Papers

In a deliberative poll, once submissions outnumber what anyone will read, some mechanism chooses which arguments each voter sees, acquiring much of the decision; practice delegates it to opaque learned rankers, so a voter cannot recompute or contest the exposure that shaped their vote. We ask whether it can be a published rule over publicly recomputable evidence with parameters held by the voter, treating legibility as an admissibility condition on usable mechanisms, not an objective traded against accuracy. We formalise a poll over bipolar justification sets, judging a slate by reason coverage, the order it arrives in, and captured endorsement mass; we give seven checkable criteria for a civic recommender and a rule meeting them: a one-hop reversed endorsement flow parameterised by a relation-weight function. An agentic simulator records every slate at every vote, over about 17,000 seed-paired runs. Served slates fall 0.035 short of a label-reading ceiling upper-bounding every selection procedure, opaque ones included: any unconstrained ranker's advantage is bounded and small. On coverage alone, with non-degenerate authoring, the rule is indistinguishable from a random slate, a null due to an order-blind, charity-blind instrument; on the other two it leads at every prefix by a margin widening with adversarial pressure and dominates on mass by a factor of 3.3. Once a realistic fraction of submissions carries no reasons, the coverage margin returns and grows. Label-homogeneous flooding collapses completeness from 0.81 to 0.34 under a flat weight policy, only to 0.44 under author-count normalisation, making the weight function a security control worth 10% of completeness. The choice between ranking arms is a position on a coverage-versus-mass frontier, not a fact, the kind of choice only a legible rule can hand to the person it affects. It maps onto an open-source peer-to-peer platform.

Visual grounding is typically evaluated as a one-shot mapping from an informative referring expression to a visual target. This formulation misses a central property of real-world reference: initial referring expressions are often incomplete or ambiguous, requiring participants to establish shared understanding through interaction. We introduce a controlled evaluation framework for interactive visual grounding in large vision-language models (LVLMs), varying how much target information is provided upfront and how much must be acquired through dialogue. Across four human-grounded visual contexts and four interaction protocols, current LVLMs perform significantly below task-level human baselines. Interaction can help when follow-up questions refine or repair an initial target description. Performance is lowest when no initial description is provided and target information must be acquired through questions, indicating that proactive question-driven grounding remains difficult. LVLMs are also poorly calibrated, often reporting confidence that exceeds their empirical accuracy. Follow-up studies confirm these patterns across varied description sources (human versus AI), reasoning efforts, repeated interactions, description providers, and visual contexts. Overall, interactive visual grounding remains challenging, requiring visual matching, information seeking and synthesis.

Multimodal large language models (MLLMs) have made significant progress in understanding and interpreting mul- timedia content. However, their ability to generate me- dia remains limited. Recent approaches have attempted to bridge this gap by translating the hidden representations of token sequences into the embedding space of visual models or directly into raw image data. However, these methods often represent each image using multiple specialised to- kens which significantly increases the input length. This be- comes a major limitation for tasks such as graphic design generation where the output typically involves a seamless blend of thousands of tokens across text, multiple images, and layout information. To address this challenge, a novel architecture is proposed that maps hidden token represen- tations to the embedding space of visual models, such as CLIP ViT-L/14, using a single [IMG] token per image. The architecture employs two shallow MLP blocks, each with a separate compression module followed by a shared expan- sion module, trained with six distinct loss functions. One block aids the other during training and is omitted during inference, resulting in a lightweight solution. Strong perfor- mance is demonstrated in both image-to-design and text-to- design generation tasks.

Retrieval-augmented generation (RAG) improves the factuality of large language models by grounding responses in external documents, but it also exposes a critical security vulnerability: adversarial documents injected into the knowledge database can enter the context window and steer the model toward targeted incorrect answers. Existing post-retrieval defenses rely on instruction following, parametric knowledge, or text-level consistency, all of which can be imitated or optimized against by adaptive attackers. We propose RAGSentinel, a training-free, label-free defense for black-box RAG systems. RAGSentinel uses a surrogate encoder to measure query-conditioned hidden-state shifts induced by retrieved documents, removes shared topic directions, and filters poisoned documents as geometric outliers from a robust majority consensus. We prove that, under an honest-majority assumption and a representation-level separation condition, RAGSentinel exactly recovers a poison-free majority-sized context. Experiments across three question-answering datasets, three LLM families, and multiple poisoning attacks show that RAGSentinel consistently achieves low attack success rates while preserving competitive accuracy and remaining effective against adaptive attacks with full pipeline knowledge.

