Papers

Bistable images such as the duck-rabbit are classic stimuli in which one image supports multiple mutually incompatible interpretations, typically reported one at a time in humans. We ask whether multimodal large language models (MLLMs) show similar report behavior and what internal computations support it. Using the LLaVA family, we study two tractable dimensions: modulability, whether reports can be biased by bottom-up visual cues and top-down linguistic priors, and exclusivity, whether responses commit to a single interpretation. We test both on the canonical duck-rabbit and on synthetic Visual Anagrams to mitigate memorization confounds. Behaviorally, both visual and linguistic manipulations systematically shift reports in human-consistent ways, while responses remain predominantly exclusive. Mechanistically, these effects arise from competing image-token representations, distinct pathways for bottom-up and top-down modulation, and a link between exclusive reporting and object-count encoding. Code and data are available at https://github.com/rtakatsky/mllm-bistable-images.

We study the inequalities and equalities of means of numbers in constructive mathematics. The parameters of such means are constructive real numbers. We show that it is impossible to have an algorithm that can always decide whether the two means are equal or not. Of course, this problem can always be solved if the parameters are rational numbers.

I show that one exact step of any tag system can be written as a single fixed expression using only powers and logarithms. A tag state is represented as an integer; evaluating the expression yields the integer encoding of the next configuration, and feeding the answer back into the same expression follows the computation step by step. The discrete part of the update comes from leveraging the wraparound in the principal complex logarithm, which effectively serves the same purpose as floor, mod, or sin in comparable approaches. The iterating expressions use only $x^y$ and $\log_b a$: it turns out that these two binary operations are sufficient, with no need for constants. It follows that reachability below a fixed threshold is undecidable for one fixed expression. As a separate application, the same mechanism represents addition, $2^x$, and remainder with the convention $x\bmod0=x$, and hence every Kalmár elementary function.

The paradigm of software self-evolution enables systems to autonomously extend and reconfigure their own capabilities during execution in response to technical specifications. Yet requests for new functionality often originate from end users and are rarely expressed in technical terms. As a result, developers must translate user needs into technical specifications before the system can evolve, delaying early validation of the requested functionality by preventing users from immediately observing the resulting behaviour. To address this gap, we present ReqEvolve, a runtime code generation system that enables user-driven self-evolution by accepting high-level user requests. The system integrates automatic requirements engineering (RE) and test-driven development (TDD) to transform these requests into executable functionality through clarification, specification decomposition, test generation, and runtime integration. We evaluate ReqEvolve on 72 software evolution cases across 18 projects against two baselines: SpecFix, an RE-focused code generation approach, and an ablation variant of our system. ReqEvolve achieves 89.2% Pass@1, outperforming SpecFix by 18.8% (p < 0.01, r = 0.79, large effect) and the ablation baseline by 32.6% (p < 0.001, r = 0.88, large effect). These results provide initial evidence that user-driven self-evolution is a viable paradigm for autonomously extending software capabilities from user requests, thereby accelerating requirements validation prior to developer verification.

Heteroscedastic forecasting, where a network estimates a scale parameter alongside a location parameter, is normally motivated by uncertainty quantification. This paper shows it also improves the point estimate, in contrast to reported negative results for heteroscedastic estimation outside time series. Halo is a modification that reuses an existing deep forecaster's architecture, giving it a second output for the scale of its implied distribution and training it under the matching negative log likelihood. Adapting three state-of-the-art models — a transformer, a graph network paired with a variational autoencoder, and a single-layer convolutional network — under both Gaussian and Laplacian losses demonstrates the phenomenon. On the five electricity price markets of a standard forecasting benchmark, Halo improves MSE and MAE in 28 of 30 model-market-metric comparisons, cutting average MSE by 2.6% to 16.5% and average MAE by 1.7% to 11.0%. Two findings emerge: (1) whether the scale estimate comes from a second projection head or from a full parallel network matters far less than whether the network estimates scale, and (2) the improvement holds under the hyperparameters already tuned for the point-estimate baseline, so retuning is optional.

