# Research Papers — 2026-07 · bundle page 48 of 222 > Full-text markdown of papers #2351–2400 of 11075 published in 2026-07, newest first. AI-readable mirrors of the full 11075-paper 2026-07 archive (50 papers per file). Canonical page: https://aipod.dev/papers/archive/2026-07/page/48 [← Previous bundle](https://aipod.dev/papers/archive/2026-07/page/47.md) [Next bundle →](https://aipod.dev/papers/archive/2026-07/page/49.md) ### TriShieldRAG: A Three-Ring Defense-in-Depth Framework Against Knowledge Corruption in Retrieval-Augmented Generation Authors: Susil Kumar Mohanty, Rohit Patel, Kosuru Yuvaraj, Jeenal Chaudhary, Disha Singhania Published: 2026-07-26 Categories: cs.CR, cs.AI, cs.CL, cs.LG arXiv: https://arxiv.org/abs/2607.23838v1 Retrieval-Augmented Generation (RAG) lets a large language model answer questions using documents retrieved from an external knowledge base at query time. This makes RAG useful for private data, fast-changing information, and reducing hallucination, but it also means the model's answer is only as trustworthy as whatever the retriever hands it. If the knowledge base accepts writes from more than one party, an attacker needs only a handful of adversarial documents to steer the model toward a chosen wrong answer. PoisonedRAG demonstrated this: as few as five crafted documents flip an undefended system's answer roughly 90% of the time, and three natural single-stage defenses (perplexity filtering, query paraphrasing, knowledge-base expansion) leave attack success at 30% or higher. We built TriShieldRAG to close that gap. Rather than relying on one checkpoint, we place three independent, formally specified rings across the pipeline: an Ingest Guard that screens documents for lexical and statistical poisoning signatures; a Retrieval Scorer that re-ranks the retrieved set by a provenance and consistency-weighted trust score; and a Cross-LLM Consensus stage that polls three architecturally diverse language models (Claude, Mistral Small, Llama 3.2) and allows one bounded re-retrieval on disagreement. We derive the conditions under which Rings 2 and 3 are expected to work: a minority-poison assumption and an explicit provenance-tag assumption. Our reported configuration is consistent with this analysis, though we have not yet run the controlled poison-fraction sweep needed to confirm it independently. Evaluated against the non-adaptive attacker from the original PoisonedRAG, over a 5,000-document Wikipedia knowledge base with 10 target questions, the full pipeline reduces attack success rate from roughly 91% to roughly 13% while preserving accuracy on benign queries. --- ### Latent-LoRA: Compact Latent-Space Adapters with Gradient-Free Routing for Continual Learning Authors: Reza Rahimi Azghan, Gautham Krishna Gudur, Giulia Pedrielli, Pavan Turaga, Hassan Ghasemzadeh Published: 2026-07-26 Categories: cs.LG, cs.CL arXiv: https://arxiv.org/abs/2607.23837v1 Large language models generalize well to individual tasks but lack an inherent mechanism for learning them sequentially, leading to catastrophic forgetting. To mitigate this, LoRA-based continual learning methods allocate a separate low-rank adapter per task, yet existing approaches either require task identity at inference or sum all adapters indiscriminately, letting irrelevant branches distort the output. Recent gating-based solutions route inputs to the correct adapter but introduce trainable parameters that themselves need protection against forgetting. In this work, we observe that pooled token embeddings from a frozen LLM embedding layer already separate task distributions throughout the learning sequence. A Gaussian mixture model fitted on these embeddings, without any gradient-based training, is sufficient for task-agnostic adapter selection at test time. This eliminates the need for a learned gating module. On the adapter side, constraining each task's parameters to the principal subspace of the pretrained weights via SVD yields a compact latent-space parameterization. Within this subspace, orthogonal regularization directly controls inter-task interference. The resulting system, Latent-LoRA, is replay-free, requires no trainable routing component, and uses substantially fewer parameters per task. Experiments across five model scales and two established continual learning benchmarks show state-of-the-art performance with near-zero forgetting. --- ### Consistent Evidence, Robust Recognition: Faithful Attribution Regularization under Geometric Transformations Authors: Xianghao Jiao, Ruoyu Chen, Wei Wang, Jiazi Hu, Jiawei Liang, Shangquan Sun, Shiming Liu, Qunli Zhang, Xiaochun Cao Published: 2026-07-26 Categories: cs.CV arXiv: https://arxiv.org/abs/2607.23835v1 Attribution methods are widely used to characterize the evidence underlying model predictions, yet their potential to improve model behavior remains underexplored. Attribution inconsistency under label-preserving geometric transformations may indicate transformation-sensitive evidence reliance, motivating attribution regularization. However, such supervision is valid only when attribution faithfully reflects the evidence driving predictions. Existing self-supervised methods typically align gradient-based maps such as Grad-CAM, whose limited faithfulness means that attribution consistency need not imply consistency of the underlying decision process, leaving transformation robustness unresolved. We propose an annotation-free attribution regularization framework based on submodular search over image regions. By measuring how candidate subsets affect model outputs, the search extracts compact, class-discriminative evidence as search-derived supervision. We further introduce a submodular ranking loss with path-consistency and termination-alignment terms that respectively align spatially corresponding candidate rankings along paired search trajectories and encourage the transformed trajectory to satisfy the stopping criterion at the target terminal step. The loss provides a differentiable surrogate for regularizing both final attributions and the otherwise discrete evidence-selection process. Experiments on ImageNet-100 show that our method substantially improves attribution stability, Insertion, and Deletion on ViT-B/16 with only a 0.28-point accuracy drop, with similar gains on ViT-L/16. On ImageNet-1K, it improves transformed-input accuracy on ResNet-50 and ConvNeXt-B while limiting the clean-accuracy drop to 0.30 points, demonstrating more consistent evidence reliance with minimal performance loss. Code will be released soon. --- ### DriveDNA: A Large-Scale Multimodal Naturalistic Driving Dataset and Benchmark for Driving Style Identification Authors: Yuhang Wang, Lingyao Li, Hao Zhou Published: 2026-07-26 Categories: cs.LG arXiv: https://arxiv.org/abs/2607.23822v1 Driving style captures stable, driver-specific patterns in how a vehicle is driven. In naturalistic data, however, this signal is hard to isolate because drivers are observed in different vehicles, on different roads, and under different conditions, so models may mistake vehicle- or situation-specific regularities for driver-specific style. We introduce DriveDNA, a large-scale naturalistic dataset and benchmark for personalized driving-style modeling, comprising 4,121 drives from 465 drivers across 115 vehicle models and totaling 975 hours of human-controlled driving at 10 Hz with forward video, collected from community drivers in everyday use. DriveDNA defines driving style as a consistent, driver-specific behavioral pattern in how a vehicle moves under similar conditions. The benchmark evaluates this signal through three core tasks: few-shot driver re-identification, personalized behavior prediction, and condition-matched comparison, and provides behavioral annotations plus 276,248 rule-generated maneuver events across six classes with large-scale human auditing. We evaluate baselines spanning classical descriptors, supervised and self-supervised time-series encoders, multimodal fusion, probabilistic prediction, and zero-shot foundation models under a fixed multi-seed protocol. Learned representations substantially outperform classical descriptors on unseen drivers (AUROC .935 vs. .707) and retain driver-specific information under matched driving conditions, while descriptor performance approaches chance. Video-only models achieve comparable re-identification accuracy but exhibit severe route leakage, showing that strong recognition may arise from contextual shortcuts rather than driving behavior. These findings show that reliable driving-style evaluation must assess both the behavioral value of learned representations and their robustness to vehicle, drive, and condition confounds. --- ### SCTA: An Agentic Framework for Stable and Interpretable Target Gene Discovery from Single-Cell RNA Sequencing Authors: Shuyu Chen, Chen Zhu, Ye Zhang, Yang Li, Qiqi Xie, Haohan Wang Published: 2026-07-26 Categories: cs.LG, q-bio.GN arXiv: https://arxiv.org/abs/2607.23821v1 Identifying therapeutic target genes from single-cell RNA sequencing (scRNA-seq) data remains a fundamental challenge in translational biology. Unlike bulk assays, scRNA-seq captures heterogeneous cellular states and rare subpopulations, but this same heterogeneity makes target discovery highly sensitive to analytical choices throughout the pipeline, including preprocessing, cell population selection, differential expression analysis, and downstream biological interpretation. As a result, existing workflows and general-purpose analysis agents often produce unstable or difficult-to-interpret target hypotheses, limiting their reliability for disease-focused discovery. We present SCTA (Single-Cell Target Agent), a decision-centric agentic framework for stable and interpretable target gene discovery from scRNA-seq data. Rather than treating analysis as a single general-purpose reasoning task, SCTA decomposes target discovery into specialized agents aligned with key decision points in the single-cell pipeline and constrains downstream reasoning with structured biological evidence. In a representative ablation study on hereditary chronic pancreatitis, we demonstrate that SCTA's full evidence integration yields the most stable target selection across independent runs among the tested configurations, while recovering biologically coherent, disease-relevant mechanisms validated in prior studies. These results suggest that decision-aware agent orchestration tailored to the structure of single-cell analysis can improve the robustness, interpretability, and practical utility of target discovery in precision medicine. --- ### Kalypso: Relational LLM Serving Authors: Hojae Son, Md Ashraful Islam, Huy Gia Cao, Hui Guan, Marco Serafini Published: 2026-07-26 Categories: cs.DB, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2607.23815v2 Large language models are increasingly used as semantic operators for