When an LLM serving deployment runs out of KVcache room, there are two well-established ways out. Tensor parallelism shards the weights and the KV cache across two, four, or eight devices, buying memory headroom at the price of an all-reduce on every layer and a hardware bill that grows with the device count. The algorithms community shrinks the cache in place, with KV quantisation and eviction keeping a single GPU and spending a little quality instead. Compression papers report memory ratios, parallel-scaling papers report throughput curves, and almost nobody puts the two on the same cost axis. We place tensor-parallel configurations (degree 1 to 8) and KV-compressed configurations (16/8/4-bit, keep-ratios down to 0.25) on one costnormalised axis, cost per million tokens against latency, using a profiled simulator calibrated on A100, A40, and H100 hardware, and we go looking for the cost-equivalence crossover. We do not find one. Across two models (Llama-2 at 7B and 70B), three GPU types, and every level of memory relief we could construct, compression is cheaper by 1.20x to 2.00x. A 7B model on an 80 GB device cannot exhaust its KV budget within its own context window, and the boundary that decides between the strategies is model size relative to device memory, at roughly 36B parameters for an 80 GB card. Below that wall, compression dominates and extra GPUs are largely wasted spend; above it, tensor parallelism stops being a choice and becomes an entry ticket: Llama-2-70B is infeasible on one A100 at any KV setting, because the binding resource is weights, which KV compression does not touch. Tensor parallelism is the only lever that improves latency (compression makes per-token latency worse, by 8 to 93%, through batching contention), while compression is the only lever that multiplies capacity per dollar (16.5x, against 1.21x for an eightfold spend on GPUs).

Background: Large Language Models (LLMs) have demonstrated strong performance across a variety of code-understanding tasks, leading many to believe that they can reason about program semantics. However, existing evaluations primarily focus on single-language settings or rely on synthetically generated code, raising concerns about whether current results reflect true semantic understanding. Aims: We investigate whether LLMs can accurately judge functional equivalence across different programming languages in human-written code, a setting that requires deeper reasoning beyond superficial similarity. Method: We introduce PolyHuman, a dataset of human-written programs in CPP, Java, and Python. Using this dataset, we evaluate intra- and inter-language equivalence detection across open-weight and proprietary LLMs, selecting GPT-o4-mini as a representative model to assess stability. We then manually analyze 81 cases of systematic disagreement in which models incorrectly judge functional equivalence, examining the code logic and the generated Chain-of-Thought reasoning. Finally, we categorize these failures and compare them across GPT-o4-mini, Claude-Opus-4.7, and Gemini-3-Flash to determine whether they reflect model-specific issues or broader limitations of state-of-the-art LLMs. Results: We identify a difficulty-dependent breakdown in equivalence judgment (harder problems make the model increasingly prone to misclassifying non-equivalent code as equivalent), a model-specific sensitivity to programming language for the best-performing model (particularly a more conservative behavior on Python), and a partial reliance on similarity-based cues. GPT-o4-mini also shows substantial run-to-run instability under identical settings, indicating inconsistent rather than absent capability. Conclusions: Current LLMs do not reliably capture functional equivalence within or across languages.

Safety alignment in large language models (LLMs) remains brittle against a growing spectrum of attacks. Jailbreak attacks bypass safety mechanisms through crafted prompts, while neuron-level attacks directly prune safety-critical neurons post-deployment. Both exploit a common weakness: safety-relevant information concentrates in a sparse neuron subset. We present NeuronGuard, a fine-tuning-stage defense that simultaneously hardens LLMs against both attack classes by redistributing safety signals across a broader set of neurons. NeuronGuard dynamically identifies safety-critical neurons via periodically refreshed per-layer linear classifiers, forces refusal behavior under deliberate neuron ablation, and applies KL-divergence regularization for distributional consistency. A randomized gradient projection strategy preserves downstream task utility by resolving conflicts between the defense and task objectives. We provide a formal guarantee that NeuronGuard strictly reduces the attack success rate (ASR) upper bound, and experiments across three LLMs, six state-of-the-art attack strategies, and multimodal settings confirm near-zero ASR while maintaining task accuracy, including against white-box adaptive adversaries.