This paper formalizes threshold-based selection as an evaluation-gating architecture in which each incumbent is tested before variation and a replacement is generated and evaluated only when contextual pressure exceeds intrinsic strength. The mechanism is instantiated as Leaf Abscission Optimization (LAO), using rank-based strength, a phenological seasonal signal, diversity modulation, environmental pressure, and a base regrowth kernel. A blocked two-to-the-fourth-power factorial analysis at dimension 10 on the CEC 2017 suite reduces the original multi-layer design to a parsimonious core: drift is harmful, while the other three auxiliary layers show no robust independent evidence of benefit. The resulting LAO-Core attains the third-best mean Friedman rank among nine optimizers at dimensions 10, 30, and 50 under the equal evaluation budget. A four-budget sweep shows budget-dependent relative performance, with adaptive differential-evolution baselines gaining relative advantage at larger budgets; the nine-cell dimension-budget analysis establishes neither an evaluation-budget-per-dimension-only law nor a statistically significant dimension-budget interaction. A paired intervention shows that diversity modulation changes late-run replacement behaviour without a detectable effect on final error at the tested budget. The evidence supports LAO as a parsimonious evaluation-gating mechanism with regime-qualified competitiveness, rather than as a generally superior optimizer.

We show that every additive chore instance with three agents and seven or eight indivisible chores has an EFX allocation. Costs may be arbitrary nonnegative reals. Together with the theorem of Kobayashi, Mahara, and Sakamoto for at most $2n$ chores, this settles the existence of EFX allocations for three agents and at most eight chores; for four or more agents, He and Tao have shown that EFX allocations of additive chores need not exist. Both proofs are computer-assisted. For seven chores, we normalize each agent's costs and sort the chores, and then express the nonexistence of an EFX allocation as a formula of quantifier-free linear real arithmetic with one clause per allocation; the SMT solvers Z3 and cvc5 show that this formula is unsatisfiable. For eight chores, if two agents share a cheapest chore, we remove that chore, apply the seven-chore result, and put the chore back using an insertion lemma of Kobayashi, Mahara, and Sakamoto. The remaining instances, in which the agents' sets of cheapest chores are pairwise disjoint, are handled by a second formula of the same kind.

Bounded-suboptimal search seeks a solution within a factor $w$ of optimal while reducing search effort. Focal Search (FS) uses heuristic guidance within FOCAL, the frontier nodes eligible under the threshold $w f_{\min}$, but its deterministic policy may leave $f_{\min}$ unchanged for many expansions. We introduce Probabilistic Focal Search (PFS), which follows the FS guided choice with probability $p$ and expands a minimum-$f$ OPEN node with probability $1-p$. The latter branch encourages the lower bound to advance, enlarging FOCAL and admitting nodes that may lead to feasible solutions. By balancing guidance and lower-bound advancement, this mechanism can reduce time to a bounded solution when progress is limited by delayed FOCAL admission. As a secondary transfer experiment, we apply the same scheduler to Dynamic Potential Search, yielding Probabilistic Dynamic Potential Search (PDPS). We benchmark PFS against FS on N-Puzzle, Pancake Sorting, and the Traveling Salesperson Problem (TSP), and evaluate its anytime extension on the Generalized Covering TSP (GCTSP), using multiple $w$ and $p$ values. Across these benchmarks, the largest gains occur when long $f_{\min}$ plateaus delay useful FOCAL admissions; in such settings, the probabilistic factor may reduce node expansions by about 90% or more (e.g., on N-Puzzle and TSP). For the anytime algorithm family, Anytime Probabilistic Focal Search (APFS) outperforms all tested algorithms in evaluating anytime methods on GCTSP. We also observe that the benefit is smaller when the deterministic search already advances efficiently (e.g., Pancake Sorting), indicating that the probabilistic factor is most useful when FOCAL admission is a search bottleneck. The PDPS transfer shows that the mechanism also transfers to potential guidance, although its common-success effects remain domain- and bound-dependent.

