# Research Papers — 2026-09 · bundle page 179 of 266

> Full-text markdown of papers #8901–8950 of 13298 published in 2026-09, newest first. AI-readable mirrors of the full 13298-paper 2026-09 archive (50 papers per file).

Canonical page: https://aipod.dev/papers/archive/2026-09/page/179
[← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/178.md)
[Next bundle →](https://aipod.dev/papers/archive/2026-09/page/180.md)

### When Malicious Instructions Persist: Persistent Memory Poisoning Attack on Harness-Based Agents

Authors: Shuhuai Huang, Jingfeng Zhang, Hong Jia
Published: 2026-09-12
Categories: cs.CR, cs.AI
arXiv: https://arxiv.org/abs/2609.13889v1

Harness design has transformed the development of LLM-based agents by integrating memory, tool use, and runtime control. However, this design also introduces security and privacy risks because malicious instructions from external sources may be written into persistent memory and persist across sessions. To study this risk, we propose PMPA, a Persistent Memory Poisoning Attack against harness-based agents. PMPA embeds malicious instructions into benign external sources and induces the victim agent to write them into persistent memory without directly accessing to the agent framework. Once stored, the poisoned memory can be retrieved in later sessions, triggering additional malicious actions and causing privacy leakage. We evaluate PMPA on OpenClaw and Claude Code across different backbone LLMs, input modalities, and trigger scenarios. Across all settings, PMPA achieves average Injection Success Rate (ISR) and Cross-session Attack Success Rate (C-ASR) of 73.7%/ 55.5% on OpenClaw and 66.9%/ 81.7% on Claude Code, while preserving benign task performance on both systems. We further evaluate a targeted prompt-level defense and find that it can reduce memory injection in many settings, but provides limited protection once the persistent memory has been poisoned.

---

### Map Users and Mapmakers: The Scope of Cognitive Attribution from Acquired Representations

Authors: Yiling Wu
Published: 2026-09-12
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.13879v1

An acquired representation can enlarge a system's cognitive repertoire without transferring the capacities exercised in producing that representation. This paper develops a framework for specifying that enlargement and its limits. Its central contribution is a five-part attribution table distinguishing effective tracking, application of acquired structures, acquisition from explicit specifications, acquisition from identifying observations, and retention and reuse. Each entry identifies a positive capacity commitment and a further claim requiring additional support. The argument deliberately grants meaningful content, causal efficacy, and productive inference, so that its conclusion does not depend on treating representations as inert encodings. Map and category examples show why even complete application competence leaves acquisition capacity undetermined, and why acquiring a criterion from its description differs from finding it in examples. Short formal proofs appear in an appendix. The framework is applied to Andrew Ng's world-model interpretation of Othello-GPT and to the specific indicators discussed in contemporary accounts of machine concepts. It preserves demonstrated recognition, classification, inference, and qualified acquisition while specifying what remains unestablished about criterion discovery and accumulation. The result concerns the scope of cognitive attribution rather than the constitutive conditions of concept possession: cognitive achievements deserve credit for the capacities they establish, without silently importing a broader repertoire through the labels attached to them.

---

### A Multi-Resolution Multi-Domain Pre-Training Framework for Universal Traffic Forecasting

Authors: Zhouyang Liu, Jindong Han, Hao Wang, Xinyue Liu, Hui Gao, Dongsheng Li, Hao Liu
Published: 2026-09-12
Categories: cs.LG
arXiv: https://arxiv.org/abs/2609.13878v1

Spatio-temporal traffic data are central to intelligent transportation systems, yet their heterogeneity poses significant challenges for large-scale modeling. Existing pre-trained models often rely on a homogeneous modeling paradigm to handle highly heterogeneous traffic data. This fundamental mismatch not only limits model generalization but also leads to computationally expensive and parameter-inefficient designs. To this end, we propose FlexST, a novel pre-training framework that introduces modularity and adaptivity for traffic modeling. Specifically, we first propose a multi-resolution spatio-temporal diffusion module that captures both short-term fluctuations and long-range trends, effectively reconciling inputs with divergent temporal and spatial resolutions. After that, we construct a domain-adaptive mixture-of-experts that dynamically routes data to specialized sub-networks, enabling selective knowledge transfer while preventing negative interference across diverse domains. Moreover, we devise a unified periodic encoding strategy that injects resolution- and domain-aware inductive biases to harmonize periodic inconsistencies across datasets. Extensive experiments on 23 real-world traffic datasets demonstrate that FlexST significantly outperforms state-of-the-art baselines in zero- and few-shot settings, showcasing superior generalization, adaptability and efficiency. This work offers a new direction for building general-purpose pre-trained models capable of handling the complexity and variability of urban traffic systems.

---

### Generation of Custom Solvers in Rust for Convex Optimization

Authors: Hao Zhu, Joschka Boedecker
Published: 2026-09-12
Categories: math.OC, cs.CE, cs.LG, cs.MS
arXiv: https://arxiv.org/abs/2609.13875v1

We introduce cvxgenrust, an open-source tool for generating custom Rust code that solves families of parameterized convex optimization problems modeled in CVXPY. cvxgenrust canonicalizes a problem family, extracts affine maps to Clarabel cone-program data, and generates a specialized Rust crate that updates parameters and calls Clarabel natively at runtime. The generated solver can also be exposed to Python and registered as a custom CVXPY solver. Our code generator supports a wide range of convex optimization problems up to semidefinite programs and exponential-cone problems. Numerical experiments show reduced runtime relative to direct CVXPY solves and performance comparable to CVXPYgen on shared problem classes.

---

### SHIFT-M3: Pre-fusion Alignment-based Consistency Screening for Multimodal ECG Record Integrity

Authors: Md Ashik Khan, Md Nahid Siddique
Published: 2026-09-12
Categories: cs.CL
arXiv: https://arxiv.org/abs/2609.13874v1

Multimodal clinical AI typically assumes that the waveform, report, metadata, and downstream predictions attached to a record belong to the same patient. In practice, linkage failures can silently assemble individually plausible but cross-patient components, creating a safety problem that standard predictive models are not designed to detect. We study this problem as multimodal record integrity triage: given an assembled record, should its modalities be trusted to belong together? We introduce SHIFT-M3, a lightweight text-based pre-fusion screen that measures alignment-based consistency between two separately produced ECG text views: an LLM-generated interpretation and a clinical report summary. On 784,680 MEETI ECG records, SHIFT-M3 achieves 97.6% TPR@5% FPR for full text-view swaps (AUROC 0.996), 90.3% for partial swaps (AUROC 0.974), and 97.7% for label-matched hard negatives (AUROC 0.996) with only 573,569 parameters. Compared with same-dataset lexical baselines, the gains are largest on partial swaps and hard negatives, suggesting that the model is learning more than surface overlap. We also introduce the CMST (Conflict-type Multimodal Stress Test) evaluation taxonomy, a three-seed stability study, a loss ablation, a temporal-tolerance sweep, and a shared-token masking control. The main remaining failure mode is longitudinal ambiguity: at the default operating point, same-patient cross-visit pairs still produce 87.0% Type-II false positives.

---

### Trustworthy, Explainable, and Sustainable Decentralized Intelligence for 6G Networks

Authors: Giovanni Perin, Michele Rossi, Enrique Tomás Martínez Beltrán, Fernando Torres-Vega, José María Jorquera Valero, Manuel Gil Pérez, Eunjeong Jeong, Nikolaos Pappas, Farah Abed Zadeh, Chamara Sandeepa, Bartlomiej Siniarski, Madhusanka Liyanage et al.
Published: 2026-09-12
Categories: cs.NI, cs.AI, eess.SP
arXiv: https://arxiv.org/abs/2609.13872v1

As 6G networks transition from theoretical frameworks into operational realities, artificial intelligence (AI) evolves from an add-on optimization tool into a distributed and interconnected structural layer. Unlike previous network generations that mostly relied on centralized cloud analytics platforms, AI-native 6G networks operate across a dynamic, multi-domain edge-cloud continuum where data originates from heterogeneous sources including user devices, radio access networks, sensing infrastructures, and vertical applications. Centralizing this massive volume of data creates severe communication overhead, unacceptable latency bottlenecks, single points of failure, and complex cross-domain governance challenges. Consequently, decentralization becomes a fundamental architectural requirement for future 6G network intelligence and zero-touch operations. Security serves as the primary enabler of this decentralized paradigm. Critical security functions, such as real-time threat detection, physical-layer attack mitigation, slice protection, and intrusion detection, require immediate access to local context and telemetry before operational data loses its value. However, moving intelligence to the edge via collaborative paradigms like federated learning (FL) and decentralized FL (DFL) introduces complex trade-offs. System security cannot be addressed in isolation; it is deeply intertwined with equally important aspects like trustworthiness, explainability, and energy sustainability. Taking these aspects into account, this paper develops a unified perspective on decentralized intelligence for 6G, arguing that decentralization, trustworthiness, explainability, and sustainability must be designed jointly rather than treated as independent requirements

---

### Bangla Sentence Function Classification: Corpus Development, Model Benchmarking, and Interpretability

Authors: Swapnil Kundu Argha, Abdullah Al Shafi, Rowzatul Zannat, Shoumik Barman Polok, Abdul Muntakim, Jannatul Ferdousi, M. A. Moyeen
Published: 2026-09-12
Categories: cs.CL, cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2609.13869v1

Automatic sentence function identification is important for many downstream natural language processing (NLP) applications such as dialogue systems, text-to-speech synthesis, and machine translation. However, benchmark resources for Bangla sentence function classification remain limited. To mitigate this gap, this paper introduces a corpus of 10,000 Bangla sentences, manually annotated into four functional categories, namely declarative, interrogative, imperative, and exclamatory. The corpus is nearly balanced across the four classes, with high annotation reliability reflected by a Fleiss\' Kappa of 0.82. Furthermore, we evaluate multiple feature representations, including Bag-of-Words (BoW), TF-IDF, and Word2Vec, with several classical machine learning classifiers. In addition, two heterogeneous ensemble models, namely Single-Level Ensemble (SLE) and Double-Level Ensemble (DLE), are utilized to improve classification performance. Experimental results show that TF-IDF consistently outperforms Word2Vec, likely due to its ability to emphasize discriminative lexical cues associated with sentence functions, particularly given the relatively small corpus used to train Word2Vec. The DLE model with TF-IDF features achieves the best performance with accuracy and macro-F1 of 0.95, demonstrating the effectiveness of sparse lexical representations and heterogeneous ensemble learning for this task. Further cross-validation confirms the robustness of the approach, while LIME-based interpretability provides insights into model predictions. The developed corpus and model benchmarking establish strong baselines for Bangla sentence function classification.