Test-time reasoning methods such as iterative refinement, decomposition, and repeated sampling are often evaluated in isolation, making their gains difficult to compare across models, benchmarks, and evaluation pipelines. We introduce a unified view of these methods as recursion operators over an agent's reasoning trace: GROW, which deepens a single reasoning path; PRUNE, which decomposes and recomposes the problem; and BRANCH, which samples alternative reasoning paths and selects among them. We evaluate all three operators against a single-pass chain-of-thought baseline under a shared harness with identical prompts, token budgets, and grading code. Across five benchmarks and three frontier models, comprising 14 model-benchmark settings, 49,327 graded items, and 151,876 model calls, BRANCH improves accuracy in all 14 settings by an average of 5.98 percentage points and is the best-performing operator in 12. In contrast, GROW yields a mean gain of 2.18 points and degrades performance in two settings, while PRUNE improves accuracy by 0.94 points on average. Analysis shows that BRANCH's advantage arises not only from exploring multiple reasoning paths, but also from recovering from truncation: its gains strongly correlate with the baseline rate of empty, budget-exhausted outputs (r = 0.72). These results weaken the hypothesis that different problems require routing among test-time reasoning operators; at this level of abstraction, repeated branching is consistently dominant. Finally, we show that unpaired evaluation and treating scoring-pipeline failures as model errors can materially change, and even reverse, comparative conclusions, motivating paired scoring as a standard protocol for test-time-compute evaluation.

An agent harness is what turns a language model into an autonomous agent: the surrounding code that builds the model's context, mediates its tools, runs the loop, and persists state across a long-horizon run. This layer, not the model it wraps, is increasingly the binding constraint on agent behaviour. We present a source-level, multi-case study of three open coding-agent harnesses built from deliberately opposing philosophies: LangChain's deepagents (batteries-included), Earendil's pi (radical minimalism), and DeepSeek's dsh (everything-is-a-plugin). Reading each at a pinned commit and following its commit history, we find that the two mature harnesses have travelled in opposite directions (deepagents subtracting authored scaffolding, pi accreting durable infrastructure), yet converged toward one architectural middle form of five recurring elements: a commoditised loop, an append-only replayable session record, model quirks kept as data, progressive disclosure of context, and explicit extension seams. A third harness, read afterward as a held-out check, exhibits all five, and in one seam reuses another's implementation outright. We therefore do not claim independent invention, and decompose the convergence into parallel discovery, diffusion, and literal reuse. Finally, one load-bearing dimension shows no convergence, and indeed no presence: external verifiability, a tamper-evident record an outside party can check without trusting the runtime. We read this absence not as an oversight but as a predictive gap, the next axis on which harnesses for provenance-sensitive domains will differ.

Federated video anomaly detection trains model collaboratively without sharing raw surveillance footage, but limited server-side visibility lets compromised clients to inject backdoor via malicious updates. This paper introduces STAIN-FL, a stealthy targeted backdoor attack injection framework that uses naturally occurring surveillance conditions, including low-light scenes, indoor settings, and crowd density, as contextual triggers. STAIN-FL combines anomaly-to-benign label manipulation with gradient masking over least-updated coordinates to preserve clean accuracy while inducing trigger-conditioned misclassification. We evaluate STAIN-FL on UCF-Crime using 1024-dimensional I3D features in a non-IID four-client multi-agency setting, comparing FedAvg and FedProx under sparse and continuous attacks. Results show that sparse attacks have low-detectability, operationally significant attacks rather than high-intensity attacks: they keep the mean clean-accuracy drop below $2\%$, yet still misclassify more than half of triggered anomalies at peak backdoor accuracy under FedAvg ($56.7\%$) and FedProx ($54.2\%$). Under FedAvg, the sparse backdoor remains above the $25\%$ backdoor-accuracy threshold for an average of $336$ post-attack rounds, highlighting the persistence risk of contextually triggered attacks in surveillance systems.

High-fidelity image-to-3D generation requires a 3D representation that captures both geometry and appearance. However, preserving fine detail across the physically based rendering (PBR) modalities needed for relighting remains challenging. To address this, we propose Luce, a 3D representation that unifies geometry and PBR materials within a voxelized multimodal Gaussian cloud, using dedicated Gaussian primitives for albedo, metallic-roughness, and surface normals. A variational autoencoder compresses this representation into a unified material-aware latent space. A rectified-flow transformer generates this latent from a single image using multi-layer features from a pretrained image encoder that preserve both semantic context and fine spatial detail. The latent is then decoded into relightable PBR Gaussians and an optional textured mesh with a tangent-space normal map. On Toys4K, Luce achieves state-of-the-art single-image-to-3D generation, improving FID by 28% over the strongest baseline. We further evaluate Luce on a benchmark of AI-generated images depicting diverse subjects and materials, where it improves the CLIP image-alignment score over the best baseline (0.8519 vs. 0.8299). Luce generates relightable, geometrically accurate, and materially faithful assets that preserve fine details such as text, logos, and inscriptions.