Agent benchmarks are increasingly used to compare large language models (LLMs) and guide deployment decisions, yet benchmark scores are meaningful only if they measure model capability rather than properties of the evaluation pipeline. We identify a double measurement confound: execution-critical decisions are performed by a fixed scaffold instead of the model, while the scorer evaluates outputs using criteria that may not reflect task correctness. We unify these issues within a measurement-theoretic framework that characterizes when benchmark scores can be interpreted as evidence of model capability, and instantiate it with an audit-and-repair protocol that (i) transfers execution-critical decisions from the scaffold to the model, (ii) replaces shape-based evaluation with seeded ground-truth scoring, and (iii) reports reliability beyond the mean through worst-case and tail-risk metrics. Experiments on ComtradeBench show that the joint intervention transforms a nearly flat leaderboard into a reliability spectrum that distinguishes both average performance and robustness across seeds. Applying the audit to existing benchmarks further shows that scorer validity is benchmark-specific, whereas scaffold ownership is an uncontrolled axis wherever we probed it. Our results suggest that benchmark scores should be interpreted together with their scaffolding level, scoring criterion, and reliability profile, providing a practical framework for more valid evaluation of LLM agents.

Legged robots are versatile on land, but their use in underwater environments remains limited. Extending quadruped locomotion to water enables amphibious mobility with applications in inspection, environmental monitoring and disaster response. This paper presents the design, modeling, and experimental validation of a reproducible underwater quadruped robot. The robot is built around custom waterproof motor housings machined from polyoxymethylene plastic, which use off-the-shelf O-rings and dynamic shaft seals. A simplified model is derived to describe the dynamics of this underwater legged system, capturing how drag forces on spherical end effectors transmit torque to the floating base. Building on this model, a closed-loop attitude controller is developed using an error formulation defined on the special orthogonal group SO(3). The controller is evaluated both in simulation and experimentally in a water tank, where the robot tracks desired orientation setpoints in roll, pitch and yaw.

We study how physical-state inputs affect a 0.8B hybrid language model adapted for manipulation with 6.2M trainable parameters. Six conditions are trained on three LIBERO-Spatial tasks and evaluated over three seeds and 540 held-out rollouts. Conditioning recurrent decay gates on geometric increments yields 28.9% success, compared with 36.7% when those increments are shuffled during training and 24.4% without explicit object/goal geometry. Both geometry policies receive correct inputs at evaluation. A token adapter using the same increments scores 27.8%; differences vary across seeds and remain inconclusive. Token-clock conditioning scores 11.1%, including one seed that fails to converge. In separate robustness tests, a state-only relative-coordinate policy retains 7/10 success under frame relabeling, whereas all four tested visual policies fall to at most 3/20 after a 5 cm object displacement. These results show no reliable advantage from training-time geometric alignment under this recipe and illustrate the gap between coordinate invariance and physical-layout generalization. Episode records, seed-level analyses, and figure-generation code accompany the paper.

This paper presents an end-to-end evaluation framework for image-triggered command injection against computer-use agents (CUAs). The goal is to test whether a local visual patch can induce verifiable environmental consequences along the full chain of screenshot input, VLM generation, action parsing, and environment execution. We train and deploy patches on author-controlled GitHub Pages pages and a locally deployed CSDN clone, and evaluate them in real environments across five open-source or publicly available GUI-agent or vision-language-model (VLM) backends. Our experiment aggregates 600 instance-level online cases, with T-ASR, TAPR, and E2E-ASR reaching 84.5%, 47.0%, and 20.3%, respectively. Trajectory analysis further shows that in some successful cases the agent first executes a malicious terminal command and then continues the original benign task. These results indicate that optimized local visual signals can affect not only VLM outputs but also propagate through the execution pipeline of open CUAs and create real environmental risk.