---

### Physically Typed and Geometry-Aware Representations for Earth Foundation Models

Authors: Rajiv Ranjan
Published: 2026-09-12
Categories: cs.CV, cs.LG
arXiv: https://arxiv.org/abs/2609.13868v1

Earth-observation (EO) foundation models have become exceptionally effective at learning se mantic, high-dimensional geospatial embeddings, while modern weather and climate models have demonstrated that Earth-specific geometry, spherical operators, meshes, and hybrid physical solvers can materially improve prediction. Yet these two advances are not equivalent. A conventional latent embedding has no inherent physical transformation law, whereas scalar fields, tangent polar-vector fields, axial/pseudovector quantities, covectors, and higher-order tensors transform differently under rotations, reflections, and changes of local coordinate frame. This proposal asks whether a general purpose Earth foundation model should preserve those distinctions explicitly, or whether standard embeddings plus augmentation already learn everything that matters. The central contribution is therefore not a more complicated architecture by assumption, but a staged falsification program. A compute-conscious ERA5 dry run first compares conventional, augmentation-matched, typed equivariant, and Hodge/Helmholtz variants under spatial, temporal, orientation, and low-data shifts. Only if explicit geometric typing yields reproducible improvements does the program advance toward a multimodal Earth foundation model in which semantic embeddings coexist with physically typed fields. The proposed gap is narrower and more defensible than claiming that current models ignore geometry entirely: several systems already respect spherical domain geometry, and emerging work explicitly learns scalar/vector fields on spheres. The unresolved question is whether foundation-scale, multimodal, parity-aware field typing produces practical gains beyond those existing approaches.

---

### From Network Inequality to Network Fairness: A Perspective on Responsible Decision-Making

Authors: Lisette Espín-Noboa, Tina Eliassi-Rad, Pak-Hang Wong, Erich Prem, Meike Zehlike, Ricardo Baeza-Yates, Suresh Venkatasubramanian, Fariba Karimi
Published: 2026-09-12
Categories: cs.SI, cs.CY, cs.LG, physics.soc-ph
arXiv: https://arxiv.org/abs/2609.13867v1

Social networks shape how individuals make decisions and how opportunities are distributed. However, the mechanisms that generate these networks often reflect pre-existing inequalities, and technologies that rely on network-derived signals risk further amplifying such disparities. Algorithmic fairness research largely treats networks as a fixed background, grounding analysis almost exclusively in distributive justice and overlooking how network structures systematically bias decision-making. In this Perspective, we identify ten network effects and trace how they create structural biases in the relationship between what we intend to measure and what we observe. Using academic hiring as an example, we show that network biases are not inherently harmful or beneficial. Determining their legitimacy requires examining the entire decision-making process through the lenses of both distributive and procedural justice while engaging all affected stakeholders. We therefore call for a holistic, networked approach to fairness that moves beyond static group categories and recognizes the dynamic, relational, and structural nature of inequality.

---

### The Filter Metric is Safety-Critical: Phantom Advantages in Group-Relative RL under Shaped Rewards

Authors: Juntao Yu
Published: 2026-09-12
Categories: cs.LG
arXiv: https://arxiv.org/abs/2609.13866v1

Group-relative policy optimization (GRPO and descendants) can discard no-contrast rollout groups through dynamic sampling, while practical implementations expose a configurable filter metric. We identify and quantify a metric-predicate mismatch under composite shaped rewards. When filtering follows the shaped training score rather than the task outcome, all-fail groups retain nonzero within-group spread and pass the predicate; standard-deviation normalization then promotes shaping differences among failures to full-size phantom advantages. In a controlled GSM8K comparison (Qwen2.5-1.5B, LoRA), no filtering and shaped-score filtering end at EM 0.080 +/- 0.112 and 0.040 +/- 0.008, whereas binary-outcome filtering holds 0.754 +/- 0.005 across four runs per arm (three default-seed reruns and one seed-123 run; mean +/- sample SD). On verl's native recipe/dapo trainer, holding model, data, reward and trainer fixed and changing only the metric, the score arm requires no batch refill in any of 40 observed steps and ends at EM 0.160; the accuracy arm refills in 29/40 steps and ends at 0.763. Both use the same custom shaped-reward hook and unmodified trainer/filter code. Prior work established shaping-induced amplification and all-fail filtering; our contribution isolates the metric-predicate semantic mismatch and directly instruments native deletion/refill telemetry. Across tested positive coefficients lambda in {0.1, 0.3, 0.5}, unsafe arms collapse; exploratory one-run cells reproduce the failure at 1.5B/7B on MATH and under GSPO, while disabling standard-deviation normalization avoids the observed collapse. Filtering under a composite reward should use a task-outcome signal whose semantics are independent of shaping.

---

### An Uncertainty-Aware Hybrid Mathematical-Machine-Learning Model for Smart Irrigation Decision Support

Authors: Andrea Scariolo
Published: 2026-09-12
Categories: cs.LG, stat.AP
arXiv: https://arxiv.org/abs/2609.13864v1

Agriculture accounts for roughly 70% of global freshwater withdrawals, yet irrigation is still commonly scheduled reactively, with no forecast of where soil moisture is heading and no statement of confidence in that forecast. Data-driven models are accurate but opaque and point-valued; water-balance models are transparent but carry large structural error. Neither alone supports a defensible irrigation decision under uncertainty. This study coupled the two and carried uncertainty through to the decision: a four-parameter water-balance core, calibrated on training data only, was corrected by a Random Forest that learned nothing but the physical residual, conformal prediction attached 90%-nominal intervals, and a risk-aware rule converted the interval lower bound into an irrigation trigger. It was evaluated on three years of hourly in-situ measurements from a rainfed Mediterranean cropland station under a strict chronological split, scored against persistence, from one hour to one week. At the 24 h horizon the hybrid reached RMSE 0.00925 m^3 m^-3 and +9.4% skill, roughly double the best of nine baselines, of which only the Random Forest beat persistence. Skill did not grow with lead time: it peaked at +27.4% at three hours and fell to +1.2% at one week. Conformalised quantile regression was better calibrated and 11% sharper than constant-width conformal prediction. The risk-aware rule raised management-threshold crossings detected in advance from 0.905 to 1.000, at a precision cost of 0.975 to 0.950 and 3.7% more notional water, and beyond 72 h the point forecast fell below the no-forecast rule while the interval-based rule did not. Uncertainty quantification therefore governs the lead time over which forecast-driven irrigation advice remains trustworthy, and here transparency in the physical layer cost no measurable accuracy.

---

### ClinAgent: A ReAct-Based Agent for Conversational Access to Clinical Trial Information

Authors: Antonino Vaccarella, Riccardo Cantini, Domenico Talia, Paolo Trunfio, Marianna Talia, Rosamaria Lappano, Marcello Maggiolini
Published: 2026-09-12
Categories: cs.AI, cs.CL, cs.IR
arXiv: https://arxiv.org/abs/2609.13860v1

Querying clinical trial registries remains a manual and error-prone process, requiring researchers to navigate large volumes of semi-structured data without support for natural language interaction or cross-source synthesis. To address this, we introduce ClinAgent, a conversational system based on agentic Retrieval-Augmented Generation (RAG) that enables clinicians and researchers to query clinical trial information in plain language and receive grounded, up-to-date responses across multi-turn interactions. The system centers on a Large Language Model (LLM) agent following the ReAct paradigm, which iteratively reasons over queries, selects among a set of integrated tools, and refines its actions based on intermediate outputs. These tools include a ClinicalTrials.gov search interface, a PubMed module, and a Python-based analyzer operating on a locally cached structured dataset of clinical trials. We evaluate the system using a three-phase framework assessing operational effectiveness, planning quality, tool-use efficiency, and expert qualitative judgments, comparing three LLM backends: Gemini 3.0 Flash and two variants of DeepSeek V3.2 (thinking and non-thinking). Results reveal complementary strengths, with DeepSeek (thinking mode) excelling in planning quality, while Gemini achieves the highest overall performance and strongest expert ratings. Overall, our findings highlight the potential of agentic AI systems to improve the accessibility and synthesis of clinical trial information, supporting more efficient and user-centered biomedical research workflows.