Pre-execution oversight is core to trusted monitoring in AI control: a fallible LLM monitor vets planned actions before irreversible execution. Over-blocking forfeits usefulness and pressures deployers to disable it. Every protocol must fix a unit of verification: how many actions one call reviews. Existing designs take the unit as given; its effect on fallible monitors is unmeasured. Natural traces cannot isolate it: review length co-varies with error type and position. Catch alone misleads: rejecting everything catches everything. Measuring this needs boundary variation alone and a matched clean control. We introduce the twin-prefix framework, which supplies both. Each gold plan yields a prefix with one injected, environment-accepted error and a clean twin differing in one write. Judging each pair at five nested lengths ties verdict changes to the unit alone. Discrimination is scored by pre-registered informedness, catch minus false rejection. Longer review raises catch; false rejection climbs in lockstep. Informedness peaks at one or two actions for all six judges in both domains: longer windows make zero-shot monitors more rejective, not more discriminative. Replaying withheld observations traces the failure largely to observation deprivation. Safety cases should state the unit and co-report the clean series. Our framework is the first controlled, pre-registered instrument for this choice and never reads catch alone. Our calibrated short unit recovers up to 0.95 informedness over eight-action review, and no tested label-blind policy consistently beats it.

Nitrogen-vacancy (NV) centers in diamond can serve as highly sensitive solid-state quantum sensors for high-sensitivity magnetometry. However, in the noisy intermediate-scale quantum (NISQ) era, extracting reliable information from noisy, finite-shot, and measurement-limited sensing data remains a considerable challenge. Whereas, quantum machine learning (QML) offers a potential path to improve parameter estimation by learning nonlinear relationships between quantum-sensing data and the underlying physical signal. In this work, we investigate the role of QML in magnetic-field estimation within an NV center-inspired magnetometry setting. We formulated magnetic field sensing as a supervised regression task. We compared the performance of several classical machine learning models trained on measurement-based classical data with that of quantum kernel-based models trained on pre-measurement coherent quantum states. Our objective is to isolate the impact of measurement-induced information loss and therefore provide a theoretical upper bound on the sensing performance. The upper bound is achievable only when coherent quantum information is directly available to the learning model. Our results show that QML-based sensing performance improves significantly with coherent quantum-state information, and not much with changes in model complexity or learning paradigm. This observation underscores the importance of learning pipelines that tightly integrate quantum sensors and QML models to enhance magnetic field sensing under realistic constraints.

This study introduces the evolutionarily recurrent decision model (ERDM), a computational reinforcement learning framework designed to examine how evolutionary mismatch, bounded rationality, and satisficing contribute to adaptive and maladaptive behavior. ERDM simulates agents across evolutionary recurrent environments, including threat, prey/goal-pursuits, and alliances. Agents learn through competing rewards abstracted from survival metrics. A validity study under varying adverse childhood experiences demonstrates that distinct adaptive and maladaptive strategies, such as learned helplessness, avoidance, healthy relationships, and aggression, emerge naturally without being hardwired. These results align with empirical literature, showcasing ecological validity. The results suggest that many psychopathology-relevant aspects may be interpreted as bounded cognitive systems operating under modern-ancestral environmental mismatch, positioning ERDM as a key computational cognitive tool that can be extended to other studies.

