The paper introduces “Spend Where It Matters”, a two‑stage, training‑free framework that combines particle search with Metropolis‑Hastings refinement for budget‑constrained audio‑visual alignment. In the first stage, particles are adaptively allocated across denoising checkpoints based on online variance estimates, reducing estimation variance while focusing compute where it matters most. The second stage uses MH refinement with local noise‑and‑denoise proposals to further improve reward, starting from the strong candidate produced by the particle search. Experiments on “JavisBench‑Mini” show that this combined approach yields higher rewards than either particle search or MH alone, achieving significant relative gains over baseline methods. The method also demonstrates benefits for image generation on SD1.5, improving aesthetic and image reward scores compared to uniform particle allocation.
By Shivanshu Shekhar, Raghavendra Addanki, Tong Zhang, Uttaran Bhattacharya
The paper introduces TQS, a Trajectory Quality Score that evaluates coding-agent interaction trajectories based on structural signals such as information focus, intent–action alignment, and policy dynamics, rather than just final outcomes. TQS operates in linear time without requiring an LLM judge and is applied to 155,144 successful trajectories from a public corpus. Fine‑tuning agents on trajectories ranked highest by TQS consistently outperforms models trained on random or low‑ranked subsets on the SWE‑Bench‑Verified benchmark, demonstrating the value of process‑level quality filtering.
By Sida Li, Yuanxin Liu, Shicheng Li, Linli Yao, Xu Sun, Haoli Bai, Xiaohui Li
The paper investigates how visual interventions—training‑free edits applied to multimodal large language models—affect hallucinations and prediction quality. By applying a visual‑residual edit across successive decoder layers and matching its RMS strength, the authors identify an Evidence Anchoring Boundary (EAB) where the intervention’s influence drops sharply; effects before the EAB are 13.78–26.10× larger than after it. They further compare counterfactual subtraction across different question types and introduce calibrated selective routing, which only alters predictions when calibration supports an intervention, improving attribute accuracy by 1.56–1.81 percentage points while avoiding the harms of universal subtraction.
By Qiji Zhou, Junyi Hu, Peiyu Wang, Yue Zhang
The paper shows that Tabular Foundation Models (TFMs) are highly overparameterized, with up to 96% of transformer blocks removable without significant loss in performance. Using a closed‑form, label‑free affine translator, the authors demonstrate that halving the depth of eight leading TFMs costs at most $0.002$ median AUROC, while a single‑block TabPFN‑3 remains competitive. Removing depth saves up to 1.5 B parameters, speeds GPU inference up to 6×, and reduces memory usage by 6× with negligible accuracy loss.
By Irene Cannistraci, Julia E. Vogt
Neurons++ is a visual‑stream framework that reconstructs both static images and dynamic videos from fMRI data. It builds on the earlier Neurons model by extending from video‑only to image‑and‑video reconstruction and replacing parallel task conditioning with a serial, cortex‑inspired decoding pipeline. The approach first pre‑trains on image‑fMRI pairs for stable spatial‑semantic initialization, then fine‑tunes separately for image and video tasks, organizing fMRI representations into spatial, semantic, language‑level, and temporal cues, and achieves state‑of‑the‑art video reconstruction with competitive image results while enabling neural‑to‑visual interpretability and ROI‑guided semantic modulation.
By Haonan Wang, Yuchen Wang, Jingyu Lu, Qixiang Zhang, Lehan Wang, Xuanqi Huang, Xiaomeng Li
The study compares human Project Manager facilitation with Llama 3.1 8B Instruct responses in scenario‑based collaborative design meetings, analyzing 199 paired contexts from the AMI Meeting Corpus. Five behavioral dimensions—contextual semantic distance, lexical elaboration, empathy‑and‑hedge lexical density, topic divergence, and phase‑vocabulary alignment—were examined. Human facilitation showed higher contextual semantic distance and empathy‑and‑hedge density, while Llama facilitation exhibited greater topic divergence, lexical elaboration, and instruction‑conditioned phase alignment.
By Mohammed Azeez Khan, Aaron Dsouza, Ashutosh Mishra, Ummul Faiz Z. Bibi, Neha K. Nair, Amar Kumar Behera
The paper demonstrates that hidden capabilities in AI models can be provably concealed by embedding encrypted neural networks as backdoors within the host model’s weights. These hidden circuits are designed to decrypt and activate only when a secret trigger is provided, and they remain difficult to elicit or interpret even with full access to the model weights, thanks to cryptographic guarantees. The authors implement these constructions in PyTorch and test their resistance against various elicitation techniques on a chemical reaction prediction task, showing that current audit methods cannot reliably uncover such hidden behaviors.
By Andis Draguns, Sumeet Ramesh Motwani, Raymond Douglas, Christian Schroeder de Witt
The article discusses implementing architectural guardrails to ensure AI agents safely read and utilize untrusted sources within their workflows. It outlines strategies for protecting agents from unreliable or malicious information while maintaining functional integration. The focus is on designing robust safeguards that prevent the propagation of errors or harmful content.
By Thuwarakesh Murallie
The study shows that tabular foundation models (TFMs) can be dramatically simplified without losing performance. By replacing up to 96% of transformer blocks with a label‑free affine translator, researchers found that half‑depth models still outperform full‑depth TFMs and tuned GBDTs across 171 classification datasets. The simplification reduces parameters by up to 1.5 B, speeds GPU inference up to 6×, and cuts memory usage by 6× with minimal impact on AUROC.
The paper investigates how large audio‑language models (LALMs) hallucinate objects by affirming sounds that are not present. Using a black‑box causal test, the authors isolate evidence regions via CLAP similarity, remove them with spectral gating, and compare against random windows across a 2×2×2 grid on three LALMs. The study finds that evidence localization improves detection over random removal, that raw detection scores are model‑dependent, and that a contrastive decoding strategy combining blank and language‑prior cues boosts Qwen2‑Audio accuracy from 57.1 % to 71.9 %, outperforming Audio‑Aware Decoding alone.