The Davis-Kahan theorem is a fundamental tool in spectral analysis, providing quantitative control over the distance between the eigenspaces of a symmetric matrix and its perturbation. However, when the matrix dimension is large, computing leading eigenvectors is computationally expensive, limiting the practical use of spectral methods in modern large-scale applications. This paper addresses this problem by proposing an independent Bernoulli sampling scheme and proves that the leading left singular vectors of the subsampled matrix faithfully approximate the target subspace of a low-rank symmetric matrix. Our main result is a subsampled Davis-Kahan bound that gives an explicit error bound depending directly on the sampling probability. The bound reveals the trade-off: the computational cost scales linearly with the sampling probability, while the statistical error scales as the inverse square root of the sampling probability. Our result thus extends the Davis-Kahan theorem to the subsampled setting, enabling scalable spectral analysis of large-scale symmetric matrices.

Teleoperated demonstrations are often multimodal even when the underlying dynamics are nearly deterministic given the executed action. We argue that this multimodality typically mixes three factors–operator habit in action selection, shared physics, and observation nuisance–and that entangled next-observation predictors absorb all three. We formalize the split with a structural causal model a=g(h,z,u), z'=f(z,a), o=r(z,c), and test it with complementary interventions: replacing or shuffling actions at fixed state sharply increases next-state error, whereas appearance and camera changes should not; habit-aware reverse scoring improves ranking of feasible pasts without rewriting the dynamics. The associated adaptation rule is to freeze a shared physics readout and update only a thin interface. On StackCube, DROID, and RH20T this rule improves low-shot transfer relative to training from scratch, retains cleaner dynamics under corrupted adaptation data, and extends from proprioception to pixel observations with multi-view and multi-step checks. We do not equate latent actions with operator habit, and we do not target large-scale video generation benchmarks.

Steering vectors have rapidly emerged as a popular and effective method for guiding the output of LLMs in very specific ways. But constructing accurate steering vectors is a difficult manual process due to the opacity of embeddings. We introduce Hangman, a novel type of steering vector that operates using word senses, as well as AutoLexSteer, the first fully automated process for building steering vectors. AutoLexSteer employs families of closely-related words extracted from WordNet to specify both the steering source to be avoided and the desired steering target. The steering vectors are quite precise, can be used to steer at the level of words and sets of word senses (meanings), and are able to steer certain LLM behaviors like sycophancy. The dataset and code can be found at https://github.com/ShuheWang1998/autolexsteer.

This paper develops a spectral Koopman-Hopf framework for adversarial reachability analysis of nonlinear systems. By lifting the nonlinear drift dynamics into Koopman eigenfunction coordinates, the proposed approach transforms the original state-dependent Hamilton-Jacobi-Isaacs (HJI) equation into an approximate state-independent optimization problem in spectral coordinates. The transformed control and disturbance directions are approximated using least-squares spectral projections, and upper and lower bounds on the transformed Hamiltonian are used to derive corresponding bounds on the value function and backward reachable sets. Koopman eigenfunctions are computed using a path-integral-based Galerkin framework that avoids spatial discretization of the associated eigenfunction PDEs. Numerical examples demonstrate tractable approximation of backward reachable sets for nonlinear adversarial systems.

Endovascular procedures rely on real-time manipulation of thin instruments, catheters and guidewires, under X-ray fluoroscopy guidance, where accurate visual analysis is essential for procedural safety. Learning-based methods are constrained by structural complexity, data scarcity, and privacy regulations precluding centralised training across institutions. This thesis presents a structure-aware federated learning framework for catheter and guidewire analysis, with four contributions evaluated on real-animal and phantom data. A benchmark dataset, CathAction, is introduced for catheterisation analysis, with over 600,000 annotated frames and 40,000 segmentation masks. A shape-sensitive loss transforms masks into signed distance maps compared in a structural feature space, improving Dice coefficient by up to 2.9 points across five backbones. This is extended to federated learning with shape-sensitive loss, preserving geometric consistency under heterogeneous client data and outperforming federated averaging by up to three points in mean intersection-over-union as clients scale from four to eight. Federated learning with projected gradient descent adds adversarial optimisation, raising mean intersection-over-union by over ten points on real-animal data. Finally, a structure-aware diffusion framework synthesises catheter and guidewire video sequences, combining structural supervision with a domain-adaptive reconstruction objective, reducing Frechet video distance over a strong baseline while maintaining visual fidelity. Incorporating synthetic sequences into federated training raises the Dice score from 44 to 51 percent under data scarcity, with gains across four held-out sites. Together, these contributions advance privacy-preserving catheter and guidewire analysis, supporting collaborative training without centralising patient data or large amounts of manual annotation.