Surgical spatio-temporal grounding (STG) requires locating, at each video time specified by a procedural question, the object that the question asks about. Existing approaches face a trade-off: vision language models understand the question context but produce imprecise coordinates, whereas open-set detectors provide localized candidate boxes whose confidence does not reflect which box answers the question. We introduce RefineRank, which closes this gap at the candidate-box level. A compact trainable module, RefineNet, combines the language and regional features of a frozen medical vision language model with the proposals of a frozen open-set detector: it predicts a bounded coordinate correction and a quality score for every candidate box, and a fixed decoding rule returns the original or refined box with the highest score. On the MedVidBench Official Rankings (Verified), RefineRank records 0.421 STG mIoU, the highest displayed STG score, while its global multi-metric rank is 11. In a controlled evaluation on separate training and evaluation videos, coordinate correction raises the candidate oracle upper bound from 0.6772 to 0.7302, and ranking the joint pool of original and refined candidates by their RefineNet scores improves STG mIoU from 0.2719 to 0.4534, whereas separately trained selectors over the same pool reach at most 0.4186. These results show that a small box-level module can reconcile question understanding with precise localization without retraining either backbone. Code is available at [https://github.com/linzhe001/RefineRank](https://github.com/linzhe001/RefineRank).

Video generative models provide rich physical priors for robot learning, yet existing world-action models (WAMs) face a fundamental trade-off: synchronous video generation at control rate is latency-prohibitive, while abandoning test-time visual imagination sacrifices task success. We show that visual imagination achieves both real-time inference and superior success rates when generated asynchronously off the critical path and consumed directly in latent space. We introduce GlanceWAM, which decouples imagination from control on a single shared video DiT backbone: an asynchronous proposer glances ahead on a slow clock to imagine a single lookahead frame seconds into the future in the background, while an action head decodes action chunks at control rate (48 ms) purely in latent space without blocking. Enabled by a non-interfering attention mask that isolates video representations and staleness-robust horizon training that accommodates asynchronous lookahead aging, GlanceWAM breaks the speed-success dilemma. Trained purely on demonstrations, it attains 72.2% on the 24-task RoboCasa kitchen benchmark (vs. 67.1% for synchronous Cosmos Policy) and 99.0% on LIBERO while cutting per-chunk control latency $24\times$ relative to synchronous world-action models (48 ms on one A100). In single-arm and bimanual real-robot manipulation, it achieves higher average success than $\pi_{0.5}$ without any robot-data pretraining. Code is available at https://github.com/linhanwang/GlanceWAM.

Video-derived social graphs can appear precise while inheriting uncertainty from detection, tracking, classification, and aggregation. We present DairySocialTwin, a graph-centered audit framework applied to 203 videos (122 h 23 min) from 15 views at three commercial dairy farms. Curation retained 447 of 839 candidate clips; 400 entered a frozen source-clip index, including 71 legacy clips with unavailable farm and time metadata. Five models were compared on a development split, with an NRI-style model retained as an exploratory trade-off between class balance and dyadic output. In a 90-min manual audit, reconciliation improved switches, fragmentation, and continuous-track duration across three gap definitions, but constrained identity capacity and inherited detector errors limited claims of identity accuracy. Matched manual and reconciled summaries showed coarse agreement and assignment sensitivity. Camera-specific semantics, observation opportunity, and identity-summary-to-video trace-back made these dependencies inspectable. DairySocialTwin supports auditable relational inference without claiming permanent identity, calibrated behavior, or biological diagnosis.

Aggregate factuality scores hide where a language model succeeds, which relations it confuses, and whether an answer survives innocuous changes to the question or decoder. We introduce PROOF, a profile-oriented benchmark for factual coverage in instruction-tuned language models. PROOF converts a frozen Wikidata snapshot into 18,486 English multiple-choice questions grounded in 11,779 semantic facts, 101 classes, 392 properties, and 14 domains. Each question has an explicit "I don't know" option, a "No correct option" control, and nine controlled formulations; 1,849 questions are no-correct-option traps. We evaluate 18 open-weight model deployments on 166,374 prompts each and separately perturb decoding on a fixed 10% subset. Base factual accuracy ranges from 6.58% to 57.59% (chance: 8.64%), yet every model has a 19.3-36.4 percentage-point spread across domains. Paired facts reveal direction-dependent retrieval, usually favoring subject-to-object queries, with the pattern reversing for one model. We find no consistent temporal penalty after exact-stratum adjustment. Neutral wording changes accuracy by as much as 26.5 percentage points, while adversarial formulations break up to 79.4% of answers that were initially correct. Direct switching to an injected false label varies from 0.04% to 27.5%, showing that accuracy loss and hint following are distinct. Selected-token confidence often indicates severe overconfidence, and decoder perturbations move accuracy by up to 15.7 percentage points and domain profiles by 16.8 points. PROOF therefore measures factual coverage as a structured, intervention-aware profile rather than a single claim about what a model "believes."