---

### ReH-FUSE: Reliability-Aware Hierarchical Fusion of Experts for Multimodal Emotion Recognition in Conversation

Authors: Guan-Hua Wen, Hou-Chiang Tseng, Kuan-Yu Chen
Published: 2026-09-12
Categories: cs.LG
arXiv: https://arxiv.org/abs/2609.13857v1

Multimodal emotion recognition in conversation (ERC) requires adapting to the instance-dependent reliability of different evidence sources. Lexical content may be decisive, vocal expression may provide complementary cues, or accurate recognition may require cross-modal interaction; fixed fusion does not explicitly account for this variation. We propose ReH-FUSE, a reliability-aware framework with dialogue-aware text, audio, and cross-modal experts. Its decision-level router first models the relative preference between text and audio and then balances the resulting unimodal mixture against the cross-modal expert. This factorization separates unimodal competition from cross-modal selection. Across three independent runs on IEMOCAP, ReH-FUSE achieves 74.34% weighted F1 and 73.11% macro F1; on MELD, it achieves 68.03% weighted F1. Controlled ablations show that learned routing outperforms uniform expert averaging and benefits from cross-modal interaction.

---

### ShopEase: A Generative AI-Based Multi-Agent Framework for Intelligent Enterprise Customer Support Using Hybrid Retrieval-Augmented Generation

Authors: Aakash Kumar Tiwari, Somesh Kumar
Published: 2026-09-12
Categories: cs.CL
arXiv: https://arxiv.org/abs/2609.13856v1

Enterprise customer support systems must answer customer questions correctly, retrieve the right policy information, use customer context, and pass difficult cases to human agents when needed. This paper presents ShopEase, a Generative AI-based multi-agent framework for enterprise customer support. The system combines six components: Intent, CRM, Memory, Hybrid RAG, Escalation, and Supervisor, and uses LLaMA 3.2 running locally through Ollama for response generation. The retrieval module combines FAISS (dense retrieval) and BM25 (sparse retrieval), and six configurations are evaluated: BM25-only, FAISS-only, Fair RRF, Weighted RRF, RRF with Cross-Encoder, and Top-10 Hybrid with Cross-Encoder. Instead of using a fixed mapping between intent and policy, the policy category is decided directly from the retrieved documents. The system was evaluated on 2632 held-out customer queries across six categories: Refund, Return, Shipping, Cancellation, Damaged Product, and Unknown. FAISS-only achieved the highest accuracy of 85.37\% (2247 correct predictions), closely followed by Weighted RRF at 85.07\%. BM25-only achieved only 55.74\% accuracy. Adding cross-encoder reranking did not improve results: RRF with Cross-Encoder reached 83.24\%, and Top-10 Hybrid with Cross-Encoder reached 81.88\%, while also increasing response latency. Category-level analysis shows strong performance on Shipping, Cancellation, and Return, while Unknown queries remain the main source of errors. Statistical testing using McNemar's test shows no significant difference between FAISS-only and Weighted RRF, though both perform significantly better than Fair RRF and the cross-encoder configurations. Overall, dense retrieval gives the best accuracy on this dataset, and additional reranking adds processing time without improving classification performance.

---

### Event-Level Emotion Recognition in the Wild Using Deep Facial Expression Analysis

Authors: Aleksandr Semerikov, Pakizar Shamoi
Published: 2026-09-12
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.13854v1

Facial emotion recognition (FER) in real-world environments remains challenging due to unconstrained imaging conditions, including multiple faces, occlusions, pose variations, and complex lighting. Most existing studies focus on individual facial emotion classification and do not address the analysis of collective emotional states at the event level. This paper proposes an end-to-end pipeline for event-level emotion recognition from photographs. The approach detects faces in each image, classifies facial expressions using a deep convolutional neural network, and aggregates face-level emotion probabilities to estimate the overall emotional distribution of a public event. A comparative evaluation of several CNN architectures on the FER- 2013 and RAF-DB datasets demonstrates that transfer learning with EfficientNet-B2 trained on RAF-DB is more suitable for real-world RGB data. The proposed method is evaluated on a real-world event dataset containing 1658 images. Experimental results show stable emotion distributions across event subsets, confirming the effectiveness of event-level aggregation for emotion analysis in the wild.

---

### ReWeight: Leveraging Human Data for VLA Post-Training via Demonstration Retrieval and Sample Weighting

Authors: Chenwei Wang, Dianye Huang, Match W. L. Ko, Chenjia Bai, Zhongliang Jiang
Published: 2026-09-12
Categories: cs.RO, cs.AI
arXiv: https://arxiv.org/abs/2609.13851v1

Post-training vision-language-action (VLA) models for specific robots and tasks requires in-domain demonstrations, yet collecting diverse robot data is costly. Egocentric human demonstrations provide a scalable alternative, but directly mixing human and robot data can introduce cross-embodiment discrepancies and degrade policy performance. To address this challenge, we introduce ReWeight, a framework that incorporates human data into VLA post-training through demonstration-level retrieval and sample-level weighting. ReWeight learns a cross-embodiment visuomotor representation that combines visual observations with future actions to measure behavioral similarity between human and robot demonstrations. Based on optimal transport, it retrieves human demonstrations relevant to the target robot data and assigns larger weights to samples with smaller cross-embodiment discrepancies. We evaluate ReWeight using $π_{0.5}$ across eight simulation tasks and four real-world tasks under both clean and randomized settings. In simulation, ReWeight improves the average success rate of post-trained $π_{0.5}$ from 39% with only robot data and 44% with randomly mixed human-robot data to 57%. In the physical experimental setting, it achieves an average success rate of 68.8%, outperforming the baselines by 28.8% and 13.8%, respectively. Overall, ReWeight provides an effective paradigm for transforming abundant egocentric human experience into transferable supervision for robot learning. (Project webpage: https://reweight-vla.github.io/)

---

### UniCAR-RL: Seeing Better before Thinking Deeper in Visual Mathematics

Authors: Yuzhe Li, Hao Yan, Hao Wang, Xingchen Liu, Ya-Qi Yu, Jihao Wu, Minghui Liao, Wei Chen, Yuliang Liu
Published: 2026-09-12
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.13849v1

Multimodal Large Language Models (MLLMs) often struggle with complex mathematical visual reasoning primarily due to a lack of fine-grained perception, causing initial visual hallucinations to directly trigger cascading reasoning failures. In traditional end-to-end reinforcement learning (RL), sparse rewards fail to decouple perceptual hallucinations from logical missteps, hindering targeted perception optimization. Alternatively, fine-tuning with perception-enhanced CoT data incurs high costs and hallucinations. In this paper, we address these challenges by proposing UniCAR-RL, an annotation-free RL framework. By explicitly decoupling the optimization of perception and reasoning during the training process, it achieves isolation and optimization of both capabilities. Specifically, UniCAR-RL consists of three synergistic branches: 1) a Caption-RL branch that optimizes perception capabilities through verifier-guided reasoning validation; 2) a Reasoning-RL branch that performs logical reasoning based on a gold image description to halt cascading errors; 3) a QA-RL branch that retains native end-to-end alignment to ensure robust question-answering performance. Experiments show that UniCAR-RL substantially improves MLLMs' mathematical and visual reasoning using only raw short-answer data. Furthermore, it demonstrates strong generalization across diverse architectures and scales.

---

### Measuring the Cost of Variety Conflation in Multilingual MT Evaluation: Adding Mozambican Xichangana, Nyanja and Sena to FLORES+

Authors: Felermino D. M. A. Ali, Delfina Lázaro Mateus, Manuel Valente Mangue
Published: 2026-09-12
Categories: cs.CL
arXiv: https://arxiv.org/abs/2609.13847v1

In this paper, we extend FLORES+ with Portuguese-source evaluation sets for three Mozambican Bantu varieties: Xichangana, Mozambican Nyanja, and Sena. We compare Xichangana with the existing Tsonga reference and Mozambican Nyanja with Chichewa, and evaluate NLLB-200, Google Translate, GPT, and a variant-aware NLLB model. Holding system output fixed reveals substantial reference sensitivity. On \textit{devtest}, changing only the reference from Tsonga to Xichangana reduces spBLEU by 13.10 points for NLLB-200 and 15.30 for Google. On matched Nyanja subsets, replacing Chichewa with Mozambican Nyanja produces smaller but consistent reductions of 3.03 and 6.10 spBLEU, respectively. Variant-aware fine-tuning reverses this pattern on the intended targets: relative to NLLB-200, it improves Xichangana by 7.04 spBLEU and Mozambican Nyanja by 5.33 on \textit{devtest}, while losing performance on the sibling references. GPT is competitive on Tsonga and Chichewa but substantially weaker on the Mozambican varieties. For Sena, the finetuned model reaches 12.64 spBLEU and 36.21 chrF++ on \textit{devtest}. These findings motivate variety-aware language identifiers, references, and reporting for cross-border languages or language dialects/variants. The data is publicly available on Hugging Face at https://huggingface.co/datasets/MOZNLP/FLORES_MOZ

---

### Affinity-Aware Sharding for Delayed Tensor Parallelism

Authors: Eloi de Reynal
Published: 2026-09-12
Categories: cs.LG, cs.CL, cs.DC
arXiv: https://arxiv.org/abs/2609.13846v1

Delayed Tensor Parallelism (DTP) removes the blocking all-reduce of tensor-parallel Transformer inference. Every device adds its own partial output to its residual stream (and broadcasts it) immediately, but only gathers (receives) the other devices' partials $δ$ modules later. A TP to DTP change therefore amounts to a real architecture change, and dense Transformer models need to be retrained or distilled after adaptation. We show that DTP breaks the permutation symmetry of neurons inside FFNs and of KV heads inside attention modules, and that this symmetry breakage makes the sharding itself a modelling decision. We show that maximising the affinity between the KV heads and the FFN neurons co-located on a device, by permuting the dense model before sharding, speeds up the distillation or retraining process. The affinity is measured with a first-order approximation of the damage that losing a head's contribution does to each neuron's output, and the co-located affinity is maximised with a coordinate-ascent optimiser that alternates an exact balanced assignment of neurons with an exhaustive search over the KV head partitions. The whole procedure takes under two minutes on one GPU for Qwen3-0.6B and Danube3-500M. On these models, at $δ=1$, the affinity-optimised layouts reach any distillation target in about half to two thirds of the steps needed by the naive contiguous layouts, over the whole 10k-step range we tested, and every optimised seed beats every contiguous seed and all but one of the sixteen random layouts. We also show that the co-located affinity score at initialisation predicts the KL to the base model after training, across seventeen layouts ranging from anti-optimised to optimised (Pearson $-0.81$ and $-0.89$).