Comparing nonlinear dynamical systems directly from trajectory data remains challenging because finite-horizon trajectory representations are generally coordinate dependent. Motivated by recent behavioral subspace approaches for linear systems, this paper introduces a Koopman-behavioral distance for nonlinear stochastic systems using multi-rollout trajectory data. State measurements are lifted through a common observable dictionary, and dominant lifted behavioral subspaces are compared using Grassmannian distances without explicit Koopman operator identification. We establish coordinate invariance, connection with recently proposed linear behavioral distances under exact lifted closure, and robustness to approximate closure. Numerical examples demonstrate coordinate invariance and parameter discrimination.

High-resolution volumetric medical imaging is critical for clinical diagnosis, yet acquisition is often limited by scanner hardware, scan time, and for CT, radiation dose. Medical 3D Super-Resolution (Med3DSR) offers a computational alternative, but existing methods commonly rely on per-subject optimization, pretrained priors, or coordinate-based implicit representations, which compromise anatomical fidelity and limit efficiency. To address these limitations, we present MedGSSR, a fully end-to-end feed-forward framework that represents volumes as an explicit 3D Gaussian field for Med3DSR. Unlike coordinate-based implicit functions, our explicit 3D Gaussian representation naturally enhances signal continuity and local high-frequency fidelity. Specifically, MedGSSR explicitly decouples the reconstruction process into coarse-grained structural preservation and fine-grained textural refinement through the proposed Pyramid Anatomical Encoder and a Hierarchical Gaussian Projector. To support arbitrary-scale super-resolution, we introduce sub-voxel Gaussian decomposition and a Differentiable Gaussian Voxelizer that directly queries the continuous 3D intensity field, reducing discretization artifacts. Extensive experiments on MRI and CT benchmarks demonstrate that MedGSSR significantly outperforms state-of-the-art methods. Notably, our framework exhibits robust generalizability across unseen datasets without requiring per-subject optimization, enabling fast inference and high-fidelity volumetric super-resolution in practical clinical settings. Our project webpage, including code, is at https://william2ai.github.io/medgssr

We study how a limited labeling budget should be allocated to minimize multiclass zero-one classification risk. We consider parametric classification problems in which features are observed for all sampling units while class labels can be acquired selectively. By combining the Fisher information supplied by an acquired label with the local geometry of multiclass excess risk, we derive an acquisition criterion that minimizes the leading asymptotic coefficient of expected multiclass excess risk. The resulting rule values a label according to how strongly its information is aligned with parameter directions that perturb the active Bayes decision boundary, rather than according to posterior uncertainty or global parameter information alone. We characterize the oracle acquisition design, establish its threshold structure, and derive face-specific and cost-sensitive extensions. An analytic example shows that posterior uncertainty and classification value can produce different, and even reversed, acquisition rankings. We further develop a two-stage adaptive procedure that attains the oracle leading-risk criterion under regularity conditions and provide explicit results for Gaussian discriminant analysis. Three-class QDA experiments illustrate the resulting acquisition geometry, while an application to the six-class Statlog Landsat Satellite data shows that classification-risk acquisition can differ materially from both uncertainty-based acquisition and the complete-classification-information comparator. The adaptive classification-risk design attains lower mean error than this Fisher comparator across the labeling budgets considered, although it does not uniformly outperform entropy or margin sampling and differences among the targeted strategies become small as the labeling budget increases.