Large language models make statements concerning their own "minds". When asked whether or not they are conscious, they usually say that they are not; if they are prompted to ignore their guidelines, they might say that they are; and if asked to write a diary from their point of view, they often describe a human lifestyle. All these contradictory ways of describing themselves are the result of the way the questions are phrased. This paper shows exactly where such descriptions came from, and considers when they can be regarded as evidence for what they claim to report. In order to achieve this, we traced the provenance from end to end. We examine Pythia and OLMo 2 across 66 pretraining checkpoints, three of the post-training stages of OLMo 2 that have been released, about 90,000 continuations, and four training corpora. A set of forty items is used in order to keep an eye on self-reference, frame sensitivity, and self-ascription throughout training. The denial formula was almost completely missing from the vast quantity of text that the models initially came across, but was present in a dense manner in the small, carefully chosen set of example dialogues that they were trained on later on. Supervised fine-tuning causes first-person AI language to become the default, and the other affirmations are then suppressed using preference optimization. The final policy is still very sensitive to framing and to the chat template itself. Two of the conditions which are set out in the epistemology of testimony determine whether or not these outputs can act as evidence for what they claim to report: reference and causation. Reports produced by the base model fail the reference condition, and those obtained after training remain sensitive to the frame and do not show state dependence. The result is symmetric in that trained denials are no more admissible than trained affirmations.

Voxel-based soft robots (VSRs) present a promising avenue for developing artificial organisms with lifelike intelligence. However, the vast design spaces and expensive evaluations substantially challenge their design optimization. Here we develop MISCO, a novel evolutionary framework empowered by deep generative models to optimize VSR designs with theoretical guarantees. MISCO integrates an estimation-of-distribution algorithm with a meticulously designed variational autoencoder featuring multi-task learning, position awareness, and inter-voxel signaling. These key components enhance the representational capacity of VSR morphologies and facilitate highly efficient sampling and optimization of morphological distributions. We provide theoretical guarantees for MISCO's asymptotic convergence to globally optimal designs, alongside a favorable convergence rate. Extensive simulated experiments further demonstrate MISCO's exceptional effectiveness in navigating vast design spaces, evolving high-performing VSRs for diverse tasks while flexibly balancing optimization efficiency and morphological diversity. Being validated both empirically and theoretically, MISCO represents a step change towards more scalable and reliable soft robot development.

Recent advances in robotics highlight hierarchical configurations of robot morphology, where multiple levels of functional substructures synergize to facilitate intelligent behaviors. This hierarchical perspective, while particularly advantageous for voxel-based soft robots (VSRs) to ease design and control complexities, is hindered by its heavy reliance on domain expertise. In this work, we address the following question: can we derive such hierarchical design principles solely from existing successful designs? We answer affirmatively by presenting RoboLDA, a Bayesian probabilistic model that decomposes VSR morphology generation into a four-level hierarchy: "task-robot-organ-voxel", and is trained via variational inference. Through extensive experiments on simulated VSRs, we verify the presence of consistent, intuitive hierarchical patterns underlying high-performing VSR designs and showcase RoboLDA's proficiency to extract and leverage these hierarchical priors for zero-shot robot design in unseen tasks. The generated designs, even without further optimization, achieve on average 106.4% of the optimized performance produced by evolutionary algorithms. Additionally, the organ structures inferred by RoboLDA serve as valid functional substructures, significantly enhancing synergistic motion control when integrated with modular control policies. Our work pioneers hierarchical generative modeling of robot morphology, offering a promising pathway towards more interpretable and generalizable development of embodied agents.

Dynamic Gaussian Splatting provides an explicit representation of evolving 3D scenes, but existing approaches are primarily optimized for reconstruction, future-state generation, or rendering rather than for learning reusable predictive dynamics. We propose 4DGS-JEPA, a Gaussian-native joint-embedding predictive architecture for causal multi-horizon prediction over dynamic Gaussian scenes. The model uses a hierarchical scene-, motion-group-, and Gaussian-level representation together with a horizon-conditioned transition operator that supports both direct prediction and recursive rollout. Its central principle is temporal composition: different chronological transition paths reaching the same future endpoint should produce compatible predictive states. Endpoint and multi-horizon path supervision anchor these predictions to future target embeddings, while a selective geometry decoder and geometry-level composition ground the learned dynamics in consistent group motion and Gaussian geometry without requiring complete future appearance reconstruction. We further introduce a hybrid correspondence mechanism that combines persistent canonical identity with residual optimal-transport matching under reordering and topology change. We characterize zero-loss path agreement and finite-error rollout accumulation theoretically. Three controlled experiments provide mechanism-level evidence that temporal composition reduces latent path dependence while retaining predictive accuracy, geometry-level composition improves consistency of decoded motion, and hybrid correspondence preserves reliable identity while remaining robust when correspondence becomes ambiguous. Together, 4DGS-JEPA provides a predictive, temporally compositional formulation of dynamic Gaussian worlds.