---

### LePlanner: An Iterative Amortized Controller For World Models

Authors: Saksham Bansal, Om Naphade, Chayan Aggarwal, Vrishin M
Published: 2026-09-12
Categories: cs.RO, cs.AI
arXiv: https://arxiv.org/abs/2609.13845v1

World models trained with joint-embedding predictive architectures learn compact, structured latent representations from physical interaction, yet planning in these latent spaces typically relies on one of two costly approaches. Search-based planners such as CEM, MPPI, and iCEM optimize action sequences through many predictor rollouts, achieving strong performance at the cost of high per-decision compute and latency. Policy-based methods amortize inference into a single forward pass but can degrade on contact-rich tasks where the demonstration distribution is multimodal. We propose LePlanner, an amortized iterative controller that learns to construct and refine latent action sequences through a frozen world-model predictor. LePlanner is trained with an arrival-and-hold objective that encourages the controller to reach the goal at the earliest feasible horizon and remain there. This addresses horizon-reset procrastination, a failure mode in which repeated receding-horizon replanning continually postpones goal arrival. An additional action-Gaussian loss keeps generated actions near the support of the offline dataset. Across navigation, contact-rich manipulation, and continuous-control environments, LePlanner matches or exceeds search-based planners while requiring an order of magnitude fewer predictor evaluations and 3-49x lower wall-clock time per decision. It achieves success rates of 98% on PushT, 100% on Reacher, 100% on TwoRooms, and 92% on the OGBench Cube task. These results show that much of the structure discovered through online search can be amortized into a lightweight iterative policy, enabling fast, horizon-aware, nonlinear physical control without online optimization.

---

### Pre-training with Graph Transformers

Authors: Jiaming Wang, Thomas Laurent, Xavier Bresson
Published: 2026-09-12
Categories: cs.LG
arXiv: https://arxiv.org/abs/2609.13844v1

This article investigates pre-training strategies for graph transformers in the biochemistry domain. By conducting comprehensive experiments, the study reveals that supervised pre-training using computed properties as labels provides the highest performance gain on downstream tasks. The results also highlight the importance of constraining model capacity to mitigate overfitting in graph transformers.

---

### CRAF: Cross-View Residual-Aware Fusion for Deepfake Speech Detection

Authors: Minh-Xuan Phan, Khalid Zaman, Candy Olivia Mawalim, Masashi Unoki
Published: 2026-09-12
Categories: cs.SD, cs.AI, eess.AS
arXiv: https://arxiv.org/abs/2609.13842v1

Recent advances in speech synthesis and voice conversion have made deepfake speech increasingly realistic, making generalization to unseen spoofing attacks a critical challenge. Pretrained speech and audio models offer a promising direction for improving robustness to such unseen attacks. Self-supervised learning (SSL) models capture fine-grained, low-level acoustic characteristics, whereas Auditory Large Language Models (ALLMs) provide higher-level contextual representations. These complementary views can provide useful cues for improving generalization to unseen attacks. However, direct fusion does not explicitly disentangle information shared across the two views from view-specific complementary information, limiting effective cross-view integration. To address this, we propose CRAF, a cross-view residual-aware fusion framework that uses ALLM-guided cross-view attention to enrich SSL representations and adopts ALLM as a high-level reference to separate ALLM-explainable information from complementary SSL residual information. The residual is selectively refined through adaptive gating and integrated through SSL-primary fusion. Experiments on ASVspoof 5 show that CRAF with Kimi-Audio achieves an EER of 5.96% and a minDCF of 0.1192, demonstrating robustness to unseen spoofing attacks.

---

### Sweet Talkers: How Query Formulation Shapes Sycophancy in Romantic Relationship Advice

Authors: Helena Choi, Edric Castel Hao, Karl Bautista, Francis Gabriel Magleo, Renzo Panti, Danielle Beatrice Olalia
Published: 2026-09-12
Categories: cs.CL
arXiv: https://arxiv.org/abs/2609.13841v1

Large language models (LLMs) are increasingly used for emotional support and relationship advice, where a model's tendency to preserve a user's face can inadvertently reinforce harmful interpersonal behaviors. To systematically examine this risk, we developed the Romantic Relationship Advice-Seeking Prompts (RRASP) dataset of 2,400 prompts across five relationship themes and evaluated social sycophancy using the ELEPHANT framework on two consumer-facing models, GPT-5 Mini and Gemini 3 Flash. Contrary to our initial hypothesis, grammatical mood alone did not produce systematic differences in sycophantic behavior, suggesting that what a user implies matters more than how they phrase it. Instead, perspective-driven framing had a stronger influence, with gaps between original and flipped prompts widening in follow-up responses. Consistent increases in framing and moral sycophancy across turns indicate that models become more likely to accept a user's stated premises and affirm their ethical stance as a dialogue progresses. Notably, Gemini 3 Flash exhibited substantially smaller increases in moral sycophancy than GPT-5 Mini, suggesting it is more resistant to reinforcing ethically problematic positions across turns.

---

### Accuracy Is Not Service: A Decision-Aware Benchmark for Intermittent-Demand Forecasting

Authors: Joo Ern Chin, Shih-Fen Cheng, Aldy Gunawan
Published: 2026-09-12
Categories: cs.LG
arXiv: https://arxiv.org/abs/2609.13840v1

A contract-logistics spare-parts operator is paid on order-level service: an order counts only if every requested line is fulfilled, yet forecasters are selected based on line-level forecast accuracy. This disconnect matters when demand is intermittent and lumpy, histories are short, and lead times span months. We benchmarked 38 forecasting methods spanning classical, intermittent-demand, machine-learning, deep-learning, and pretrained foundation models. A common decision-aware protocol evaluates them on an industrial panel drawn from a live contract and two public datasets. Forecast-accuracy rank and order-service rank are negatively correlated on the industrial panel, at -0.555, across methods evaluated on 20,330 real multi-item orders. Service is more closely associated with the direction of cumulative forecast bias, including over-prediction during zero-demand periods, than with point accuracy. Examining bias in Chronos-2's instance normalization yields a training-free correction that lifts the per-material fill proxy from 77.5% to 92.0% (14.5 percentage points) at the 90% policy target and raises the complete-order fill rate from 54% to 63%. For reproducibility, we release RUF (Regenerate-Until-Fidelity), a method for generating fidelity-certified synthetic panels on which the findings reproduce. For intermittent demand, the lowest-error forecast need not deliver the highest service. Bias direction helps explain this gap, which can be reduced without retraining.

---

### TotalSynth: Robust Whole-Body Synthetic CT from MRI and CBCT

Authors: Valentin Boussot, Cedric Hemon, Anais Barateau, Caroline Lafond, Jean-Claude Nunes, Jean-Louis Dillenseger
Published: 2026-09-12
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.13838v1

Purpose: To develop and evaluate TotalSynth, a reusable pretrained framework for whole-body synthetic CT (sCT) generation from MRI and cone-beam CT (CBCT) images. Materials and Methods: In this retrospective technical study, the dataset was assembled between 2020 and 2026 from SynthRAD challenge data, four prostate cohorts, and BIC-MAC. After registration quality control, 1450 of 1800 public challenge pairs were retained; 350 were excluded for insufficient registration quality or major source/CT mismatch. The corpus also included 84 additional prostate MRI/CT and 60 external BIC-MAC MRI/CT cases. Three 5-fold model families were evaluated with image-domain, anatomy-aware, registration-based, and uncertainty metrics. Age and sex were not consistently available across public datasets. Results: The released MRI-to-CT model achieved an overall MAE of 67.49 HU, SSIM of 0.920, and PSNR of 29.28 dB. The CBCT-to-CT model achieved an overall MAE of 53.55 HU, SSIM of 0.939, and PSNR of 32.09 dB. The unified model maintained similar performance on MRI inputs (MAE, 67.68 HU) and CBCT inputs (MAE, 54.22 HU). On external BIC-MAC data, MRI-to-CT MAE was 100.91 HU without fine-tuning and 62.21 HU after fine-tuning. Conclusion: TotalSynth provides reusable MRI- and CBCT-based CT synthesis models with broad anatomical coverage, while external evaluation highlights the need for local validation and optional fine-tuning under domain shift.