filtering, extracting, ranking, joining, and transforming unstructured data. Existing semantic query processing systems invoke request-centric LLM serving systems that are unaware of the query plan, leaving substantial performance opportunities unused. This paper introduces relational LLM serving, an abstraction that makes LLM serving aware of semantic query structure while preserving query semantics and output accuracy. The key opportunity is pipelined execution across semantic operators: when intermediate tuples flow directly from one operator to the next, their KV-cache state can be reused instead of recomputed. We present Kalypso, a relational LLM serving system that exposes an API for semantic query plans and executes them using an adaptive, memory-aware scheduling algorithm. Kalypso addresses a new online scheduling problem in which pipelined operator execution is coupled with GPU memory pressure management to reuse KV-cache state in the serving engine before eviction. Its scheduler continuously adjusts memory allocations to balance upstream parallelism, downstream progress, and GPU utilization. Our evaluation shows that Kalypso improves query completion time over baselines using request-centric LLM serving, with speedups up to 4.57x across diverse workloads, demonstrating that query-aware LLM serving can substantially improve the efficiency of semantic query execution. --- ### Kalypso: Relational LLM Serving Authors: Hojae Son, Md Ashraful Islam, Huy Gia Cao, Hui Guan, Marco Serafini Published: 2026-07-26 Categories: cs.DB, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2607.23815v1 Large language models are increasingly used as semantic operators for filtering, extracting, ranking, joining, and transforming unstructured data. Existing semantic query processing systems invoke request-centric LLM serving systems that are unaware of the query plan, leaving substantial performance opportunities unused. This paper introduces relational LLM serving, an abstraction that makes LLM serving aware of semantic query structure while preserving query semantics and output accuracy. The key opportunity is pipelined execution across semantic operators: when intermediate tuples flow directly from one operator to the next, their KV-cache state can be reused instead of recomputed. We present Kalypso, a relational LLM serving system that exposes an API for semantic query plans and executes them using an adaptive, memory-aware scheduling algorithm. Kalypso addresses a new online scheduling problem in which pipelined operator execution is coupled with GPU memory pressure management to reuse KV-cache state in the serving engine before eviction. Its scheduler continuously adjusts memory allocations to balance upstream parallelism, downstream progress, and GPU utilization. Our evaluation shows that Kalypso improves query completion time over baselines using request-centric LLM serving, with speedups up to 4.57x across diverse workloads, demonstrating that query-aware LLM serving can substantially improve the efficiency of semantic query execution. --- ### Earnings25: A Comprehensive 500-Hour Speech Benchmark for Finance Authors: Denglin Jiang, Haoran Zhou, Anshul Wadhawan, Brendan Fahy, Vinay Ramesh, David Weisberg, Dmitriy Derkachevskiy, Helen Sheehan, Srivas Prasad, Michele Franceschini Published: 2026-07-26 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2607.23813v1 We introduce Earnings25, a finance-domain benchmark for evaluating automatic speech recognition (ASR) on English-language earnings calls under realistic conditions. Earnings25 comprises two complementary test sets: (i) testset-full, 498 hours of full English-language S&P 500 earnings calls from Q4 2025, and (ii) testset-segmented, a 46-hour industry-balanced set of 290 segments sampled from English-language U.S. earnings calls in 2025. The benchmark provides aligned transcripts and structured metadata, including speaker roles, industry labels, and call structure, enabling speaker- and industry-aware evaluation beyond aggregate word error rate (WER). We report reproducible baselines for Whisper and Parakeet-TDT using standardized scoring. --- ### Memory Efficient Audio Synthesis with Decoupled Temporal Depth Diffusion Transformers Authors: Dongseong Hwang, Prasanth Yadla, Kaan Elgin, Shifas Padinjaru Veettil, Sivanand Achanta, Dipjyoti Paul, Ramya Rasipuram, Tyler Johnson, Emad Soroush, Chung-Cheng Chiu, Zhifeng Chen Published: 2026-07-26 Categories: cs.SD, cs.CL arXiv: https://arxiv.org/abs/2607.23811v2 Siri Expressive Voices synthesize rich, configurable speech in real time and entirely on device, powered by AFM 3 Core Advanced, Apple's most powerful on-device foundation model. This work presents the memory-efficient audio synthesis architecture behind that capability: a detokenizer that converts the semantic audio tokens emitted by the foundation model into high-fidelity audio within the tight compute and memory budget of the Apple Matrix Coprocessor (AMX). We convert semantic audio tokens to a residual vector quantization (RVQ) representation with a three-component design, a streaming encoder, a temporal decoder, and a depth decoder, that systematically decouples temporal and depth processing. A single reusable depth decoder with Diffusion Transformer (DiT)-style stage conditioning generates all RVQ levels autoregressively, replacing the dedicated per-level decoders of prior multi-decoder architectures, while causal sliding window attention with fixed-window key-value caching yields constant memory complexity independent of sequence length. Deployed on the AMX, the detokenizer sustains roughly 10 ms per generation step, about 16x faster than real time, with a peak runtime memory of only 21 MB and 329 MB of on-device assets, enabling continuous streaming synthesis of 20-320 seconds of audio. This constant, small footprint replaces the linear and quadratic memory scaling of conventional transformer- and GAN-based approaches. Ablation studies validate the key architectural components, and audio quality assessment confirms that the architecture maintains synthesis fidelity while achieving efficiency gains over existing methods. Operating at a 1-billion-parameter activation size within AFM 3 Core Advanced, it improves Mean Opinion Score by +0.28 overall (4.15 vs. 3.87) and by +0.42 on conversational speech (4.24 vs. 3.82) over the prior on-device text-to-speech system. --- ### Memory Efficient Audio Synthesis with Decoupled Temporal Depth Diffusion Transformers Authors: Dongseong Hwang, Prasanth Yadla, Kaan Elgin, Shifas Padinjaru Veettil, Sivanand Achanta, Dipjyoti Paul, Ramya Rasipuram, Tyler Johnson, Emad Soroush, Chung-Cheng Chiu, Zhifeng Chen Published: 2026-07-26 Categories: cs.SD, cs.CL arXiv: https://arxiv.org/abs/2607.23811v1 Siri Expressive Voices synthesize rich, configurable speech in real time and entirely on device, powered by AFM 3 Core Advanced, Apple's most powerful on-device foundation model. This work presents the memory-efficient audio synthesis architecture behind that capability: a detokenizer that converts the semantic audio tokens emitted by the foundation model into high-fidelity audio within the tight compute and memory budget of the Apple Matrix Coprocessor (AMX). We convert semantic audio tokens to a residual vector quantization (RVQ) representation with a three-component design, a streaming encoder, a temporal decoder, and a depth decoder, that systematically decouples temporal and depth processing. A single reusable depth decoder with Diffusion Transformer (DiT)-style stage conditioning generates all RVQ levels autoregressively, replacing the dedicated per-level decoders of prior multi-decoder architectures, while causal sliding window attention with fixed-window key-value caching yields constant memory complexity independent of sequence length. Deployed on the AMX, the detokenizer sustains roughly 10 ms per generation step, about 16x faster than real time, with a peak runtime memory of only 21 MB and 329 MB of on-device assets, enabling continuous streaming synthesis of 20-320 seconds of audio. This constant, small footprint replaces the linear and quadratic memory scaling of conventional transformer- and GAN-based approaches. Ablation studies validate the key architectural components, and audio quality assessment confirms that the architecture maintains synthesis fidelity while achieving efficiency gains over existing methods. Operating at a 1-billion-parameter activation size within AFM 3 Core Advanced, it improves Mean Opinion Score by +0.28 overall (4.15 vs. 3.87) and by +0.42 on conversational speech (4.24 vs. 3.82) over the prior on-device text-to-speech system. --- ### ACM: Agentic Context Management for Long Horizon Tasks Authors: Xiaochuan Li, Ryan Ming, Meng Chu, Shuai Shao, Rong Jin, Chenyan Xiong Published: 2026-07-26 Categories: cs.AI arXiv: https://arxiv.org/abs/2607.23809v1 Agentic tasks are inherently long-horizon and multi-turn, constantly accumulating context through interactions with the environment. Existing context compression methods inevitably incur information loss and are triggered by rigid heuristic rules, leaving them misaligned with the agent's evolving reasoning focus. We propose Agentic Context Management (ACM), a framework that equips agents with purpose-built context editing tools for lossless context management. Inspired by the interaction between short-term and long-term human memory, the agent autonomously decides when to compress its context, offloads discarded content to an external memory system, and queries it on demand for later retrieval. Building on this framework, we further develop a post-training pipeline that constructs high-quality demonstrations of context management and improves model performance on both agentic search and coding tasks. Further analysis reveals that effective context management reduces peak token pressure, enables extended explorations, and yields more consistent solutions across independent trials. Code, data, and model checkpoints are available at https://github.com/lixiaochuan2020/agentic-context-management. --- ### Indic DiarBench: A Multilingual Joint Diarization and ASR Benchmark for Indian Languages Authors: Deovrat Mehendale, Aditya Mehndiratta, Dhruv Rathi, Kaushal Bhogale, Mitesh M. Khapra Published: 2026-07-26 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2607.23808v1 In this work, we introduce Indic DiarBench, a speaker diarization and ASR benchmark dataset spanning all 22 scheduled languages of India. This corpus comprises approximately 108 hours of natural multi-speaker audio from near-field meetings, far-field recordings, and in-the-wild audios. All annotations are human-corrected with time-aligned speaker attributed transcriptions. The dataset captures conversational nuance prevalent in Indian speech, such as English code-mixing, dialectal variation, and frequent speaker overlap. To establish a