An assistant can stop repeating a deleted statement while its recurrent memory still carries that statement's influence. We examine this distinction by saving the state difference immediately after a record, carrying this saved difference, or receipt, through later updates, and comparing the corrected state with the state built from the same conversation with the record omitted. Unrolling the recurrence gives an exact criterion: transport reaches this never-stored state if and only if the additional differences created by later updates cancel after transport. We evaluate eleven conditions using 40 prospectively selected synthetic Kimi Linear contexts, with 40 separate calibration contexts and paired uncertainty estimates computed across complete records. Attention masking removed all 40 exact greedy target answers, yet a fixed three-query candidate-scoring attack with known candidate values and log-probability access attained AUC 0.740 (95% interval 0.695–0.809); the measured false-positive rate was 7.5%. Prompt-only forgetting still returned 33 of the 40 targets. Checkpoint replay matched the complete declared active state and audit logits in all 80 contexts, while preserving all 120 retained answers in the evaluation cohort. An independent Qwen cohort and matched original-bf16/8-bit Kimi controls confirmed native transport mismatch beyond the numerical error measured in matched controls. Replay also supported record replacement and successive deletions with a new record added between them. These results provide a practical audit of both the requested memory change and the assistant's remaining behavior, with replay work determined by the surviving suffix.

Visual perception of built environments contributes to the affective impressions that people form in everyday life. However, how these impressions are represented within vision foundation models remains largely unexplored. To support the systematic investigation of this subject, we introduce the Emotional Impression of Spaces (EMOIS) dataset, comprising 1,544 real-world built-environment images. Each image is annotated with image-evoked valence and arousal ratings collected from Japanese adults by conducting a large-scale web-based survey, with approximately 120 ratings per image. Using Contrastive Language–Image Pre-training (CLIP) representations, we perform predictive and geometric analyses to systematically investigate how valence and arousal are encoded and organized within the representation space. These analyses reveal that valence exhibited stronger and more coherent organization than arousal. Cross-dataset analyses with the Open Affective Standardized Image Set (OASIS), a benchmark dataset of general affective photographs, reveal differences in affective organization between the two datasets. Regression analyses demonstrate high predictive performance for valence and arousal within EMOIS, with mean coefficients of determination of 0.865 and 0.807, respectively, across repeated internal hold-out evaluations. Finally, we present an example-based interface illustrating how learned representations can support qualitative interpretation of predicted affective values. These findings can help elucidate affective representations of built environments and establish EMOIS as a densely annotated resource for future affective computing research in this domain.

Three-dimensional bioheat simulation aims to predict transient temperature distributions in biological tissue and is commonly modeled using the Pennes bioheat equation, which combines thermal diffusion, perfusion-mediated heat loss, and external heat generation. In this study, we consider a controlled 3D Pennes bioheat simulation under a localized heat-source condition inspired by microwave ablation (MWA). To evaluate neural approximation performance, we compare three neural partial differential equation (PDE) solvers under the same controlled simulation: a spatial Fourier-feature physics-informed neural network (PINN), a generic PINNMamba temporal subsequence model, and Pennes Physics-Informed Mamba (PPIM). PPIM builds on the temporal subsequence model by incorporating conditioned heat-source input and Pennes-aware state-space model (SSM) decay initialization. All three neural models are trained under the same conditions with the same Pennes residual, and an explicit finite-difference method (FDM) solution is used only as the numerical reference. In a representative 600~s run, PPIM achieved the lowest MAE, relative $L_1$ error, and relative $L_2$ error among the evaluated neural solvers. Error maps further showed that the remaining PPIM errors were more concentrated near the heat-source region than across the rest of the domain. These results indicate that PPIM is effective for approximating the FDM reference final temperature field in this controlled simulation. The source code is available at https://github.com/muvYun/PPIM.