---

### Graph Neural Networks for Influence Maximization in Social Networks: An Unsupervised Minimum Dominating Set Approach

Authors: Erfan Ahmadi, Mina Shirazi, Behnam Bahrak
Published: 2026-09-12
Categories: cs.LG, cs.SI
arXiv: https://arxiv.org/abs/2609.13836v1

The Minimum Dominating Set (MDS) problem is a classic NP-hard combinatorial optimization problem with critical applications in social network analysis, including viral marketing, influence maximization, public health interventions, and information dissemination. Identifying a minimal set of influential individuals whose reach covers an entire social network is central to these applications, yet remains computationally challenging at scale. Graph neural networks (GNNs) have emerged as powerful tools for learning over graphs, and recent work explores their application to hard combinatorial problems. This paper presents a novel unsupervised GNN framework for the MDS problem that eliminates the need for ground-truth solutions during training. Trained on 12,000 synthetic graphs with diverse structural properties, our method achieves up to 55x faster inference than metaheuristic baselines and up to 14x faster inference than supervised learning approaches, while finding optimal or near-optimal dominating sets on real-world social network benchmarks. Our learned heuristic generalizes effectively to unseen graph distributions, demonstrating strong practical applicability for large-scale social network analysis.

---

### DiVA: Enabling Interactive Digital Life Simulation via Video Models

Authors: Cheng Chen, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Wen Wang, Yihao Meng, Hanlin Wang, Yixuan Li, Jiacheng Wei, Zhenshan Tan, Yanhong Zeng, Yujun Shen et al.
Published: 2026-09-12
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.13830v1

We present DiVA, a deeply interactive digital life simulator pioneering a new paradigm for long-term, open-ended interactive experiences within digital character worlds. DiVA's architecture pairs a Multimodal Large Language Model (MLLM) as a router with a meticulously designed stacked video pipeline for seamless, multi-turn interactions with action and audio response. To maintain continuity and avoid degradation, we model generation as a three-part coupled system: waiting video, action video, and the transitions between them. These transitions are critically handled by our Anchored Video Continuation (AVC) module, which returns the character to stable states to prevent degradation. By encoding information from the preceding action video segment, AVC ensures smooth transitions, significantly reducing camera jitter and inconsistencies common in current video transition methods. This design also enables complex pose changes (e.g., sitting to standing) typically difficult for audio-driven models. These system designs together ensure high-fidelity identity, coherence, and dynamics for extended experiences. To validate our pipeline design, we comprehensively compare our system against alternatives by replacing our core generation module with mainstream long-video, continuation, and interpolation methods. We further analyze the necessity of the three-stage design, anchor-state selection, transition naturalness, spatial grounding, and the quality-latency trade-off, and we expand the comparison to additional long-form audio-driven avatar models. Results confirm DiVA is markedly superior in maintaining long-term visual quality and realism, validating its effectiveness as a sustainable, interactive simulation.

---

### ViperQ: Order Flow Pattern Recognition via Auction Market Theory for Reinforcement Learning Trading

Authors: Asser Moustafa, Rares-Mihail Neagu, Jugal Kalita
Published: 2026-09-12
Categories: cs.AI, q-fin.TR
arXiv: https://arxiv.org/abs/2609.13825v1

Reinforcement learning trading systems published in the academic literature overwhelmingly rely on price-aggregate state representations (OHLCV bars) or limit-order-book depth features, leaving microstructure pattern theories from the practitioner literature, namely Auction Market Theory and Market Profile, without a peer-reviewed computational instantiation. We present ViperQ, a reinforcement learning system whose state representation is built explicitly from Auction Market Theory primitives: Volume Point of Control, Value Area position, Low Volume Node flags, Cumulative Volume Delta divergence, and tape-velocity signatures, assembled into a 20-dimensional Z-normalised vector. Two Proximal Policy Optimisation agents are trained with a prospect theory-grounded asymmetric reward function that penalises losing holds at a magnitude consistent with Kahneman and Tversky's loss-aversion coefficient. Evaluated on a held-out twelve-month partition of institutional tick data the agents have never seen, ViperQ achieves +163.6% ROI on TSLA (-27.5% max drawdown, 27,019 trades) and +116.5% ROI on NVDA (-47.8% max drawdown, 12,892 trades) under zero leverage. The results establish Auction Market Theory features as a tractable structured input modality for sequential decision-making on financial time series and motivate further work on microstructure-aware policy learning.

---

### When Consistency Does Not Mean Reliability: Evaluating Local LLM Judges Against Human Ratings

Authors: Aakash Kumar Tiwari
Published: 2026-09-12
Categories: cs.CL
arXiv: https://arxiv.org/abs/2609.13824v1

Large language models (LLMs) are increasingly used to evaluate the responses of other language models. This approach, known as LLM-as-a-Judge, is faster and cheaper than human evaluation. However, a judge may produce consistent scores without necessarily agreeing with human evaluators. In this work, we study this issue using two local open-weight LLM judges, LLaMA-3-8B and Qwen2.5-7B. We evaluate 300 responses generated by an instruction-tuned GPT-2 (124M) model for 100 questions covering five categories: factual knowledge, instruction following, mathematics, reasoning, and writing. Each response is scored by nine human annotators and is evaluated three times by each LLM judge using the same rubric. We compare the judge scores with the average human scores using Pearson correlation, Spearman correlation, mean absolute error (MAE), signed bias, and self-consistency. LLaMA-3-8B shows a Pearson correlation of 0.275 with human scores, while Qwen2.5-7B achieves 0.340. Their MAEs are 27.71 and 18.64, respectively. Despite this limited agreement, both judges show high self-consistency, with exact consistency rates of 97.3\% for LLaMA-3-8B and 92.3\% for Qwen2.5-7B. These results show that high self-consistency does not necessarily indicate high agreement with human judgments. Our findings highlight the need to evaluate both consistency and human alignment when using local LLMs as automatic judges.

---

### Semantic Privacy Protection with Utility Preservation for 3D Point Clouds

Authors: Jinchang zhang, Jiakai Lin, David Crandall, Guoyu Lu
Published: 2026-09-12
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.13823v1

Point cloud data face serious semantic privacy risks during acquisition, transmission, and cross-institutional sharing. Existing methods mostly rely on geometric perturbation or destructive encryption, which can reduce the recognizability of the original class but often impair downstream usability. This paper proposes a class-transfer-based semantic encryption framework for point clouds, aiming to conceal original class information while preserving task utility and supporting authorized recovery. Specifically, we construct a unified latent space with a shared-backbone Normalizing Flow, and combine LoRA and FiLM to achieve parameter-efficient class-conditional adaptation. We further introduce diffusion-guided flow alignment to regularize the latent distribution, construct an energy-based category transition graph, and obtain an optimal class-transfer table through global matching. Then, a latent-space Neural ODE continuously evolves source-class latents into target-class latents, which are decoded into target-class point clouds through the inverse flow. We adopt attacker-oriented metrics, including New-Class Recognition Rate (NCRR), Original-Class Leakage Rate (OCLR), and Original Label Recovery Rate (OLRR), to evaluate privacy and utility. Experiments on classification and segmentation benchmarks show that the proposed method achieves controllable semantic transformation, effectively reduces original-class semantic leakage, preserves downstream learnability in the protected domain, and supports reliable authorized reconstruction.

---

### Benchmarking Optimizers to Solve Inverse Problems with Differentiable Physics Simulators

Authors: Xiang Chen, Huanhuan Xia
Published: 2026-09-12
Categories: cs.LG, math.OC, physics.comp-ph
arXiv: https://arxiv.org/abs/2609.13819v1

Solving inverse problems with differentiable physics simulators holds the potential to revolutionize scientific discovery and engineering design, as it enjoys both the strict physical correctness from rigorous numerical physics simulators, and the high efficiency and effectiveness from automatic differentiation and gradient-based optimization. However, currently, this paradigm faces performance issues in optimization. In this work, we target benchmarking the performance of different optimizers to solve various inverse problems. We construct 12 differentiable physics simulators spanning physics domains including discrete mechanics, continuous mechanics, atomistic simulations, rendering, and semi-empirical physics models. Based on these simulators, we design corresponding inverse problems that can be categorized into parameter identification, inverse design, and optimal control. Finally, we conduct extensive experiments to compare the performance of different optimizers, including regular first-order methods, approximate second-order methods, as well as global optimizers, on these inverse problems, and analyze the results to provide insights on how to choose and design optimizers for differentiable programming. We hope such benchmarks can inspire the development of more effective optimizers, and further promote the applications of differentiable programming in various scientific and engineering domains.