baseline for joint ASR and diarization capabilities we evaluate leading systems including commercial speech APIs and multimodal large language models. Indic DiarBench is released as an open-access resource to advance inclusive, multilingual speech technology research for Indian languages. --- ### A Frozen 12B Beats Frontier Models on Verified Work: 100% Accuracy, 0 Tokens, Bit-Exact, Forever Authors: Sietse Schelpe Published: 2026-07-26 Categories: cs.CL, cs.AI, cs.IR, cs.LG, cs.PF arXiv: https://arxiv.org/abs/2607.23806v1 Improving a language model today means retraining it: enormous compute, a new opaque model each cycle, non-deterministic output. We take the opposite path: the model stays frozen, and a persistent memory of verified solutions grows beside it. Once a problem family is solved and has passed an independent verification step that never consults the answer key, every new instance of that family is answered at zero generation tokens, bit-exact, deterministically. Across 180 fresh instances spanning nine problem families, four architectures from four vendors - dense and mixture-of-experts - each score 180/180 at zero generation tokens per answer: execution-bound capability decoupled from parameter scaling. A negative control attributes the capability fully to the memory: emptied, it solves nothing. The same verify-before-store contract holds for open-ended reasoning: 88/88 consistency-gated acceptances across all four models, machine-checked formal proof, and reasoning-method transfer at 77/80. Memory selection takes 1.4 microseconds; a full reuse completes in 6-23 ms at 36 mWh. Approximate similarity retrieval selects the wrong item 94.3% of the time on a 4,500-item verified store where exact addressing makes zero errors. The store also serves as working context at a scale no shipped engine matches: a 6,000,000-token movable window on a single 46 GB GPU at flat memory, where vLLM stops at 30,399 tokens and SGLang silently truncates past 32,000. On published benchmarks, frontier models remain far ahead of any 12B at raw from-scratch reasoning; on everything this system has solved and verified, the comparison inverts: a frontier API call pays a fresh generation pass on every query, forever, while verified reuse costs zero tokens and returns the identical bits every time. A public testbench with free, rate-limited access accompanies this report: https://corbenic-galahad-bench.hf.space --- ### How Context Attribution Handles What the Model Already Knows Authors: Quoc-Huy Trinh, Lin Zhu, Sebastian Szyller Published: 2026-07-26 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2607.23804v2 Context attribution methods for large language models (LLMs) identify which input context contributes to the model response. Recent works show the initial success in attributing the con- tributive score of the contexts. However, we observe that when the context overlaps with the training data, these methods can- not disentangle in-context from in-weight (IW) contributions, producing unreliable scores. Based on this observation, in this work, we introduce: 1) an evaluation protocol that relies on four new metrics (base-model context attribution score (BCS), cross-model context attribution consistency (CAC), attribution preservation score (APS), source separation pre- cision (SSP)) and 2) a benchmark dataset (WMDP-Cyber++) with ground-truth provenance labels to systematically assess attribution under IW overlap. In our experiments across four well-known context attribution methods, we demonstrate that they provide unfaithful attribution when the knowledge from the context also exists in the weights. Finally, we adapt these methods for source separation (IW vs. in-context learning (ICL)) and show that they cannot do the disentanglement based on the contributive score --- ### How Context Attribution Handles What the Model Already Knows Authors: Quoc-Huy Trinh, Lin Zhu, Sebastian Szyller Published: 2026-07-26 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2607.23804v1 Context attribution methods for large language models (LLMs) identify which input context contributes to the model response. Recent works show the initial success in attributing the con- tributive score of the contexts. However, we observe that when the context overlaps with the training data, these methods can- not disentangle in-context from in-weight (IW) contributions, producing unreliable scores. Based on this observation, in this work, we introduce: 1) an evaluation protocol that relies on four new metrics (base-model context attribution score (BCS), cross-model context attribution consistency (CAC), attribution preservation score (APS), source separation pre- cision (SSP)) and 2) a benchmark dataset (WMDP-Cyber++) with ground-truth provenance labels to systematically assess attribution under IW overlap. In our experiments across four well-known context attribution methods, we demonstrate that they provide unfaithful attribution when the knowledge from the context also exists in the weights. Finally, we adapt these methods for source separation (IW vs. in-context learning (ICL)) and show that they cannot do the disentanglement based on the contributive score --- ### Beyond Appearance: A Multi-cue Framework and Large-scale Benchmark for Pedestrian Association and Tracking on Mobile Aerial-Ground Platforms Authors: Ruiqi Wu, Bingliang Jiao, Ruize Han, Hangzheng Yu, Xunkai Jiang, Shining Wang, Yuanqi Hu, Wenxuan Wang, Peng Wang Published: 2026-07-26 Categories: cs.CV, cs.MM arXiv: https://arxiv.org/abs/2607.23803v1 Multi-view Multi-object Association and Tracking (MvMoAT) associates objects across camera views and tracks them over time, supporting identity persistence and forensic trajectory reconstruction in multi-platform cooperative perception. Unlike conventional multiple object tracking, MvMoAT faces frequent viewpoint shifts that distort appearance and undermine cross-view association and temporal tracking. We propose FUSION, a viewpoint-robust Feature Unification framework for multi-view aSsociation and IdentificatiON. Its Multi-cue Adaptive Combination (MAC) module adaptively integrates viewpoint-invariant cues with appearance features to improve cross-view association, while Online Multi-view Feature Synchronization (OMFS) aggregates pedestrian features across historical and cross-view frames for temporally consistent tracking. We also introduce RealMvMoAT, a large-scale benchmark featuring substantial inter- and intra-camera viewpoint variation. It contains 504.9K frames from 7 cameras (5 UAV and 2 ground views) across 10 scenes, with over 7.3M identity-labeled bounding boxes. All cameras exhibit random and substantial motion. To the best of our knowledge, RealMvMoAT is the largest MvMoAT dataset to date. Its scale, viewpoint diversity, complex platform motion, and realistic trajectories provide a comprehensive resource for future research. Experiments on RealMvMoAT and six public benchmarks show that FUSION achieves state-of-the-art performance. --- ### From RLVR to RLSVR: Task Transformation Induces Self-Verifiable Rewards for Open-Ended LLM Self-Improvement Authors: Qinsi Wang, Jing Shi, Huazheng Wang, Kun Wan, Yiran Wu, Bo Liu, Qingyun Wu, Hai Helen Li, Yiran Chen, Handong Zhao, Wentian Zhao Published: 2026-07-26 Categories: cs.AI arXiv: https://arxiv.org/abs/2607.23802v2 Reinforcement Learning with Verifiable Rewards (RLVR) has driven recent progress in reasoning-oriented large language models (LLMs) by enabling large-scale optimization. However, its applicability remains largely limited to domains such as mathematics and coding, where correctness can be deterministically verifiable. Open-ended tasks instead often rely on human preferences, reward models, or LLM-based judges, introducing evaluation bias, judge capability bottlenecks, and additional inference costs. Drawing on the principle of self-supervised learning, which constructs pretext tasks to derive supervision from the data itself, we propose Reinforcement Learning with Self-Verifiable Rewards (RLSVR), a task-transformation-based training paradigm for extending RLVR to open-ended tasks. RLSVR transforms open-ended tasks into verifiable proxy environments whose internal rules and interaction outcomes automatically generate reward signals. We instantiate RLSVR with SpyRL, a Self-PlaY Reinforcement Learning method inspired by social deduction game Who Is the Spy?. Agents receive asymmetric information, complete the same target task, and vote to identify a designated spy. Because the spy identity is predetermined, voting outcomes provide fully verifiable rewards, while successful identification remains closely related to output quality. Experiments on text summarization, creative writing, and mathematical reasoning show that SpyRL outperforms existing self-improvement methods on non-verifiable tasks and yields consistent gains on verifiable reasoning tasks. These results demonstrate that task transformation can extend scalable RLVR-based self-improvement beyond inherently verifiable domains. Models and code have been released at https://github.com/wangqinsi1/RLSVR/tree/SpyRL. --- ### Provable Speedups From Dynamic Population Sizes in Evolutionary Algorithms for Multiobjective Optimization Authors: Andre Opris Published: 2026-07-26 Categories: cs.NE arXiv: https://arxiv.org/abs/2607.23800v1 This paper investigates the role of dynamic population sizes in evolutionary multi-objective optimization. Although such approaches are widely used in practice, their benefits remain poorly understood, and rigorous runtime analyses explaining when and why they help are still scarce. To address this, we introduce the bi-objective problem class CLIMB and analyze the runtime of GSEMO and the widely used NSGA-II on this problem. Our results show that allowing a dynamic population size for NSGA-II can lead to a moderate improvement, yielding a speedup of order $Ω(\sqrt{n}/\log n)$. In particular, we prove that GSEMO and NSGA-II-DYN, a version of NSGA-II with dynamic population sizes we propose in this paper, can find the Pareto front of CLIMB in expected $O(n \log n)$ fitness evaluations, whereas NSGA-II with a fixed population size requires $Ω(n^{1.5})$ fitness evaluations in expectation. To the best of our knowledge, this is the first rigorous runtime analysis in multi-objective optimization demonstrating a super-constant speedup of GSEMO over NSGA-II. Our analysis builds on concepts from single-objective optimization, like the evolution of population diversity over time, and employs the well-known family-three method to prove the lower bound. --- ### PathScale-R1: Cross-scale Reasoning for Pathological Image Analysis Authors: Chi Phan, Tianyi Zhang, Yufeng Wu, Qiaochu Xue, Jiajie Zhang, Linghan Cai, Zeyu Liu, Sudong Wang, Yueming Jin, Dan Hu Published: 2026-07-26 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2607.23794v1 Pathological