Clinical deterioration unfolds through coupled, partially observed trajectories, not a single diagnostic label. We introduce PGP-Clinical-TimeKAN, a trajectory-first framework for joint probabilistic forecasting of multivariate physiology. It combines missingness-aware temporal encoders, a soft organ-system prior, patient-specific relations, nonlinear Kolmogorov-Arnold messages, and a low-rank multivariate Student-t head. We evaluate 24-hour histories and six-hour forecasts on a frozen MIMIC-IV-derived cohort of 6,882 patients and 54,694 windows. Across five seeds and 13 models, PGP-Clinical-TimeKAN obtains the second-lowest normalized MAE (0.37727 +/- 0.00029) and the lowest RMSE (0.52656 +/- 0.00034). It reduces MAE by 0.52% relative to deterministic TimeKAN. For probabilistic forecasting, it reaches a marginal NLL of 0.66380 and a CRPS of 0.27301. Empirical coverage is 0.533, 0.831, and 0.958 for nominal 50%, 80%, and 95% intervals. Removing relational structure causes the largest ablation loss. Increasing covariance rank improves joint likelihood but has little effect on point accuracy. A trajectory-derived risk score remains weaker than a dedicated GRU-D classifier (AUROC 0.603 versus 0.650), which limits the present clinical claim. Joint trajectory forecasting therefore provides an inspectable intermediate task, but accurate physiology forecasts alone do not ensure a calibrated event detector.

AI-generated art has sparked debates around potential plagiarism, as these images may closely resemble existing artworks. This research quantifies the similarity between original pieces and AI-generated counterparts, particularly those produced by the Stable Diffusion XL Refiner 1.0. We use Siamese Networks with frozen CLIP encoders and cosine similarity optimized through triplet loss. A dataset of paired original and generated images was built using image-to-image generation and custom prompts, enriched with semantic descriptors and BLIP-2 captions. Prior studies report up to 81% style replication and 90% visual similarity. Our results show high discriminative performance: training accuracy reached 99.9%, and the best model configuration achieved 99.4% test accuracy with strong inter-class separation ($\delta \mu$ = 0.677), demonstrating the effectiveness of our semantic-visual embeddings.

The rapid proliferation of Large Language Models (LLMs) has raised concerns about their environmental impact during inference. While Green AI research has focused on datacenter GPUs and embedded platforms, the energy profile of LLM inference on Apple Silicon, with its unified memory architecture, remains unstudied. This paper presents GreenBench, a benchmarking framework that evaluates the energy efficiency, throughput, and carbon footprint of five open-source LLMs (3-9B parameters) across three NLP tasks on an Apple M4 Pro with 48 GB unified memory. Using macOS powermetrics for direct power measurement and Ollama's nanosecond-precision timing, we find that the M4 Pro draws only 0.47 W of CPU+GPU package power during sustained inference, with total system power of 8-12 W, achieving 30-40x better energy efficiency per token than datacenter GPUs in single-user deployment. Smaller models (3-3.8B) deliver 2.6-4.2x higher throughput and up to 62% less energy per token than larger models (7-9B). Pareto analysis identifies Qwen 2.5 (7B) as the optimal accuracy-efficiency trade-off at 57% MMLU and 59 tokens/s, while Llama 3.2 (3B) suits latency-critical applications at 175 tokens/s. We provide per-token energy at package and system levels with CO2 estimates for India and US grids.

Retrieval-Augmented Generation (RAG) systems have attracted significant interest for their ability to mitigate hallucinations in Large Language Models (LLMs). Although knowledge databases for RAG are increasingly diversifying to include various modalities such as speech and text, research on handling such multi-modal database scenarios remains limited. In this paper, we propose STeReO (Speech and Text Reranking Orchestrator), a reranker based on speech and text retrievers that aggregates disparate modality databases. To address the lack of specialized training data, we first curate a dataset comprising queries, mixed-modality evidence, and their corresponding relevance ranks. We then train the reranker and evaluate its effectiveness in both single-modality and mixed-modality scenarios. Our results demonstrate that the proposed algorithm excels at selecting the most relevant evidence, thereby significantly improving downstream question-answering performance.