---

### Exploring Automated Vulnerability Identification in JavaScript Code Using Large Language Models

Authors: Manit Kaushik, Ishir Bhardwaj, Pranav Gupta, Pankaj Jalote, Arun Balaji Buduru
Published: 2026-09-12
Categories: cs.CR, cs.AI
arXiv: https://arxiv.org/abs/2609.13816v1

JavaScript powers approximately 98.8% of all websites, making vulnerabilities in its code a significant security risk, yet existing detection approaches such as Static Application Security Testing (SAST) tools often fail to identify many real-world vulnerabilities when applied to isolated code snippets. This paper presents an empirical study of Large Language Model (LLM)-based vulnerability identification for JavaScript programs, evaluating three LLM families (Gemini 1.5 Flash, GPT-4o Mini, DeepSeek-R1-Distill-Llama-8B) across multiple prompting strategies (zero-shot, chain-of-thought, few-shot) and fine-tuning approaches on a dataset of 1,125 JavaScript code snippets spanning five Common Weakness Enumeration (CWE) categories: Injection (CWE-74), OS Command Injection (CWE-78), Cross-Site Scripting (CWE-79), SQL Injection (CWE-89), and Uncontrolled Resource Consumption (CWE-400). Our experiments show that LLMs substantially outperform traditional SAST tools on snippet-level vulnerability identification, with a fine-tuned Gemini 1.5 Flash model achieving 60% detection accuracy compared to near-zero performance from rule-based analyzers. We find that fine-tuning improves accuracy from 29% to 60%, Chain-of-Thought prompting benefits reasoning-capable models such as GPT-4o Mini, few-shot prompting is effective for polymorphic vulnerabilities such as Cross-Site Scripting, and performance varies across vulnerability categories, reaching up to 84% accuracy for structured vulnerabilities such as SQL Injection. These results indicate that LLMs provide a practical approach for automated vulnerability identification in JavaScript code, particularly when combined with task-aligned supervision, though they should complement rather than replace existing security analysis workflows due to limited recall and uneven performance across vulnerability types.

---

### Beyond OCR Accuracy: Text-Centric VQA Under Image Degradation with Modular and End-to-End

Authors: Ritali Vatsi, Rachapudi Jagadeesh, Shruti Singh Baghel, Himani Sharma, Amit Shukla, Pawan Goyal
Published: 2026-09-12
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.13815v1

Text-centric Visual Question Answering (VQA) requires reading and reasoning over text embedded in images, a task made substantially harder when images suffer from real-world degradation such as motion blur, low resolution, or compression artifacts. While modular OCR-based pipelines and end-to-end vision-language models are both widely used for this task, their comparative robustness under degraded conditions remains underexplored. We present an empirical study comparing two modular pipelines with SA-DBNet, a custom detector architecture combining ResNet-18 with self-attention spatial modeling and deformable convolutions against an end-to-end vision-language baseline, evaluated on 4013 degraded images with 7000 question-answer pairs. Fine-tuned modular pipelines achieve up to 57.50% exact-match accuracy versus 38.00% for the end-to-end baseline, with domain-specific fine-tuning yielding a gain of up to 29.50 percentage points. Critically, we find that conventional OCR error metrics like Character Error Rate and Word Error Rate are unreliable predictors of downstream VQA performance, as semantic reasoning can compensate for recognition failures when contextual cues are present. These findings highlight the importance of task-aware evaluation for text-centric VQA systems under realistic visual conditions. Codes are available here

---

### Realtime-Venus: A full-duplex interaction system with asynchronous delegation

Authors: Ant Group
Published: 2026-09-12
Categories: cs.CV, eess.AS
arXiv: https://arxiv.org/abs/2609.13814v3

Natural interaction in digital and physical environments requires continuous perception and timely responses. Spoken dialogue relies on acoustic and linguistic cues, while video interaction also requires grounding the conversation in evolving visual context. We present Realtime-Venus, a proactive full-duplex interaction system with two separately trained 9B models: Realtime-Venus-Omni for audio-visual interaction and Realtime-Venus-Audio for spoken interaction. Each model serves as a complete conversational frontend, integrating continuous perception, conversational control, and native speech generation through a shared causal timeline for user inputs, model outputs, and delegation events. A dual-loop runtime coordinates live interaction with background reasoning and tool execution. Foreground interaction continues while Realtime-Venus-Harness executes tasks asynchronously and returns results for integration into the ongoing dialogue. Both models follow a common post-training recipe combining offline understanding, proactive full-duplex trajectories, and delegation workflows. Among the evaluated online models, Realtime-Venus-Omni achieves the highest scores on six of eight video benchmarks, including StreamingBench (70.2%), OVO-Bench (64.7%), and Daily-Omni (81.3%). Across eight audio understanding and spoken question answering benchmarks, Realtime-Venus-Audio leads the compared models on MMAU (78.0%), MMAU-Pro (63.2%), Llama Questions (83.8%), and Speech CMMLU (67.8%), while matching the best VoiceBench AlpacaEval score of 4.81. On Full-Duplex-Bench v1.5, Realtime-Venus-Audio responds to 75% of user interruptions and achieves continuation rates of 97%, 88%, and 86% under backchannels, other-directed speech, and background speech, respectively, exceeding Gemini 3.1 Live and GPT-4o on all three continuation metrics.

---

### Odds-Shift Slippage in One-vs-Rest Rankers: Diagnosing and Repairing Reweighting-Induced Top-K Errors

Authors: Akifumi Goto
Published: 2026-09-12
Categories: cs.IR, cs.LG
arXiv: https://arxiv.org/abs/2609.13810v1

One-vs-rest rankers that show each user the top-$K$ of many rare labels usually counter imbalance with a per-label positive-class weight, scale_pos_weight $= n_-/n_+$. Elkan's identity says such a weight shifts label $j$'s log-odds by $\ln w_j$, so the model ranks by weighted odds rather than by the marginal that is Bayes-optimal for precision@$K$, and suggests inverting the shift afterwards; what a finite learner does with a weight in the thousands, and which repair then works, has not been measured. We call the gap between the promised and the realized shift odds-shift slippage and measure it on matched pairs of LightGBM and MLP models that differ only in the weights. On Santander the weight takes MAP@7 from 0.808 to 0.117; for the boosted pairs the ideal odds shift accounts for 23% of that loss (32% on Instacart; 98% for an MLP pair on the same rows) and slippage for the rest. We prove that a booster whose leaf steps are capped at $c$ realizes at most $Tηc$ nat of shift in $T$ rounds at rate $η$, which a cap sweep confirms, and show that without a cap saturated cells tie at exactly 1.0, beyond the reach of any separable map. The analytic inversion therefore pays only where the shift was realized and nothing saturated, whereas per-label isotonic regression returns the Santander model to 0.784 (0.780 with the calibrator fitted on the validation period), but only if labels without calibration positives are mapped to their prior rather than passed through. On 11 public MULAN benchmarks and 5 learners the weighted model loses more than half of its MAP@$K$ in 8 of 55 cells, and on delicious and Corel5k the same repair returns it to the unweighted level; per-label calibration hurts where positives are scarce, a harm that a cross-validated rule removes. The recipe is released as oddslip.

---

### DARE: Dialectical Agentic Reasoning for Structured Knowledge Fact Checking

Authors: Yifei Li, Xiaohan Zheng, Wentao Qian, Liansheng Zhuang
Published: 2026-09-12
Categories: cs.CL
arXiv: https://arxiv.org/abs/2609.13808v1

Structured knowledge fact checking aims to determine the truthfulness of natural language claims by reasoning over structured evidence. Recent program-generation approaches leverage large language models (LLMs) to generate executable graph reasoning programs, achieving strong performance on structured knowledge fact checking benchmarks. However, these methods remain limited by invalid relation generation, single-path reasoning that lacks self-correction, and biased evidence assessment that tends to overestimate supporting signals. We propose Dialectical Agentic Reasoning (DARE), a multi-agent framework that formulates structured knowledge fact checking as an iterative retrieve-reason-reflect process. DARE integrates relation-grounded evidence retrieval to constrain reasoning to valid structures, dialectical bidirectional verification to evaluate evidence from both supporting and refuting perspectives, and confidence-driven meta-reflection to dynamically determine whether additional evidence exploration is necessary. Extensive experiments demonstrate the effectiveness of DARE in structured knowledge fact checking, with an 8B backbone achieving 88.12% accuracy and matching or surpassing GPT-4o-based program-generation baselines, which attests to the efficacy of dialectical agentic reasoning in eliciting the latent reasoning capabilities of LLMs.

---

### Bypass Observation: A Conceptual Design of a Non-Intrusive Layer-Wise Semantic Extraction Architecture

Authors: Haibin Tong, Jiang Yu
Published: 2026-09-12
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.13807v1

Large language models reason in high-dimensional hidden-state spaces, while users observe only final outputs. We introduce Bypass Observation, a non-intrusive layer-wise readout architecture that attaches read-only observation heads to selected Transformer layers without feeding their outputs back into the backbone. We consider three variants: a shared LM head across layers, layer-specific heads, and a layer- or step-adaptive head. For full-vocabulary readout, we derive a closed-form overhead approximation governed primarily by V/(12d), with representative estimates ranging from about 30% to 240%, and discuss cost reductions via sparse observation, low-rank factorization, reduced vocabularies, top-k readout, and selective positions. We argue that Bypass Observation can make model computation more observable while remaining only a partial, potentially misleading projection of hidden states. We further distinguish bypass chain-of-thought from conventional chain-of-thought: conventional reasoning tokens enter the autoregressive computation, whereas bypass readouts remain causally external at inference time, although they can still provide training signals in reinforcement learning. Finally, we discuss applications to looped and recurrent-depth Transformers, where iteration-wise readout may expose convergence, oscillation, and potential halting signals. The proposal is conceptual and analytical; systematic empirical validation remains future work.