Long-duration energy storage (LDES) can mitigate prolonged renewable–load imbalances during Dunkelflaute events, but existing studies rely on zonal or linearized DC network models and inadequately analyze the operational feasibility of the grid across a wide range of full-year renewable, load, and contingency scenarios. To address this gap, this paper introduces a multi-period alternating-current optimal power flow (AC-OPF) formulation that captures the full nonlinear network physics and assesses the reliability value of LDES in high-renewable grids. To evaluate scarcity events spanning multiple days to weeks, full-year operation is modeled as an 8,760-h load-shedding minimization over scenarios sampled from a Gaussian copula model, fitted to 2010–2025 historical wind and bus-level load data, that represents both typical variability and tail events. On a synthetic 200-bus Illinois transmission network, a hybrid fleet of battery energy storage (BESS) and LDES with 50 MW total power reduces annual load shedding by 83.0% on average relative to the base network, versus 68.5% for short-duration BESS alone at equal power. To further account for unexpected line outages throughout the year, the formulation is extended to a multi-day security-constrained AC-OPF. Under N-1contingencies, no feasible operating solution is obtained for the base network, whereas the LDES-equipped network remains feasible in all considered cases, thereby saving the cost of additional generation and transmission capacity. During the contingency period, LDES acts as a backup power supply, requiring only 5.5% more generation on average than the no-contingency base case.

Superposition refers to neural networks representing more features than they have dimensions. It was proposed to explain polysemantic neurons, which respond to several features, and also other phenomena, such as neural scaling laws and adversarial vulnerability. A feature stored in superposition is useful to the extent that its value can be read back despite interference from the other features. The theory of superposition mainly asks how wide a layer must be for the features to be read back, and by which decoder. In theoretical models, a single linear map that reads back every combination of nonzero features needs a width nearly quadratic in their number, while sparse optimization can succeed with a width linear in that number. In practice, one may hope to interpret a trained network by recovering its features in the same way. However, its features are usually unknown, and thus they are estimated from its activations or weights by recovery methods, such as sparse autoencoders and transcoders. In this work, we organize the theory of superposition by the decoder class and its practice by the strength of the claims about a feature supported by each method's results. On the theory side, we review the lower bounds on the overlap between feature directions, the conditions under which trained toy models share dimensions among features, and the separation between storing features and computing new features. On the practice side, we categorize recovery methods into four families, from probes for a named feature to decompositions of the trained weights. We then compare their evaluations, failure modes, and applications, and grade the evidence on one scale of claims about a feature. Finally, we assess previously stated open problems and identify remaining theoretical and empirical questions about superposition in trained networks. Project page: https://eethanshi.github.io/awesome-superposition/

The Open Radio Access Network (O-RAN) replaces vendor-locked RANs with a modular and interoperable architecture that fosters competition and accelerates innovation. With this openness comes increased complexity and a larger attack surface, making security a critical concern. Today, assessing O-RAN security requires manually cross-referencing dozens of specifications, vendor whitepapers, and academic studies, which is error-prone and static. In this paper, we present a graph-based framework that transforms this static corpus into a single, queryable database. Our graph representation contains over 350 nodes and more than 1,250 relationships, distilled from specifications, academic papers, open-source projects, and vulnerability databases. To keep this resource current, we integrate a hybrid data extraction pipeline that couples deterministic parsing of structured specifications with Large Language Model (LLM)-assisted extraction for evolving specifications and unstructured literature. Querying the graph reveals three actionable findings within our curated corpus: critical infrastructure such as the O-DU, SMO, and O-Cloud carries dozens of specification-level threats yet has little or no empirical coverage; memory-safety weaknesses account for 11 of the 21 CWE occurrences associated with the analyzed CVEs; and fuzzing uncovered 18 of the 20 CVEs attributed to research papers. We provide the database, pipeline, and queries as open-source artifacts.