diagnosis is inherently multi-scale, requiring the integration of global tissue architecture at low magnification with cellular morphology at higher magnification. However, existing pathology benchmarks and vision-language models (VLMs) are still largely developed under single-scale settings, limiting their ability to learn clinically meaningful multi-magnification reasoning. Moreover, naively constructed visual question answering (VQA) tasks may be susceptible to text-only or superficial visual shortcuts, leading to unreliable assessments of visual understanding. To address these limitations, we introduce a benchmark and training framework for shortcut-resistant cross-scale pathology reasoning. We design an Adversarial Text-only Screening strategy for semantic reasoning questions and a Structure-controlled Distractor Sampling strategy for visual grounding questions, encouraging models to rely on cross-scale visual evidence. Based on this pipeline, we construct PathScale-VQA, a high-quality cross-scale pathology VQA benchmark with 10,373 multiple-choice questions grounded in 1,368 diagnostic paths across multiple magnification levels. Building on the semantic reasoning set, PathScale-R1 is optimized through Difficulty-driven Reasoning Distillation supervised fine-tuning followed by reinforcement learning with a Scale-aware Reasoning Structure reward, which encourages the use of evidence across magnifications. Extensive experiments demonstrate state-of-the-art performance of PathScale-R1 on cross-scale reasoning tasks and effective transfer to conventional single-scale pathology VQA. Our code is available at https://github.com/iMVR-PL/PathScale-R1. --- ### GNN-based Multi-Agent Control of Traffic Shockwaves in Sparse Vehicular Ad-hoc Networks Authors: Prachi Nandi, Madhuri Malakar, Sonakshi Satpathy, Pabitra Mohan Khilar Published: 2026-07-26 Categories: cs.NI, cs.LG, math.NA arXiv: https://arxiv.org/abs/2607.23792v2 Traffic shockwaves are stop-and-go waves that propagate upstream through the streams of vehicles and are one of the major causes of traffic congestion, fuel inefficiency, and increased accident rates in modern transportation systems. Although Connected and Autonomous Vehicles (CAVs) offer a promising opportunity to mitigate such shockwaves, most existing control strategies rely on global traffic state information, making them impractical for early-stage deployment of Vehicular Ad-hoc Networks (VANETs). In this paper, we propose a decentralized Multi-Agent Reinforcement Learning (MARL) framework that integrates a Graph Neural Network (GNN) to enhance the control architecture of connected and autonomous vehicles. The proposed approach enables vehicles to learn cooperative control policies using locally available information and interaction with neighboring vehicles. The effectiveness of the proposed scheme is evaluated using a scalable simulation environment under realistic highway traffic conditions. Simulation results show that the proposed GNN-based MARL framework can reduce the propagation of traffic shockwaves by up to 80%, even when only 10% of the vehicles are connected. --- ### Maximum Satisfiability of Simple Temporal Problems Authors: Johannes K. Fichte, Johanna Groven, Peter Jonsson, Victor Lagerkvist, Jorke M. de Vlas Published: 2026-07-26 Categories: cs.CC, cs.AI arXiv: https://arxiv.org/abs/2607.23785v1 The Simple Temporal Problem (STP) is a core framework for quantitative temporal constraints. As STP data can be inconsistent, we study MAXSTP: compute a maximum-cardinality consistent subset of constraints. This extension is NP-hard, and we analyze its parameterized complexity under measures that capture practically relevant instance features: the number of variables $n$ (instance scale), the maximum coefficient magnitude $k$ (numeric range), and structural parameters of the constraint graph such as treewidth $tw$ (decomposability) and vertex cover size $vc$ (density). We show that MAXSTP is W(1)-hard parameterized by $n$, implying that $n$ and parameters that depend on $n$ (including $tw$ and $vc$) are insufficient for fixed-parameter tractability. For combined parameters, we give an $O^*(k^n)$-time algorithm, yielding single-exponential solvability for fixed $k$. While $k+tw$ remains W(1)-hard, MAXSTP is in XP via an $O^*((n\cdot k)^{tw})$ algorithm. Our results suggest that MAXSTP is often computationally harder than optimizing qualitative CSPs. We verify that many such problems (including RCC-8 and Allen's algebra) are FPT when parameterized by $n$ or $tw$. However, we also demonstrate that FPT algorithms for MAXSTP are indeed possible but with other parameters such as $k + vc$. --- ### A Few Words Go a Long Way: Language Guided Robot Policy Synthesis Authors: Daphne Chen, Archit Ritesh Jain, Eric Goossen, Emma Romig, Michael Murray, Nick Walker, Maya Cakmak Published: 2026-07-26 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2607.23784v1 While vision-language-action models have demonstrated impressive zero-shot manipulation capabilities, they remain fundamentally black box policies that are difficult to interpret, adapt, or correct when they inevitably fail. In this work, we propose ARCHITECT, a framework that treats robot policy acquisition as an interactive program synthesis task. ARCHITECT leverages the reasoning capabilities of LLM coding agents to synthesize modular robot programs that utilize a suite of perception and control tools. Unlike end-to-end models where distribution shift leads to unpredictable, cascading failures, our modular architecture allows users to isolate failures and localize feedback at the level of abstraction required. We introduce an iterative process where a human supervisor provides natural language corrections to steer the policy. These corrections are grounded in the policy code by program execution traces and distilled into a persistent skill library, a form of long-term in-context learning which enables the agent to accumulate a repertoire of reusable, interpretable behaviors. In a benchmark evaluation on a Franka Panda robot, ARCHITECT outperforms state-of-the-art VLA models and program synthesis baselines on complex, long-horizon tasks, including articulated object manipulation and cloth folding. Our results demonstrate that the synthesized skill library enables the system to transfer to novel tasks with decreasing human intervention, providing a steerable and data-efficient alternative to black-box robot learning. Website: https://robo-architect.github.io/ --- ### Scale Weight Decay and Train Better Authors: Anuj Apte Published: 2026-07-26 Categories: cs.LG, cond-mat.dis-nn, cs.AI, math.OC arXiv: https://arxiv.org/abs/2607.23777v1 The discovery of scaling laws has motivated training neural networks on ever increasing quantities of data. This is typically done with a constant decoupled weight decay which causes the network weights to shrink steadily over the course of training. Taking inspiration from the Robbins--Monro conditions, we propose to scale weight decay by the fraction of the peak learning rate $η/η_{\max}$. We prove that this scaled weight decay preserves the asymptotic stationarity guarantees of the corresponding unregularized methods for both stochastic gradient descent and the non-Euclidean spectral optimizer Muon, thereby avoiding the additional asymptotic bias introduced by constant decoupled weight decay. This retains the stability benefits of weight decay without changing the asymptotic optimization target. Using a steady-state analysis, we explain why under standard weight decay the weight norm shrinks steadily as training proceeds, whereas under scaled weight decay it settles to a roughly constant value. When applied to the training of mixture-of-experts models, Muon with scaled weight decay (Muon-SW) consistently outpaces Muon with identical hyperparameters, reaching the same validation loss $\mathbf{30\%}$ faster at our largest scale across models from $72 - 930$ million parameters trained at $\sim 600$ tokens per active parameter. If this trend continues to hold, the method promises to substantially accelerate the pre-training of frontier models while requiring only a few lines of code to implement. --- ### Outcome-Fair Restless Multi-Armed Bandits for Stochastic Deadline Scheduling Authors: Shakti Sharma, Rahul Meshram Published: 2026-07-26 Categories: eess.SY, cs.AI arXiv: https://arxiv.org/abs/2607.23772v1 We study a restless multi-armed bandit (RMAB) problem for a stochastic deadline scheduling application. RMAB problems are solved using the Whittle index policy. The goal in RMAB is to maximize the expected cumulative discounted reward maximization. The Whittle index policy maximizes reward, but is not fair among two classes. In this paper, we introduce fairness criteria and study an outcome-fair model for RMAB which allows fairness for jobs and users structurally disadvantaged demographic classes. We formulate an outcome fair stochastic deadline scheduling problem as RMAB, and we develop the outcome fair Whittle index policy. We define a virtual queue mechanism that dynamically enforces long-term completion rate guaranties across demographic groups. We analyze a standard Whittle index policy and the outcome-fair index policy. We demonstrate the performance of our algorithms with numerical examples. We compare policies---Whittle index policy (no fairness), input-fairness Whittle index policy, outcome fair Whittle index policy. We observe that the outcome-fair Whittle index policy provides better fairness among classes compared to other policies. We demonstrate a trade off between fairness and profit. This decreases as the server capacity increases. --- ### Training Language Models to Cooperate with Inference-Time Controllers Authors: Moumita Choudhury, Vanshaj Khattar, Jing Liu, Toshiaki Koike-Akino, Ankush Chakrabarty, Shlomo Zilberstein, Ye Wang Published: 2026-07-26 Categories: cs.AI arXiv: https://arxiv.org/abs/2607.23771v1 Large language model (LLM) performance increasingly depends not only on the base model, but also on the inference-time controller used to organize reasoning. Existing post-training methods, however, typically optimize for a single fixed interaction pattern, despite real deployments relying on diverse controllers such as Chain-of-Thought, self-consistency, debate, planning, and verification pipelines. This creates a training--deployment mismatch and limits transfer to new workflows. We introduce CALM (Controller-Aware Language Models), a post-training framework that explicitly places controllers in the training loop. We formulate controller-aware post-training as multi-task reinforcement learning over controller-induced interaction protocols, where controllers are compositions of reusable local reasoning modules. This structure also induces a module-level decomposition of mixed-controller training under a turn-level GRPO objective, enabling a