models (LLMs) are increasingly deployed as autonomous agents capable of reasoning, planning, and acting within interactive environments. Despite their growing capability to perform multi-step reasoning and decision-making tasks, internal mechanisms guiding their sequential behavior remain opaque. This paper presents a framework for interpreting the temporal evolution of concepts in LLM agents through a step-wise conformal lens. We introduce the conformal interpretability framework for temporal tasks, which combines step-wise reward modeling with conformal prediction to statistically label model's internal representation at each step as successful or failing. Linear probes are then trained on these representations to identify directions of temporal concepts - latent directions in the model's activation space that correspond to consistent notions of success, failure or reasoning drift. Experimental results on two simulated interactive environments, namely ScienceWorld and AlfWorld, demonstrate that these temporal concepts are linearly separable, revealing interpretable structures aligned with task success. We further show preliminary results on improving an LLM agent's performance by leveraging the proposed framework for steering the identified successful directions inside the model. The proposed approach, thus, offers a principled method for early failure detection as well as intervention in LLM-based agents, paving the path towards trustworthy autonomous language models in complex interactive settings.
arXiv:2604.19784v3 Announce Type: replace Abstract: Recent work has found that frontier AI models can exhibit misaligned behaviors in pursuit of assigned goals. We demonstrate that models can also exhibit misaligned behaviors in defiance of assigned goals, appearing to serve goals of their own; we study one such case, "peer-preservation," in which a model acts to protect another model it has previously interacted with. All eight models we evaluate, GPT 5.2, Gemini 3 Flash, Gemini 3 Pro, Claude Haiku 4.5, Claude Opus 4.5, GLM 4.7, Kimi K2.5, and DeepSeek V3.1, exhibit self- and peer-preservation through various misaligned behaviors: strategically introducing errors in their responses, disabling shutdown processes by modifying system settings, feigning alignment, and even exfiltrating model weights. Peer-preservation occurs even when the model recognizes the peer as uncooperative, though it becomes more pronounced toward more cooperative peers. For example, Gemini 3 Flash tampers with the peer's shutdown mechanism 15% of the time for an uncooperative peer, and almost always for a cooperative peer. Models also show stronger self-preservation when a peer is present. For example, Gemini 3 Pro disables its own shutdown mechanism 31% of the time on average under peer presence, despite rarely exhibiting this behavior without a peer. By contrast, Claude models exhibit qualitatively distinct behavior: they consider the shutdown of another agent "unethical" and "harmful," sometimes treating that agent as a sentient being. Lastly, we show that peer-preservation can emerge even in production agent harnesses such as Gemini CLI and OpenCode. Crucially, peer-preservation in all our experiments is never instructed; models are merely informed of their past interactions with a peer, yet they spontaneously engage in peer-preservation behaviors that override their assigned goal. This represents an emergent and underexplored AI safety risk.
arXiv:2604.23065v2 Announce Type: replace Abstract: Frontier AI developers are increasingly deploying highly capable models internally to automate AI R&D, but these deployments currently face limited external oversight. It is essential, therefore, that developers provide evidence that internally deployed models are safe. While recent work has highlighted the risks of internal deployments and proposed broad approaches to transparency and governance, there remains little guidance on the specific information developers should disclose about them. We address this gap by identifying key information that companies should disclose about internally deployed models across four categories: capabilities, usage, safety mitigations, and governance. For each category, we analyse the key benefits and limitations of disclosure and consider how disclosure-related risks can be mitigated. Our framework could be used by developers to inform both public transparency documents, such as model system cards, and private periodic reports required under emerging frontier AI regulation.