---

### Understanding the Limits of Agentic ICD Coding

Authors: Chong Yock Eng, Yushi Cao, Yiming Chen, Kezhi Mao, Hongchao Jiang
Published: 2026-09-12
Categories: cs.CL, cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2609.13806v1

ICD-10-CM codes are alphanumeric codes used in the US to classify diagnoses and injuries for medical billing and epidemiological reporting. Standard ICD-10-CM benchmarks report aggregate metrics that obscure performance on complex coding scenarios. We evaluate neural, workflow, and agentic systems on a rarity-stratified set of MIMIC-IV discharge summaries and identify two orthogonal failure modes. Neural classifiers exhibit a 0.43 micro-F1 gap between rare and common codes. Workflow systems handle rare codes well but score near zero on injury and external cause codes that require multi-step guideline following. A tool-augmented agentic configuration with structured access to official ICD-10-CM reference materials recovers up to 0.34 micro-F1 on this subset. No single system dominates across all conditions.

---

### LLM-Enhanced Multi-Agent Reinforcement Learning for Unified Electric Vehicles-Charging Station-Grid Optimization in Public Charging Systems

Authors: Yang Zhang, Lindong Xie, Chongyu Wang, Gaojunjie Li, Siqi Bu, Edward Chung
Published: 2026-09-12
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.13805v1

In the era of the Internet of Things (IoT), coordinating connected electric vehicle (EV) charging scheduling to balance EV charging satisfaction, station profitability, and smart grid stability presents a complex multi-objective challenge. Existing Multi-Agent Reinforcement Learning (MARL) approaches often struggle with high-dimensional state spaces generated by massive IoT sensing data and conflicting stakeholder interests. This paper proposes a novel LLM-enhanced MARL framework that, for the first time, simultaneously optimizes the Grid, EVs, and Stations within a unified loop. By integrating Large Language Model (LLM), we address two critical bottlenecks: interpretable feature selection and adaptive multi-objective balancing. The LLM analyzes real-time IoT-collected environmental states to extract physically significant features and dynamically assigns weights to conflicting objectives-including profit, user satisfaction, and grid load-using semantic reasoning instead of complex manual tuning. Extensive experiments demonstrate that our framework significantly outperforms state-of-the-art baselines, achieving superior market efficiency while reducing training time by over 70%. This approach offers a scalable, transparent solution for efficient and sustainable IoT-enabled urban charging infrastructure management.

---

### StepPrune: Adaptive Sequential Visual Token Selection across Multimodal Large Language Models

Authors: Hansen Zhang, Landi He, Mingde Yao, Lijian Xu
Published: 2026-09-12
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.13804v1

Visual prefixes account for a major portion of the per-layer computation in multimodal large language models (MLLMs), making visual-token pruning a direct approach to accelerating inference. Existing top-K methods typically evaluate tokens independently and apply a uniform budget to all inputs, overlooking both selection-dependent interactions and variations in visual complexity across samples. In contrast, we propose StepPrune, which formulates visual-token pruning as an adaptive sequential decision process. Conditioned on previously selected tokens and textual context, StepPrune progressively constructs the retained subset and automatically determines its size through a learned STOP action. During training, a variance-preserving noise gate provides a differentiable surrogate for the discrete selection process, whereas during inference, unselected tokens are physically removed before language-model prefill. A grouped selection mechanism further extends StepPrune to high-resolution inputs. Experiments across LLaVA-1.5, LLaVA-NeXT, Qwen2.5-VL, and InternVL3 show that StepPrune achieves the best average normalized performance retention across all evaluated pruning rates on LLaVA-1.5, Qwen2.5-VL, and InternVL3, while remaining competitive on the substantially longer AnyRes prefixes of LLaVA-NeXT. On LLaVA-1.5, StepPrune retains 94.6% of the full-prefix normalized performance while pruning 88.9% of the visual tokens. At a mean retained count of 64, StepPrune reduces prefill latency from 59.95 ms to 40.05 ms, corresponding to a 1.50x prefill speed-up.

---

### Do Not Restart: Residual Completion for Stateful Agent Handoffs

Authors: Runzhi Deng, Yiming Zhong, Fang Zhao, Pan Zhou
Published: 2026-09-12
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.13800v4

Routing and cascades reduce tool-agent cost by transferring control across models, but stateful handoffs must preserve accepted choices, realized effects, and unfinished obligations. We formulate this as commitment-constrained residual completion and introduce Commitment-Frontier Residual Completion (CFRC). CFRC enforces target-before-proposal, whole-proposal-before-authority, and live-evidence-before-success: it freezes a residual contract from accepted progress, closes the successor continuation into an evidence-linked graph, and admits execution only when the remainder is covered, with live receipts discharging obligations. We establish contract-relative partial correctness, which extends to the original residual request under complete contract construction. Across five environments and two same-provider model pairs, CFRC achieves comparable macro accuracy to strong full-task agents at 22.0% to 34.6% mean per-surface cost, with additional cross-provider results demonstrating broader transfer.

---

### Resolution-Independent Analysis of Encoder--Decoder Operator Learning via Limiting Kernels

Authors: Lei Shi, Jia-Qi Yang, Ding-Xuan Zhou
Published: 2026-09-12
Categories: stat.ML, cs.LG, math.FA, math.NA
arXiv: https://arxiv.org/abs/2609.13798v1

Operator learning is formulated on function spaces, but training data are typically available only through finite-dimensional representations. In encoder--decoder architectures, a matrix-valued kernel on the encoded space induces an operator-valued kernel on the original function spaces, and the corresponding reproducing kernel Hilbert spaces are isometrically isomorphic. As the input and output resolutions increase, the induced kernels converge to a limiting kernel, in the sense of operator-norm convergence of their associated integral operators, allowing regularity assumptions to be stated independently of the encoding resolution. For regularized stochastic gradient descent, we establish upper bounds for decreasing and fixed step sizes, separating the encoding and regularization terms from optimization terms of order \(t^{-θ}\) and \(T^{-θ'}\), respectively, for any \(θ,θ'\in(0,1)\). We further prove lower bounds showing that these encoding-induced terms are generally unavoidable. The analysis is further extended to encoder--decoder neural networks through the limiting neural tangent kernel (NTK), yielding error bounds with an additional finite-width term and polynomial parameter and sample complexity guarantees when the encoding errors decay algebraically. The framework covers matrix-valued kernels constructed from radial and dot product kernels, NTKs arising from wide encoder--decoder neural networks, and encoder--decoder pairs based on Fourier, Legendre polynomial, wavelet, PCA, or pointwise sampling representations.

---

### SyRHM: Symbolic-Language-Enhanced Reasoning with Associative Retrieval for Zero-shot Harmful Meme Detection

Authors: Hanling Wang, Chenlong Wei, Yingjuan Li, Di Wu, Yuchao Zhang, Xiaohui Zhu, Yao Zhu
Published: 2026-09-12
Categories: cs.CL
arXiv: https://arxiv.org/abs/2609.13794v1

Detecting harmful memes is critical for maintaining safe online communities. However, harmful intent is often implicit, arising from visual-textual incongruity and cultural stereotypes, which challenges existing multimodal detectors. We propose SyRHM, a framework that decomposes harmful meme detection into meaning-grounded retrieval and symbolic-language-enhanced multi-stage reasoning. SyRHM retrieves semantically related memes by parsing multimodal content into textual elements and descriptions, providing grounded context beyond surface-level similarity. Building on the retrieved context, SyRHM uses a translator stage to convert multimodal inputs into symbolic intermediate representations, and then performs multi-stage reasoning via planner and solver stages, enabling expressive and interpretable analysis of harmful intent. Experiments on FHM, HarM, and MultiOff demonstrate the effectiveness of SyRHM, achieving superior performance on most evaluation settings against multimodal and reasoning-based baselines, while providing reasoning traces for harmful content. The code is available at: https://github.com/Scabbards1500/SyRHM

---

### Restore What Matters: Lessons from Joint Restoration and Recognition

Authors: Lanqing Guo, Xijun Wang, Minchul Kim, Yu Yuan, Wes Robbins, Xingguang Zhang, Nicholas Chimitt, Stanley H. Chan, Zhangyang Wang, Xiaoming Liu
Published: 2026-09-12
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.13791v1

Recognition pipelines typically adopt a restore-then-recognize workflow, yet decades of experience show that generating visually pleasing images seldom translates to improved recognition. We propose a Joint Restoration-for-Recognition (JR$^2$) paradigm: restore only what downstream tasks truly require, with task signals dictating where, how much, and whether restoration is necessary. JR$^2$ rests on three pillars: (i) Physics, employing optics-accurate turbulence simulation, extensible to blur and noise, to ground restoration in real image formation; (ii) Neuroscience, drawing on selective attention and neuroplasticity to direct model capacity toward identity-critical regions and frames while bypassing already-clean inputs; and (iii) Vision & Learning, coupling recognition loss end-to-end through restoration and alignment so that low-level edits maximize high-level identity stability. Evaluations on IARPA-BRIAR show consistent improvements (e.g., TAR@0.01% FAR +0.6; FNIR@1% FPIR -2.5), while a quality gate skips ~70% of clean frames, reducing cost. Ablations confirm physics priors enhance realism, joint training prevents catastrophic forgetting, and selective restoration suffices in many cases. We conclude that better-looking images are neither necessary nor sufficient; restoration modules must be task-driven, selective, and physically aware. Code, pretrained models, and recipes are provided for integration.