systematic study of controller and module-aware training strategies. We evaluate CALM on held-out controller compositions and broader controller shifts, showing that controller-aware post-training improves generalization across inference-time workflows beyond single-controller optimization. --- ### A Comparison of Data Augmentation Methods for Training Deep Neural Networks on Synthetic Aperture Sonar Authors: C. J. Moore, Gregory D. Vetaw, Jordan Malof Published: 2026-07-26 Categories: cs.LG arXiv: https://arxiv.org/abs/2607.23770v3 In this work we study Automatic Target Recognition (ATR) for Synthetic Aperture Sonar (SAS) data with a focus on deep neural networks (DNNs). The main challenge in training DNNs for SAS-ATR arises from the limited quantity of labeled target examples due to the significant costs and time required to collect real-world SAS data. One successful general strategy for mitigating the problem of limited training data is augmentation, which generates additional synthetic training data by introducing realistic variations to available data. Prior research has investigated a variety of augmentation strategies for SAS-ATR, including conventional image augmentations (e.g., contrast changes, cropping) as well as augmentations motivated the specific physics of SAS data. Building on prior work, we systematically compare many of these existing augmentation strategies for training DNNs for SAS-ATR. We also investigate the impact of augmentation when combined with modern DNN architectures such as transformers. The results indicate that augmentation can improve target recognition accuracy, although benefits vary, and not all augmentations are beneficial. --- ### WISERouter: LLM Routing with Workload Budget Constraint Authors: Yifei Li, Zihui Gao, Laks V. S. Lakshmanan Published: 2026-07-26 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2607.23765v1 Large language models (LLMs) achieve impressive performance across multiple domains, but using the most capable model for every query is prohibitive at scale. LLM routing exploits diversity in model capability and cost by assigning each query to a suitable model to balance utility and budget. Current methods have two limitations: (i) they either use heuristics that do not always enforce the budget constraint or impose a fixed per-query budget that cannot adapt across the workload and leads to suboptimal performance; (ii) they require supervised learning on a dense dataset with statistics for every query-model pair, which is expensive to collect. To address these challenges, we formulate LLM routing as a constrained contextual multi-armed bandit problem and introduce WISERouter (WR for short), a framework that supports offline learning from historical interactions as well as online learning with exploration. We further prove that WR-Online achieves a sublinear regret bound of $O(\sqrt{T})$ over a time horizon $T$. Empirical results on RouterBench and SWE-Bench demonstrate that (i) WR-Offline surpasses existing baselines in performance under a fixed budget and adheres more closely to budget constraints, and (ii) WR-Online achieves comparable performance to the baselines, while using substantially less exploration data. --- ### Escaping the Euclidean Void: Manifold-Informed Flow Matching for Sequential Recommendation Authors: Dengzhao Fang, Jingtong Gao, Yu Li, Xiangyu Zhao, Yi Chang Published: 2026-07-26 Categories: cs.IR, cs.AI arXiv: https://arxiv.org/abs/2607.23762v1 Conventional recommenders capture users' preferences by optimizing observed user-item relations, whereas continuous generative recommendation additionally learns the trajectory of synthesizing a target item. Flow matching drives this process by gradually shaping initial noise into a definitive next-item representation through intermediate states in a continuous embedding space. However, item catalogs are discrete and sparsely supported, meaning even a straight Euclidean path can cross continuous regions that contain little evidence of valid item semantics. Formalizing this failure as the Euclidean void, we propose MIRAGE, a Manifold-Informed Rectification framework for Accelerated Generation of Embeddings in sequential recommendation, which rectifies the learned embedding geometry around an unchanged straight probability path. By leveraging an item co-occurrence graph as a proxy for the underlying semantic manifold, MIRAGE aligns interpolated path states with local anchors, reorganizing the embedding space to ground the trajectory in valid item support. MIRAGE retains the original probability path and uses the graph only during training, thereby enabling accurate and efficient one-step inference. Extensive experiments on four real-world datasets reveal that MIRAGE consistently outperforms state-of-the-art baselines, effectively boosting performance on sparsely observed targets while achieving robust overall accuracy. Our code will be made publicly available upon publication. --- ### RoadVGGT: Road-Structure-Aware Feed-Forward Road Surface Reconstruction Authors: Han Jiao, Chen Liu, Jiakai Sun, Zhanjie Zhang, Mengyuan Yang, Yimeng Li, Mofan Zhou, Kun Zhan, Lei Zhao Published: 2026-07-26 Categories: cs.CV arXiv: https://arxiv.org/abs/2607.23758v1 Large-scale road surface reconstruction supports high-definition mapping, autonomous-driving perception, annotation, and simulation. Existing road-specialized optimization methods can produce high-quality road representations, but they typically require per-scene training and scene-dependent coverage design around the driving trajectory, limiting scalable reconstruction over newly collected roads. To address these limitations, we introduce RoadVGGT, a road-structure-aware feed-forward framework that reconstructs compact Gaussian road surfaces without test-time per-scene optimization. RoadVGGT uses a geometric foundation model to exploit multi-view images together with provided pose and depth observations, and predicts dense pixel-aligned Gaussian attributes through a learned Gaussian head. To make these dense predictions usable for large road surfaces, we align them into a consistent metric world coordinate system and fuse redundant Gaussians on the road-aligned XY plane through confidence-weighted grid fusion. Category-aware grouping and road--sidewalk junction protection further control fusion around vulnerable road structures. The resulting representation supports RGB and semantic bird's-eye-view maps, elevation estimation, and novel view synthesis. RoadVGGT eliminates the need for per-scene optimization in prior methods, reconstructs complete road surfaces with a compact Gaussian representation, and improves image quality, semantic mapping, and elevation accuracy. Extensive experiments demonstrate the potential of geometric foundation models for scalable feed-forward road surface reconstruction. --- ### DAP-Pose: Deep Temporal Alignment and Physics-aware Cross-modal Sensor Fusion for Robust Pose Estimation Authors: Jianhan Lin, Yuchu Qin, Jiateng Yuan, Wenbo Zhang, Shuai Gao Published: 2026-07-26 Categories: cs.CV, cs.RO arXiv: https://arxiv.org/abs/2607.23755v1 Robust and accurate pose estimation with multi-modal sensors is fundamental for autonomous vehicles and mobile robotic systems in complex environments. In this paper, we propose DAP-Pose, a unified end-to-end model for robust multi-modal pose estimation. DAP-Pose introduces a Bi-level Cross-modal Fusion (BCF) module that captures complementary semantic and geometric motion cues from visual, inertial, and GNSS measurements. To handle temporal offsets, we designed a Deep Temporal Alignment (DTA) module that explicitly aligns asynchronous streams in latent space, enabling coherent motion modeling without strict hardware synchronization. Furthermore, we incorporate physics-aware constraints via manifold geometry and GNSS-guided absolute metric scale, enforcing motion consistency and mitigating drift. Experiments upon the public KITTI benchmark dataset were conducted to evaluate the performance of DAP-Pose against existing methods. DAP-Pose achieved the state-of-the-art performance, with the lowest average translation error ($t_{rel}$) of 1.31% and rotation error ($r_{rel}$) of 0.46$^{\circ}$. Furthermore, it accurately estimates poses and maintains robust performance under severe artificially injected temporal misalignment. --- ### On the post-hoc Evaluation of PDE Discovery: A Multifaceted Challenge of Scientific Advancement Authors: Baptiste Mathevon, Farah Cherfaoui, Amaury Habrard, Marc Sebban Published: 2026-07-26 Categories: cs.LG arXiv: https://arxiv.org/abs/2607.23753v1 Partial differential equation (PDE) discovery aims to identify from data the governing law of a physical system. Constituting a cornerstone of scientific advancement, it has become during the past decade a major line of research in the rapidly evolving field of Physics-informed Machine Learning (PiML). Among the remaining open problems to address in this domain, the post-hoc evaluation of discovered PDEs raises the particular difficulty of being multifaceted. Indeed, it requires jointly considering predictive accuracy, physical consistency, interpretability, and out-of-distribution generalization capacity. Given that some of these properties are conflicting, it is worth noting that the wide range of existing evaluation metrics only partially address the overall problem, potentially leading to overly interpreted conclusions about the validity of a presumed new physical theory. From an abundant literature spanning machine learning, numerical analysis, information theory or symbolic regression, we propose, to our knowledge, the first taxonomy of PDE evaluation metrics, and discuss their advantages and limitations in depth. Based on the observation that evaluation is often achieved on a case-by-case basis and that a universally accepted methodology remains elusive, we further provide recommendations with the aim of promoting standardized and reliable practices, before sketching promising future lines of research in this field. We argue that this paper is intended both for ML experts who design new PDE discovery algorithms and for users of these methods aiming, in real applications, to discover and validate well-founded scientific laws. --- ### Soft-Constrained Optimization of Latent Space in Variational Autoencoders Authors: Ye Shi Published: 2026-07-26 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2607.23751v1 The usefulness of a variational autoencoder (VAE) depends on two properties of its latent space that are hard to obtain together: high encoding capacity in the individual latent variables, and a low-dimensional, disentangled organization of those variables. Weakening the Kullback-Leibler regularization raises capacity but degrades disentanglement, while strengthening it prunes latent variables away entirely. We formulate VAE training as a soft-constrained