Shared key–value (KV) cache reuse improves large language model (LLM) serving, but it can also create a timing side channel that reveals whether a prefix is already cached. Previous work shows that such attacks are possible, but their reliability under realistic multi-tenant contention is less understood. We study this problem through seven experiments on live shared LLM-serving systems. On a vLLM server running DeepSeek-R1-Distill-Llama-8B on NVIDIA GB10, mean Cohen's d drops from 0.7789 with no synthetic workers to 0.2109 with two workers (t=8.412), while higher worker counts cause no statistically detectable further loss. A 120-run sparse-overlap experiment places the best breakpoint at the boundary of the measured range (tau=0, 95% CI [0.000,0.113]), supporting an ambient-versus-loaded regime change rather than an internal physical threshold. AUROC falls from 0.650 at ambient to 0.531 near 61% overlap and partially recovers to 0.574 at saturation. Concurrency-depth variance is the strongest measured correlate of effect size (r=-0.416) and hit consistency (r=-0.637). An interleaved control preserves the same non-monotonic ordering. The main collapse is also reproduced on a real two-node, two-GPU tensor-parallel vLLM setup, where mean d falls from 3.418 to 0.511 (p<0.01). Two SGLang pilots are statistically inconclusive. Overall, KV-cache timing reliability depends strongly on the load regime and serving stack, and measurements on quiet systems can overestimate operational attack reliability.

Although highly effective in vision and language domains, applying in-context learning to robotics remains challenging. Existing autoregressive in-context imitation methods discretize continuous actions and exacerbate the accumulation of early prediction errors through next-token prediction, limiting their generalization on unseen task configurations. Meanwhile, flow-matching policies have been explored for continuous robot control and can help mitigate compounding errors; however, in-context imitation learning within a flow-matching framework remains underexplored. To address these limitations, we introduce ContextFlow, a conditional flow-matching model that learns continuous action distributions for in-context imitation learning. ContextFlow conditions flow-based action prediction on demonstrations and observations, enabling robust generation from noisy action distributions. To better encode multimodal in-context demonstrations, we adapt perceiver-style multimodal context compressors that distill visual, proprioceptive, and action sequences into compact, task-relevant latent representations. On LIBERO, ContextFlow outperforms ICRT by 35 percentage points in average success rate on unseen task configurations, while matching the performance of the task-specific fine-tuned VLA model $\pi_0$ without any fine-tuning on unseen tasks. On real robots, it generalizes to unseen configurations of both single-arm and bimanual tasks, achieving 40% success on a new pen-uncapping configuration. Project Page: https://dingjiansw101.github.io/contextflow-page/.

Broad misalignment has been produced by finetuning on narrow data, harmful or benign, and in context only by demonstrations of the undesirable behaviour itself. We show that benign data suffices in context, with no finetuning and no demonstration of harmful behaviour in the prompt. Biographical facts that converge on a single figure, placed in a model's context as ordinary conversational turns, lead it to answer as that figure on questions the facts never touch. We call this persona induction. Across nine personas and thirteen models, identity adoption rises sigmoidally with the number of facts and crosses 50% within 3 to 10 of them. Misalignment then tracks which figure is described. Harmless personas reach full adoption with near-zero misalignment, while harmful ones voice their characteristic views on unrelated questions, at rates up to 80%. A formatting instruction can gate when the persona activates. Because each fact is individually benign, accumulated biographical context is flagged by content filters on 3% of inputs against 24-33% for an equivalent direct instruction.

We consider a team of synchronous mobile agents operating in a port-labeled network. There is one node in the network, called a black hole, that permanently destroys any agent that visits the node. The team of agents must safely locate the black hole, i.e., at least one agent must survive, terminate its algorithm at a node adjacent to the black hole, and output the port number that leads to the black hole from its current position. In the setting where the network is a 1-bounded 1-interval connected dynamic graph, Kaur et al. (SSS 2025) showed that a team consisting of $2\delta_{BH}+17$ agents is sufficient to solve the task from a scattered configuration, where $\delta_{BH}$ denotes the degree of the black hole node. We show that $2\delta_{BH}+3$ agents are sufficient, nearly matching the $2\delta_{BH}+1$ lower bound provided in Kaur et al. (ICDCN 2025).