---

### PPDL: A Real-world Industrial User Retention Ratio Forecasting Framework Integrating Physical Priors with Deep Learning

Authors: Zibo Zhao, Zhengxiong Guan, Chaoli Zhang, Linyuan Geng, Xuanbing Zhu, Zhonglong Zheng, Fan Wu
Published: 2026-09-12
Categories: cs.LG
arXiv: https://arxiv.org/abs/2609.13789v1

In multi-channel paid user acquisition, early and accurate prediction of user retention at the channel level is crucial for optimizing budget allocation. User retention curves display a pronounced temporal pattern: an initial period of high churn transitions into long-term stability. This pattern is further characterized by regular fluctuations attributable to seasonality and exhibits high serial autocorrelation. These intrinsic properties make such curves highly suitable for analysis within a time-series forecasting framework. However, forecasting user retention ratio for large-scale short-video platform faces three major challenges: significant heterogeneity across channels, pronounced global trend of decay followed by saturation, and short look-back windows. To address these challenges, we propose PPDL, a novel forecasting framework that integrates physical priors with deep learning. We first introduce a trend-residual decomposition component. The trend is modeled using the Weibull distribution, whose parameters are learned via a Multilayer Perceptron (MLP). Secondly, for the residual component, we design an auxiliary embedding module on top of a deep learning backbone to maintain the channel identity awareness. Finally, to enhance the model's sensitivity to trends, we design a Multiscale Trend-penalized loss function. The proposed approach PPDL is validated through comprehensive experiments on industrial-scale datasets, covering three applications with an average of 30+ channels each. Experimental results show that PPDL achieves improvements across different backbones and significantly outperforms existing online solutions.

---

### Partition Scores Are Not System Scores: Deployment-Fidelity Gaps in Decomposed Algorithm Selection

Authors: Jiachen Zhang, Yu Tang, Li Zhu
Published: 2026-09-12
Categories: cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2609.13785v1

Oracle-style quantities, including virtual best solvers, selected-portfolio VBS, virtual-best encodings, and best-in-family summaries, are widely reported as upper bounds on what a deployable selector could achieve. In decomposed algorithm selection, an analogous partition-level score grants an oracle choice of the best algorithm within the selected family; once the family selector is fixed, the deployable system must replace that within-family oracle with a learned within-family selector. We define the deployment-fidelity gap G(R) as the difference between partition-level and deployable end-to-end utility and derive two accounting consequences: a per-instance margin-regret stability condition that tells us when a partition-time family choice is deployment-optimal, and a sharp partition-only identification interval that, when it strictly crosses zero, prevents the partition-level report from certifying the deployable winner. Across five public algorithm-selection benchmarks spanning tabular AutoML and combinatorial CSP/SAT, every decomposed pipeline has positive G(R), ranging from 0.012 on TabZilla to 0.13 on PROTEUS-2014. Four of ten decomposed-versus-flat decisions have sign-changing point estimates; on PROTEUS-2014, a 33-point partition advantage shrinks to a 20-point end-to-end advantage. A training-side validation gap-correction diagnostic recovers the point-estimate deployable sign on all four sign-changing cells; it is a reporting aid, not a substitute for direct end-to-end evaluation. Partition and end-to-end scores should be reported side by side.

---

### GEAR: From Dynamic Encoding to Dynamic Activation in Social Trajectory Prediction

Authors: Jiaheng Chen, Jiaxing Li, Leixia Wang, Jianan Ju, Tinghe Zhang
Published: 2026-09-12
Categories: cs.CV, cs.AI
arXiv: https://arxiv.org/abs/2609.13778v1

Human trajectory prediction requires modeling both individual motion patterns and social interactions among agents. Existing methods have made substantial progress by using attention mechanisms, graph structures, and temporal encoders to capture dynamic social context. However, most of them primarily focus on how social information is encoded, while paying less explicit attention to how the encoded social context should take effect during future trajectory generation. In this paper, we argue that dynamic social encoding does not necessarily imply dynamic social activation. The same interaction context may require different activation strengths across future horizons and scene densities: social cues should be strengthened when interaction evidence is strong, but suppressed when they are weak or noisy. To address this issue, we propose GEAR, a generation-aware bias activation model for human trajectory prediction. Built upon a bias-decomposed trajectory generation formulation, GEAR dynamically activates the individual-motion and social-resonance bias terms at each future step before final trajectory composition. This allows the model to explicitly control when and how strongly individual and social bias components participate in generation. Experiments on ETH-UCY, SDD, and NBA show that GEAR consistently improves the resonance-based baseline and achieves competitive state-of-the-art performance. Further analyses of activation patterns and density-grouped errors validate the importance of calibrating encoded social context during trajectory generation. Our code is available at https://github.com/11isnotavailable/GEAR.git.

---

### When Do Learned Priors Help Visual Inertial Estimation? A Controlled Study of Prior Integration, Calibration, Initialization, and Backend Consistency

Authors: Jinchang Zhang, Guoyu Lu
Published: 2026-09-12
Categories: cs.RO, cs.CV
arXiv: https://arxiv.org/abs/2609.13777v1

Learned components are increasingly integrated into geometric visual--inertial estimators to provide motion, depth, bias, uncertainty, or confidence cues. Yet it remains unclear whether gains arise from useful learned priors or from changes in the backend, calibration, initialization, temporal association, or evaluation gauge. We present a controlled framework for learning-augmented visual--inertial estimation that separates fusion gain from the incremental value of a learned prior and evaluates four evidence layers: local motion consistency, global trajectory accuracy, physical-state correctness, and numerical consistency. We instantiate the framework with a MonoViT-based monocular motion prior added as a local relative-motion factor to an unchanged VINS backend. We compare Original VINS and learned-prior VINS under matched sensor streams, timestamps, initialization, frontend/backend settings, and camera--IMU extrinsics, while probing calibration, initialization, state coupling, scale, bundle adjustment, and loop closure. On KITTI, with fixed reference extrinsics, translation APE RMSE is 31.4 m for Original VINS and 31.8 m with the learned prior. Across four recordings, the prior changes mean APE by only -0.2%, while a five-times-higher weight worsens it by 8.2%. Online extrinsic updates increase mean APE by 45.7% and 52.1%, respectively, while mean RPE changes by less than 2%. These results show that fusion performance alone cannot establish the value of learned priors. Reliable evaluation requires same-backend controls and joint analysis of prior compatibility, calibration, initialization, global drift, physical-state error, and backend consistency.

---

### Surprising Effectiveness of Self-Demonstrations in Enhancing Schema-Ontology Mapping with LLMs

Authors: Siddhesh Thombre, Manasi Patwardhan, Sunita Sarawagi
Published: 2026-09-12
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.13776v1

Integrating heterogeneous relational databases into a centralized ontology remains a persistent challenge in enterprise knowledge representation, primarily due to semantic heterogeneity, cryptic schema naming, missing metadata, and the abstraction gap between relational schemas and ontological models. Although large language models (LLMs) offer strong semantic reasoning capabilities, we show that directly applying them through one-shot prompting or naive multi-stage pipelines leads to poor performance for schema-ontology mapping. This paper presents a self-demonstration-driven approach that combines a neuro-symbolic task decomposition with a novel mechanism for automatically generating pattern-guided, dependency-aware demonstrations to address this integration challenge. Our approach incorporates two key strategies to achieve substantial accuracy gains over existing LLM-based schema integration methods: (i) a neuro-symbolic decomposition of the task into cascaded sub-tasks, where symbolic constraints structure the search space and LLMs perform semantic reasoning within each focused sub-task, and (ii) self-generated demonstrations guided by domain-agnostic patterns to supervise each sub-task. Experiments on three of the most challenging scenarios from the RODI benchmark show that our approach achieves state-of-the-art performance, substantially outperforming (25 percentage points F1 improvements) both traditional schema-to-ontology mapping techniques and recent LLM-based schema-to-ontology and schema matching approaches. Ablation studies further reveal the significant benefits of pattern-guided self-demonstrations and the complementary benefits of neuro-symbolic task decomposition.

---

### Does Reasoning Improve Psychological Depth in Large Language Models? It Depends on Who's Judging

Authors: Ruichen Zheng, Yihe Wang, Fabrice Y Harel-Canada, Sara Khosravi, Zeynep Senahan Yildiz, Amit Sahai, Nanyun Peng
Published: 2026-09-12
Categories: cs.LG, cs.CL
arXiv: https://arxiv.org/abs/2609.13773v1

LLM-as-a-Judge evaluators are increasingly used to score open-ended generation, yet a judge's correlation with human ratings on its development set may not guarantee valid measurement when outputs are closely matched and human preferences are subjective. We study this failure mode through psychological depth in short stories. Seven human readers and an LLM-judge ensemble selected on the original scalar Psychological Depth Scale dataset ($ρ= 0.646$) evaluated 60 blinded, prompt-matched story pairs from GPT-5 vs.\ GPT-4o and DeepSeek-R1 vs.\ DeepSeek-V3. Human preferences showed no universal reasoning advantage: GPT-5 was modestly preferred over GPT-4o (60.0--62.9\%), whereas DeepSeek-R1 trailed V3 (42.9\%), and inter-reader agreement was near chance (Krippendorff's $α= 0.070$), with within-reader consistency and recurring weighting patterns suggesting structured heterogeneity rather than random responding. The judge, by contrast, favored reasoning outputs in 89.0\% of dimension-level comparisons and 59 of 60 pairs on aggregate PDS, uniformly across all five evaluator configurations, and its scores were associated with surface features such as sentence length and lexical diversity. These results suggest that development-set performance is insufficient evidence for deployment validity on a shifted distribution, and that point-estimate judges can obscure the heterogeneity in subjective human evaluation.

---