optimization problem that addresses both. First, we impose an entropy-based constraint (EC) on individual latent variables, showing that the entropy of a latent code upper-bounds the mutual information it carries about the generative factors of the data. Second, we propose a weight-filter method that exploits the slack of the soft constraint to prune low-entropy dimensions during downstream training. On dSprites, the EC raises the aggregate latent-variable activation score by 43-62% over a vanilla VAE, attains the highest FactorVAE score among the \b{eta} \b{eta}-VAE variants (0.891 vs 0.847), and lowers reconstruction error by up to 38%. On MNIST, the weight filter reduces the latent dimensionality supplied to a downstream classifier from ten to two while holding accuracy above 90%, converging in 37% fewer epochs than the same procedure without the EC. We also find that low-entropy discrete factors tend to merge into a single latent variable, whereas high-entropy continuous factors are distributed across several. --- ### ZenGen: Social Mind for LLMs Authors: ZenGen Team, Ao Xiang, Bi Jingping, Chen Jiahui, Chen Lehan, Chen Yilin, Cheng Xueqi, Fan Yixing, Gan Kairong, Gao Haowen, Gao Jinhua, Gao Shuxuan et al. Published: 2026-07-26 Categories: cs.CL arXiv: https://arxiv.org/abs/2607.23740v2 As large language models move from isolated task solving toward long-term service in human environments, they require social intelligence: the ability to infer mental states, track social relations, reason over norms, and adapt behavior under context. This report presents ZenGen, an integrated framework for measuring, internalizing, and grounding social intelligence. For measurement, we introduce SoMBench, a psychology-grounded benchmark spanning 3 primary dimensions, 17 secondary dimensions, and 71 task paradigms. It controls question format, narrative perspective, and context length across 284 shared scenarios and 3,481 expert-verified instances. Evaluation of 20 representative LLMs reveals substantial headroom: the best model achieves only 72.08% overall accuracy, and none of the 17 secondary dimensions reaches the 90% near-ceiling band. For internalization, we develop ZenGen, a diagnosis-driven training recipe combining supervised fine-tuning, on-policy distillation, and rubric-based reinforcement learning. Across five social-cognition benchmarks, ZenGen consistently outperforms its base models, with ZenGen-27B-Stage2 achieving the best average score and ZenGen-32B-Stage2 remaining competitive with DeepSeek-V4-Pro. For deployment-time grounding, we build Actio, a harness-controlled inference architecture that routes four typed supports into reasoning: PRISM for procedural guidance, Starling for runtime mental-state representation, SAGE for reusable experience, and gated RAG for external social and normative knowledge. Across five base models and three benchmarks, the full harness improves 14 of 15 model-benchmark pairs and is best or tied for best in 8, demonstrating the effectiveness of typed runtime support. Together, these results show that socially intelligent LLMs require coordinated advances in evaluation, parametric internalization, and deployment-time grounding. --- ### Zing: Social Mind for LLMs Authors: Zing Team, Ao Xiang, Bi Jingping, Chen Jiahui, Chen Lehan, Chen Yilin, Cheng Xueqi, Fan Yixing, Gan Kairong, Gao Haowen, Gao Jinhua, Gao Shuxuan et al. Published: 2026-07-26 Categories: cs.CL arXiv: https://arxiv.org/abs/2607.23740v1 As large language models move from isolated task solving toward long-term service in human environments, they require social intelligence: the ability to infer mental states, track social relations, reason over norms, and adapt behavior under context. This report presents Zhijing, an integrated framework for measuring, internalizing, and grounding social intelligence. For measurement, we introduce SoMBench, a psychology-grounded benchmark spanning 3 primary dimensions, 17 secondary dimensions, and 71 task paradigms. It controls question format, narrative perspective, and context length across 284 shared scenarios and 3,481 expert-verified instances. Evaluation of 20 representative LLMs reveals substantial headroom: the best model achieves only 72.08% overall accuracy, and none of the 17 secondary dimensions reaches the 90% near-ceiling band. For internalization, we develop Zing, a diagnosis-driven training recipe combining supervised fine-tuning, on-policy distillation, and rubric-based reinforcement learning. Across five social-cognition benchmarks, Zing consistently outperforms its base models, with Zing-27B-Stage2 achieving the best average score and Zing-32B-Stage2 remaining competitive with DeepSeek-V4-Pro. For deployment-time grounding, we build Actio, a harness-controlled inference architecture that routes four typed supports into reasoning: PRISM for procedural guidance, Starling for runtime mental-state representation, SAGE for reusable experience, and gated RAG for external social and normative knowledge. Across five base models and three benchmarks, the full harness improves 14 of 15 model-benchmark pairs and is best or tied for best in 8, demonstrating the effectiveness of typed runtime support. Together, these results show that socially intelligent LLMs require coordinated advances in evaluation, parametric internalization, and deployment-time grounding. --- ### Source-Free Controlled Adaptation of Teachers for Continual Test-Time Adaptation Authors: Anurag Roy, Riddhiman Moulick, Vinay Kumar Verma, Saptarshi Ghosh, Abir Das Published: 2026-07-26 Categories: cs.LG, cs.CV arXiv: https://arxiv.org/abs/2607.23735v1 In many real-world scenarios, encountering continual shifts in domain during inference is very common. Consequently, continual test-time adaptation (CTTA) techniques leveraging a teacher-student framework have gained prominence, allowing models to adapt continuously even after deployment. In such a framework, a weight-averaged mean teacher is used to produce pseudo-labels from test data for self-training. The mean teacher gets updated as an exponential moving average of the student parameters using a high value of momentum that is kept fixed even if different distributions of test data are encountered. To combat the resulting drift of the model, we propose a novel controlled teacher adaptation methodology that dynamically sets a proper momentum value depending on the quality of the incoming data. Additionally, we estimate class prototypes from the source pretrained model to help align the target data as they come in. Importantly, our method does not require access to source data or its statistics at any stage of the pipeline, making it truly source-free. We perform extensive experiments on benchmark datasets to demonstrate that our approach outperforms different state-of-the-art adaptation frameworks, many of which require access to source data. --- ### TRUAV: Distributed Multi-Agent Reinforcement Learning for Trajectory Planning and Routing Enhancement in UAV-Aided IoT-Enabled VANETs Authors: Muhammad Umar Farooq Qaisar, Lin Zhang, Zhen Chen, Wajdy Othman, Shehzad Ashraf Chaudhry, Chang Liu Published: 2026-07-26 Categories: cs.NI, cs.LG arXiv: https://arxiv.org/abs/2607.23734v1 Unmanned aerial vehicles (UAVs) have emerged as a key enabler of next-generation Internet of Things (IoT) ecosystems, offering flexible aerial relaying to extend connectivity across dynamic vehicular ad hoc networks (VANETs) in smart city environments. However, conventional centralized approaches for UAV trajectory planning require continuous global network state aggregation, making them impractical under bandwidth and energy constraints typical of dense urban deployments. In this article, we present TRUAV, a distributed multi-agent reinforcement learning framework based on independent tabular Q-learning for joint UAV trajectory planning and routing enhancement in UAV-aided VANETs. Each UAV is equipped with a local Q-learning agent that operates purely on locally observable information, including vehicle density, packet queue states, and neighbor UAV positions, thereby eliminating the need for global state exchange. A potential-game-inspired reward design encourages spatial diversity and routing-aware UAV positioning among interacting agents while accounting for energy consumption. Numerical simulations over a large urban area with 200 mobile vehicles show that the proposed TRUAV framework achieves network coverage and packet delivery ratios comparable to centralized deep reinforcement learning methods, while also improving relay delay and energy efficiency. Finally, we discuss emerging challenges and future research directions for distributed multi-agent UAV-assisted IoT systems. --- ### AI Strategy: How to Choose What AI Product to Implement Authors: Foster Provost, Panos Ipeirotis Published: 2026-07-26 Categories: cs.CY, cs.AI, cs.LG, econ.GN, stat.AP arXiv: https://arxiv.org/abs/2607.23733v1 Firms struggle to choose AI projects that pay off: two projects can look equally promising to smart, motivated stakeholders and yet deserve opposite decisions. At the residential real-estate brokerage Compass, one AI product (Likely-to-Sell recommendations) flagged sales outreach opportunities and went on to account for nine figures in annual gross commission revenue. Another championed AI product (a Time-on-Market pricing tool) was rightly shelved. A simple ROI estimate could not distinguish the two. We present expected ROI (eROI), a framework that decomposes each bet into three components and rates them separately: Value if Successful, Likelihood of Success, and Investment Required. Each maps to a question executives can answer before building: How valuable would it be if it worked? How likely is it to work? And what would it cost to implement? Separating the three breaks a common catch-22: teams cannot estimate ROI until they know whether a project will work, yet cannot know whether it will work without building it. Judging Value if Successful on its own dissolves the loop, letting a team argue that a product would be valuable if it worked while it weighs how likely that is. The framework also asks, before ranking anything, whether there are enough good ideas on the table. After ranking, it guides assembling a portfolio of bets rather than funding only the single top-ranked project. We illustrate eROI on Compass's candidate AI products. Precise ROI estimates are hard to make given the inherent uncertainty of AI projects. Coarse business-level ratings of the three components are enough to tell strong bets from weak ones. --- ### An Exact Counterexample to Carlson's Associated-Prime Depth Conjecture from a Group of Order 128 Authors: Xinan Dai, Wenhao Deng, Yingdong Shi, Tailin Wu, Yuchen Yang Published: 2026-07-26 Categories: math.GR, cs.AI arXiv: https://arxiv.org/abs/2607.23732v1 In Question~3.1 of his 1995 paper on depth and transfer, Carlson