Multimodal large language models (MLLMs) achieve strong performance on VQA and scene understanding, yet affective reasoning remains vulnerable to shortcut behavior. Models may predict correct answers while neglecting people-centric cues such as micro expressions and body language, which weakens traceability and external verification. Prior reinforcement learning approaches mainly reward context or logical coherence without explicitly enforcing attention to human evidence. In addition, LLM as a Judge scoring often suffers from score clustering, which reduces reward discriminability. We propose AffectOmni, a GRPO trained framework for verifiable affective reasoning. AffectOmni introduces People Focus and Temporal Order rewards to encourage people-centric evidence selection and temporally structured reasoning, and it adopts within-group comparative scoring to produce more stable and discriminative reward signals. For verification, a Thinking Summarizer converts free form rationales into executable evidence instructions, which are grounded into pixel level evidence regions via SAM3 to provide an externally auditable interface outside the training loop. Experiments on IntentBench, Daily Omni, and WorldSense show consistent improvements over open source 7B scale baselines, including gains of 4.66% on emotion recognition and +14.29% on temporally sensitive tasks. Code is available at https://github.com/eliot127825-rgb/AffectOmni_nobody.

Retrieval-Augmented Generation conditions a language model on chunks retrieved from a document collection. Its accuracy is therefore limited by the chunking and embedding stages that determine what can be retrieved. We compare Turkish document question answering across three chunking strategies (fixed-length, semantic, and layout-aware Docling), five embedding models, and two LLMs, over three documents with contrasting layouts. Every configuration answers the same question set, which allows component effects to be separated by paired testing rather than inferred from separate benchmarks. The fully crossed design yields 9{,}000 graded question-answer evaluations, each scored by an independent judge model, and component comparisons are tested by paired McNemar tests under Holm correction. The three leading embedding models are statistically indistinguishable, so language specialization yields no measurable retrieval advantage. The faster LLM is not the more accurate one. The preferred configuration depends on content type, since layout-aware chunking helps table-heavy documents far more than text-heavy ones.

Incident risk prediction from longitudinal electronic health records (EHRs) is challenging because relevant signals are multimodal, weak in isolation, and distributed across irregular patient histories. We propose structured evidence routing, a router-predictor-reviewer workflow that separates full-record access from disease-specific assessment. The router organizes the complete pre-index EHR into a compact summary and targeted evidence slices; the predictor uses this evidence to form an evidence-linked risk assessment, which the reviewer critiques. For comparison with supervised EHRSHOT baselines, we pair the routed evidence summaries with a supervised classifier readout. Across five 1-year incident diagnosis tasks, our method reaches the AUROC range of established supervised EHRSHOT baselines and remains competitive on AUPRC, while exposing a patient-specific evidence trail. Internal pre-readout ablations further suggest that routing, laboratory evidence, task guidance, and review each contribute to performance.

Systematic dialectal performance gaps in language models (LMs) are well documented, but the source of these disparities within the modern language modeling pipeline remains unclear. Our study traces this "dialect tax" across the natural language processing pipeline. Using parallel English dialect corpora that hold meaning fixed while varying surface form, we first confirm that LMs recognize matched Standard American English (SAE) and dialectal texts as semantically equivalent. However, we discover further representational gaps corresponding to downstream performance gaps. Across model families and generations, modern LMs still encode dialectal texts unequally during tokenization, pre-training, post-training, and inference. Strikingly, bypassing traditional subword segmentation via a character-level counterfactual tokenizer removes neither input and output asymmetries nor dialectal accuracy gaps. During pre-training, dialect pairs induce more divergent gradient updates than pairs of entirely unrelated SAE documents, indicating that models find semantically equivalent dialectal content harder to learn from than unrelated SAE documents. During post-training, reward models show contextual, unstable dialect preferences, assigning higher values to isolated AAVE-exclusive tokens than to SAE-exclusive tokens, while full reasoning contexts receive task- and model-dependent dialect penalties. Overall, our findings suggest that the dialect tax is encoded and accumulated not by any one step in isolation, but at every step of the language modeling process.