CliQ‑RRG is a two‑stage framework for generating radiology reports that aligns visual and textual data using disease‑aware embeddings and a Prior‑Guided Attention Module to capture multi‑view dependencies. In the second stage, GraphRAG‑guided retrieval adds graph‑grounded clinical evidence, and a large language model restructures the content into concise question‑answer pairs with diagnostic summaries. Experiments on two chest X‑ray benchmarks show that CliQ‑RRG outperforms previous methods in both lexical and clinical metrics, producing accurate and clinically coherent QA‑style reports.
FlyHack evaluates how mechanistic conclusions drawn from connectome-constrained neural networks depend on model choice, interventions, circuit representation, and training outcomes. Using a 50‑model ensemble of the flyvis motion‑vision system and task‑matched causal CNN controls, the study finds that key behavioral phenomena—such as reverse‑phi reversal and stationary‑sawtooth responses—vary across models and are not consistently reproduced by CNNs. At the mechanistic level, attribution accuracy and circuit conclusions also differ markedly with representation and ensemble member, indicating that mechanistic claims must be scoped to the specific model family and experimental conditions.
The paper introduces CGER-Net, a context‑grounded multimodal framework designed to reduce biased gendered economic role attribution in image‑text memes. By assessing whether sufficient evidence exists for gendered reasoning, the model gates inference, abstaining when cues are ambiguous and confidently attributing roles when explicit. Evaluation on the EconMeme‑GE dataset shows CGER‑Net cuts gender overcommitment by up to 44% on ambiguous memes while preserving accuracy on clear cases, and human judges find 79% of its rationales epistemically aligned with the evidence.
By Kushal Kanwar, Dushyant Singh Chauhan, Kapil Rana, Gopendra Vikram Singh, Nils Lukas
The paper introduces Policy Alignment Membership Auditing (PAMA), a framework for detecting whether a prompt was used during on‑policy distillation (OPD). PAMA leverages the direction of teacher‑guided policy updates, quantified by Teacher Alignment Gain (TAG), alongside student drift and uncertainty signals to assess membership. Experiments on six datasets, including the MATH benchmark, show that PAMA improves AUC by 14.6–20.6% over existing methods.
By Yilong Yang, Wenzhuo Shang, Yule Liu, Jiale Teng, Zhuo Ma
The paper introduces Uni-DSM, a unified score-driven framework that uses denoising score matching to model video anomaly patterns for both detection and anticipation. It employs a shared noise-conditioned score transformer backbone with scene-dependent embeddings and motion-aware weighting, and differentiates between VAD and VAA through autoregressive denoising and a self-distilled denoising mechanism. Experiments on multiple benchmarks show state‑of‑the‑art performance for both tasks while maintaining high efficiency.
By Congqi Cao, Zhenhe Liang, Hanwen Zhang, Yifan Zhao, Qinyi Lv, Lingtong Min, Yanning Zhang
The study investigates a bias in text‑to‑image models that consistently generate analog clocks showing 10:10, a common advertising pose. Researchers tested 52 models on images requiring clocks to display 2:35, 6:50, and 11:20, varying the prompt to include no time, digits, hand positions in words, or a drawn reference dial. Results show that adding a reference dial dramatically improves accuracy (up to 75% correct) while other prompt variations only modestly reduce the 10:10 bias.
By Luca Cazzaniga
The study examines how varying the context window size of GPT‑2 affects its alignment with human reading-time data. A U‑shaped relationship is found: very short windows (<20 tokens) capture local memory limits, while larger windows (500–1,000 tokens) provide the best overall fit. Counterfactual experiments that disrupt cross‑sentential coreference chains show that these long‑range coreference relations are crucial for the improved fit of larger context windows.
By Kohei Kajikawa, Lin Ai, Tatsuki Kuribayashi, Ethan Gotlieb Wilcox
SatFix is a feed‑forward framework that performs absolute visual localization of UAVs in satellite maps using a single oblique image or a short multi‑view clip. It replaces traditional cross‑view retrieval by regressing a continuous 3‑DoF pose—2D position and heading—directly from satellite‑grid features without requiring a 3D map or auxiliary sensors. On the newly introduced University‑Metric dataset, SatFix localizes over half of test frames within 50 m and achieves median errors of 45.66 m and 20.81° with a single view, improving further with multiple views.
By Jiarui Zeng, Kun Shi, Chiman Vong, Zhedong Zheng
The paper introduces ResOT, a training‑free method that mitigates object hallucination in large vision‑language models by repairing hidden representations at inference time. ResOT aligns hallucinated directions with faithful ones through localized distribution alignment and Gaussian optimal transport, creating a low‑dimensional residual subspace for minimal intervention. Experiments on three LVLMs demonstrate significant reductions in hallucination and improvements in image caption quality and multimodal performance.
By Chen Zhao, Xingping Dong, Jiachun Shi, Liang Peng, Chong Wang, Zhen Lei, Ran He, Bo Du
Policy Learning with a Language Bottleneck (PLLB) is a framework that lets AI agents generate linguistic rules to describe high‑level strategies. PLLB alternates between rule generation, guided by language models, and policy updates that use these rules, even when a single rule cannot fully capture a complex policy. Experiments on five tasks—including a signaling game, maze navigation, image reconstruction, and robot grasp planning—show that PLLB agents learn more interpretable, generalizable behaviors and can share rules with humans for better coordination.
By Megha Srivastava, Cedric Colas, Dorsa Sadigh, Jacob Andreas