asked whether the depth of a finite-group cohomology ring is always realized by the dimension of one of its associated primes. We give a negative answer. Let \( G=\SG{128}{859},\qquad k=\kbar. \) An exact presentation certificate proves that $\depth H^*(G;k)=2$. Okuyama's associated-prime theorem would convert an associated prime of dimension two into a rank-two elementary abelian subgroup $E\leq G$ satisfying $\depth H^*(C_G(E);k)=2$. We enumerate all $75$ rank-two elementary abelian subgroups of $G$ and obtain six centralizer types. Duflot's theorem gives depth at least three for four types, while exact ideal-quotient certificates exhibit regular sequences of length three for the remaining two. Hence every rank-two centralizer has cohomological depth at least three, so $H^*(G;k)$ has no associated prime of dimension two. The finite group presentation, the three cohomology-ring presentations, the enumeration summary, and the exact algebraic certificates are included for independent verification. --- ### Outcome-Confounded Local Supervision in On-Policy Distillation Authors: Guoqing Ma Published: 2026-07-26 Categories: cs.LG arXiv: https://arxiv.org/abs/2607.23731v1 On-policy distillation (OPD) trains a student on its own trajectories while a teacher supplies dense token-level likelihoods at student-visited prefixes. These likelihoods are often read locally: agreement appears safe to imitate, whereas disagreement appears to identify an error. We show that both readings are confounded by the outcome of the completed trajectory. We introduce an outcome-resolved diagnostic that crosses pointwise teacher-student divergence with final-answer correctness, separating safe imitation, productive divergence, harmful divergence, and agreement-on-failure. In an eight-seed mathematical-reasoning study with a Qwen3-8B student and Qwen3-32B teacher, agreement-on-failure constitutes 67.84% of pooled response-token mass; with a Qwen2.5-7B/32B pair it remains 67.68%. The result persists across threshold, sequence-level, format, and truncation audits. Even on prompts that the Qwen3 teacher solves in all four independent attempts, student accuracy rises to 86.91% but agreement-on-failure remains 14.76%. We then run three matched training probes that use the available signals to imitate, mask, or contrast whole trajectories; none consistently reduces agreement-on-failure. The result points to a localization limitation: local divergence paired with a trajectory-level outcome does not identify where a failed trajectory became unrecoverable. Addressing this limitation requires additional positional information, such as process labels, teacher continuations from student prefixes, or token-level alignment across rollouts. Our contribution is therefore diagnostic rather than a new training method. --- ### Hierarchical Soft Actor-Critic for Sparse-Reward Long-Horizon Reinforcement Learning Authors: Zahra Abdalla Elashaal, Afef Hfaiedh, Nahla Khraief, Issmail Ellabib, Giansalvo Cirrincione Published: 2026-07-26 Categories: cs.RO, cs.LG arXiv: https://arxiv.org/abs/2607.23726v1 Exploration in sparse-reward long-horizon tasks poses significant challenges for reinforcement learning. To address these challenges, we propose a two-level Hierarchical Reinforcement Learning (HRL) framework. The first level handles high-level strategic planning, while the low-level uses the continuous-control Soft Actor-Critic (SAC) algorithm, and they utilize entropy-regularized policy optimization. The proposed framework was trained and evaluated using the Search-and-Rescue-2 (SAR-2) dataset. HRL-SAC effectively addresses sparse-reward long-horizon search problems characterized by delayed rewards and continuous control, and its outperforming the flat SAC baseline reinforcement learning in terms of success rates, coverage efficiency, and convergence. These findings indicate that hierarchical entropy-regularized policies are a promising solution to tackle long-horizon sparse-reward reinforcement learning tasks. --- ### Recovery of latent inner products from an anisotropic Gaussian random geometric graph Authors: Cheng Mao, Vidya Muthukumar Published: 2026-07-26 Categories: math.ST, stat.ML arXiv: https://arxiv.org/abs/2607.23723v1 We study the problem of recovering latent inner products from a random geometric graph with anisotropic Gaussian latent points. More precisely, for an i.i.d. sample $x_1, \dots, x_n \sim N(0,Σ)$ where $Σ\in \mathbb{R}^{d \times d}$, an edge $(i,j)$ is present in the graph if and only if $\langle x_i, x_j \rangle \ge ζ$ for a threshold $ζ$. We assume the threshold $ζ$ to be chosen such that the average edge density of the graph is of constant order. To address the undesired degree fluctuations amplified by the anisotropy of the latent points, we consider the doubly centered adjacency matrix of the graph, and estimate the latent inner products using a rank-$d$ spectral approximation of the doubly centered matrix. The estimator obtains a mean squared error with a rate involving the stable rank of the covariance matrix $Σ$. Notably, the rate of estimation matches the state of the art for the isotropic case $Σ= I_d$, and permits an ill-conditioned covariance matrix with a diverging condition number. The analysis of the spectral method proceeds via the entrywise Hermite expansion of the doubly centered adjacency matrix with respect to the latent inner products. Instead of the standard trace method, it uses a decoupling argument recently introduced by Kaushik, Romberg, and Muthukumar (2025) to control nonlinear error terms. --- ### E-Bench: Benchmarking Multi-Step Tool-Use Agents in Real-World Product Scenarios Authors: Weihuang Zheng, Tianyuan Zou, Eileen Ye, Alphet Liu, Youyong Kong, Ya-Qin Zhang, Duran Zheng, Maxm Pan Published: 2026-07-26 Categories: cs.AI arXiv: https://arxiv.org/abs/2607.23722v1 Large Language Models (LLMs) are increasingly deployed as agents that interact with stateful environments over multiple steps: gathering hidden information, composing tool calls, and committing state changes. We refer to this capability as multi-step tool use. Existing benchmarks have advanced tool-use agent evaluation, but often focus on isolated API calls, short trajectories, or settings that are difficult to scale or control. We introduce E-Bench, a fully synthetic benchmark with 323 state-changing tasks across three product domains: Honor of Kings, QQ Music, and Tencent Meeting. E-Bench decouples environment synthesis from task synthesis: graph-guided database filling builds reusable, orphan-free product environments, while generator-solver asymmetry creates tasks with both an information gap and a tool gap, requiring agents to discover hidden data and compose multiple tool calls before changing state. Outcomes are graded deterministically by database-state diffs. Since both environments and tasks are synthetic, E-Bench is controllable at the environment level and scalable at the task level. Benchmarking 11 cutting-edge LLMs shows that multi-step tool use remains challenging: Pass^3 stays below 60% for the strongest models, and even with code execution in the E-Bench-Code extension, reliability (Pass^3) remains below 70%. --- ### Distributional Split Criteria for Random Forests: Extensions, Shrinkage, and the Robustness of Mean Splitting Authors: Silas Koemen Published: 2026-07-26 Categories: stat.ML, cs.LG, stat.ME arXiv: https://arxiv.org/abs/2607.23721v1 Distributional random forests replace mean-based CART splitting with criteria that compare the full conditional response distribution in candidate children. We implement and systematically study a family of such criteria inside a single honest-forest implementation: isotropic random-Fourier-feature maximum mean discrepancy (MMD), an anisotropic diagonal-bandwidth variant, an adaptive per-split frequency-selection variant, and a non-kernel sliced-Wasserstein criterion, together with post-hoc kernel-mean shrinkage of the forest weights. Using paired-seed comparisons across synthetic quantile mechanisms, real univariate benchmarks, a California-housing subsample curve, and multivariate synthetic and real responses, we characterize where each extension pays. Three findings recur. First, among distributional criteria ordinary isotropic MMD is already close to best in class: the anisotropic, adaptive-frequency, and sliced-Wasserstein extensions, and post-hoc shrinkage, do not systematically improve on it. Second, on scalar tabular regression mean-based CART splitting remains the robust default and wins many cells. Third, multivariate responses are the regime where distributional splitting clearly earns its keep, most sharply on a pure-dependence copula where the energy score separates the criteria even though marginal CRPS does not. The evidence supports a simple allocation story: distributional splitting helps only when non-location structure is both present and estimable; otherwise it dilutes split-selection power away from the mean. All criteria, the honest forest, and the paired-comparison harness are implemented in the open-source \texttt{drforest} library, whose Rust-backed split search makes broad criterion sweeps inexpensive. --- ### Formally Verified Synthesizable Floating-Point Data Types in ARCH HDL Authors: Shuqing Zhao Published: 2026-07-26 Categories: cs.CL, cs.PL arXiv: https://arxiv.org/abs/2607.23715v1 We report the design and end-to-end verification of first-class IEEE-754 binary32 (FP32) and bfloat16 (BF16) arithmetic for ARCH, a hardware description language intended to be generated by language models. Every operator - comparisons, conversions, add, sub, mul, and fused multiply-add (FMA) - is described once against a single bit-vector IR and rendered three ways from one source: synthesizable SystemVerilog, an SMT-LIB model, and a Lean 4 proof model. The three artifacts cannot drift apart structurally, and the residual per-node printer correspondence is machine-checked: a Yosys-to-SMT miter proves the emitted SystemVerilog equivalent to the SMT model for all 24 operators. Verification splits at the solver-tractability frontier: multiplier-free operators (comparisons, add/sub over all 2^64 inputs, conversions, and all binary BF16 arithmetic) are proved exhaustively equivalent to the SMT-LIB FloatingPoint theory; the SAT-hard multiplier-bearing operators (FP32 mul and FMA) are proved correctly rounded in Lean, sorry-free, against a value-level round-to-nearest-even specification over exact dyadic values. Physical characterization exposed the FMA as the timing outlier: its exact-wide 470-bit datapath does not pipeline in our flow. We reimplemented it as a bounded 98-bit guard/round/sticky datapath that pipelines to 268 MHz on Nangate45, and proved, in Lean and over all 2^96 inputs, that it is bit-identical to the exact-wide reference, so it inherits the reference's proven correct rounding. The equivalence is tractable precisely because the shared multiplier appears on both sides and cancels: neither a SAT solver nor the proof ever solves a multiplier equivalence. (The BF16 FMA is deliberately an FP32-accumulating fusion, characterized as exactly that.) All machine-checked claims are pinned to a tagged open-source release. --- ### The Intruder Threshold: A Spectral Law for LoRA Fine-Tuning Authors: Peng Xie, Amr Alanwar Published: 2026-07-26 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2607.23711v2 LoRA fine-tuning can create intruder dimensions: new leading singular vectors of the updated weight matrix $W+BA$ that are nearly orthogonal to all pretrained singular vectors and that drive catastrophic forgetting. Since their discovery, no theory has predicted, layer by layer on measured spectra, when they appear. We derive a per-layer critical update strength $s^\ast=\barθ/(γσ_1(BA))$, computed from the measured spectrum of $W$ alone through the rectangular spiked-deformation transform, together with an exact secular-equation characterization of the updated spectrum, with no fitted parameters. In a pre-specified study spanning four dense Transformer families, a state-space model, a mixture-of-experts model, and an encoder-decoder (18 adapters, 9{,}840 layer scans), the law localizes the empirical threshold within a factor of two on $82\%$ of layers, separates intruder-bearing from intruder-free layers at deployment with a mean AUC of $0.89$, holds unchanged on six third-party adapters, and predicts where WikiText-2 perplexity begins to degrade; a combination of the two pre-specified edge evaluations reaches $98\%$ and is confirmed out-of-bag on the external adapters ($0.997$). Full fine-tuning disperses its update far below the threshold of every layer, which resolves the asymmetry between LoRA and full fine-tuning. Norm-matched interventions confirm that threshold-crossing layers, rather than update magnitude, carry the forgetting, and a spike-budget rule derived from the thresholds, requiring one SVD and no validation sweeps, reduces forgetting by $62\%$ on the most fragile model at no task cost. --- ### The Intruder Threshold: A Spectral Law for LoRA Fine-Tuning Authors: Peng Xie Published: 2026-07-26 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2607.23711v1 LoRA fine-tuning can create intruder dimensions: new leading singular vectors of the updated weight matrix $W+BA$ that are nearly orthogonal to all pretrained singular vectors and that drive catastrophic forgetting. Since their discovery, no theory has predicted, layer by layer on measured spectra, when they appear. We derive a per-layer critical update strength $s^\ast=\barθ/(γσ_1(BA))$, computed from the measured spectrum of $W$ alone through the rectangular spiked-deformation transform, together with an exact secular-equation characterization of the updated spectrum, with no fitted parameters. In a pre-specified study spanning four dense Transformer families, a state-space model, a mixture-of-experts model, and an encoder-decoder (18 adapters, 9{,}840 layer scans), the law localizes the empirical threshold within a factor of two on $82\%$ of layers, separates intruder-bearing from intruder-free layers at deployment with a mean AUC of $0.89$, holds unchanged on six third-party adapters, and predicts where WikiText-2 perplexity begins to degrade; a combination of the two pre-specified edge evaluations reaches $98\%$ and is confirmed out-of-bag on the external adapters ($0.997$). Full fine-tuning disperses its update far below the threshold of every layer, which resolves the asymmetry between LoRA and full fine-tuning. Norm-matched interventions confirm that threshold-crossing layers, rather than update magnitude, carry the forgetting, and a spike-budget rule derived from the thresholds, requiring one SVD and no validation sweeps, reduces forgetting by $62\%$ on the most fragile model at no task cost. --- ### The Illusion of Secure LLM Code: Closing the Security Gap via Iterative Reprompting Authors: Ishpuneet Singh, Shreyas Mahajan, Gurjot Singh, Maninder Singh Published: 2026-07-26 Categories: cs.CR, cs.AI, cs.HC, cs.LG, cs.MA arXiv: https://arxiv.org/abs/2607.23710v1 Large Language Models (LLMs) are increasingly integrated into software development workflows, yet their ability to autonomously generate secure authentication code remains uncertain. This paper evaluates the security architecture of authentication systems generated by five prominent AI coding assistants through a bi-modal assessment framework combining static code analysis and dynamic penetration testing, mapped to NIST SP 800-63B guidelines. The study examines model behavior across four prompting strategies Basic, Secure, NIST-Based, and Reprompting to reflect varying levels of developer guidance. Empirical results demonstrate that code generated from functional or generically secure prompts consistently omits critical protections, particularly concerning brute-force resistance, session management, and robust password handling. While providing explicit, single-shot NIST context significantly improves compliance, the findings reveal that this remains structurally inadequate. Instead, iterative Reprompting: forcing models into a contextual self-auditing loop is strictly required to achieve a comprehensive, defense-in-depth security architecture. Ultimately, this study proves that current AI coding assistants do not produce secure-by-default applications, dictating that enterprise deployments must transition from single-shot prompt engineering to continuous, standards-driven verification pipelines. --- ### LabRobFail: A Benchmark for Robotic Failure Analysis in Chemical Self-driving Laboratory Authors: Haobo Wang, Baoli Sun, Anqi Zou, Dongsheng Huang, Zelin Lv, Ning Wang, Rui Li, Dongzhan Zhou, Weiyu Guo, Zhihui Wang, Wanli Ouyang Published: 2026-07-26 Categories: cs.RO, cs.CV arXiv: https://arxiv.org/abs/2607.23704v2 The deployment of embodied agents in self-driving laboratories could accelerate scientific discovery, yet their reliability is constrained by the irreversible and safety-critical nature of chemical experiments. Progress is further hindered by scarce failure data and the lack of fine-grained evaluation protocols. To address these challenges, we introduce LabRobFail, a failure-centric framework for learning and evaluating robotic failure analysis in chemical laboratories. LabRobFail-Sim injects controllable failures at the control, physics, and semantic levels, enabling the construction of LabRobFail-Data, which contains over 20,000 trajectories across 70+ task scenarios, five failure categories, and 11 fine-grained failure types. LabRobFail-Bench evaluates six capabilities spanning task understanding, failure detection, temporal localization, severity assessment, failure classification, and actionable correction. We further develop LabRobFail-VLM, a domain-specialized vision-language model that generates structured failure diagnoses and recovery instructions. On seen environments, it achieves 90.83% failure-detection accuracy and 77.21% temporal-localization accuracy, substantially outperforming general-purpose VLMs. When integrated as a real-time supervisor, it improves downstream task success rates by 4-16 percentage points, demonstrating the value of fine-grained failure understanding for closed-loop recovery and reliable laboratory autonomy. Our code and data are available at https://github.com/Su-ISE-2001/SciRobo --- ### Offline-Online Curriculum RL for Multimodal Reasoning Authors: Wendi Deng, Hang Du, Guoshun Nan, Haokun Tian, Jiaqi Yu, Xinlei Cao, Jaile Li, Jingfeng Chen, Ling Deng, Ting Li, Hao Yang, Jun Liu et al. Published: 2026-07-26 Categories: cs.AI arXiv: https://arxiv.org/abs/2607.23700v1 Multimodal large language models exhibit capabilities on reasoning tasks, yet often produce flawed intermediate steps while yielding correct final answers. This behavior undermines interpretability and reliability, suggesting reliance on spurious shortcuts rather than faithful reasoning. Although efforts have explored step-level supervision, distinguishing decisive steps from redundant ones remains challenging. We propose $O^2$-CritiCuRL, a novel curriculum reinforcement learning framework that introduces critical-step awareness through an iterative offline-online paradigm. In the offline stage, $O^2$-CritiCuRL conducts multi-rollout analysis over step-annotated trajectories to estimate step-level importance, allowing the framework to distill critical reasoning steps and filter out redundant ones. In the online stage, we employ a progressive step-level reinforcement learning strategy, where truncated chains guide the model to infer missing steps and refine its reasoning, thereby sharpening its focus on critical steps and overcoming the limitations of static supervision. Extensive experiments on multimodal reasoning benchmarks show that our method achieves state-of-the-art performance while delivering superior training and inference efficiency. Code is available at https://github.com/kk0013/CritiCuRL. --- ### Offline-to-Online Creative Optimization with Generative Models and Adaptive Testing Authors: Kevin Lee, Benjamin Letham, Zhiyuan Jerry Lin, Elodie Samson, Eric Onofrey, Poppy Zhang, Shawndra Hill, Eytan Bakshy Published: 2026-07-26 Categories: cs.AI, cs.LG arXiv: https://arxiv.org/abs/2607.23696v1 Ad creative optimization is increasingly constrained by evaluation rather than generation. Generative models can produce many plausible creatives, but reliable evaluation requires online experiments, in which only a limited slate can be tested. We study how to use data from historical A/B tests to generate and select the candidates in that slate. We developed and deployed a performance-driven offline-to-online workflow that guides creative generation with a predictive model as an inference-time critic. In the offline phase, we use a predictive model trained on historical experiments to rank and refine variants created by a generative model. A final test slate is then deployed in an online adaptive experiment. In a 50-arm field experiment, we found that the best creative generated with this method yielded 45.1% higher engagement than the best human-authored creative. Two additional experiments showed the same upper-tail pattern, with lifts of 46.7% and 36.2%. We found that despite the predictive model being too noisy to directly identify the best creative offline, it effectively guides the generative model toward creating strong candidates that can be efficiently evaluated in an adaptive experiment. The results suggest a design principle for creative optimization with generative models: use predictive models to guide generation of a slate to test, judge the slate by whether it contains high-performing candidates at a feasible test size, and use adaptive experiments to select among candidates while limiting traffic lost to weak arms. ---