The paper introduces Varest, a method for image-token compression that reduces computational load while preserving semantic integrity. Varest uses variable-resolution tokenization guided by semantic fungibility, merging tokens within coherent regions coarsely and maintaining finer distinctions at object boundaries and text. After early decoder processing, learned relevance registers prune visual tokens, enabling a single checkpoint to perform well across multiple budgets and backbones, achieving 86.10‑92.33% of full-token performance with only 5.56% of the original token count.
By Tianhao Fu, Austin Wang, Charles Chen, Roby Aldave-Garza, Victor Xie, Xiyue Jiao, Yucheng Chen, Martin Tin
ROAR is a framework that unifies and analyzes runs from heterogeneous AI‑driven research systems (ADRS). It normalizes diverse output formats into a relational schema while preserving data lineage, enabling cross‑system analytics on over 900 collected runs. The unified corpus reveals problem‑specific search behaviors, such as early gains and varying effectiveness of prior‑solution strategies, and can be used to configure future ADRS runs.
By Leo Y. Lin, Vishakha Ramani, Z. Berkay Celik, Paul Castro, Marquita Ellis
SciRerankBench is a benchmark designed to evaluate rerankers in retrieval‑augmented large language models (RAG‑LLMs) for scientific literature question answering. It covers five scientific subjects and introduces three types of question‑context‑answer pairs—Noisy Contexts, Semantically Similar but Logically Irrelevant Contexts, and Counterfactual Contexts—to test noise resilience, relevance disambiguation, and factual consistency. The study evaluates 13 popular rerankers across five LLM families, offering detailed insights into their strengths and limitations.
By Haotian Chen, Qingqing Long, Meng Xiao, Xiao Luo, Wei Ju, Chengrui Wang, Yuanchun Zhou, Xuezhi Wang, Hengshu Zhu
The paper introduces AquaOV255, a large-scale underwater segmentation dataset with 255 categories and over 20,000 images, and establishes the UOVSBench benchmark by combining it with five other datasets. It also presents Earth2Ocean, a training‑free open‑vocabulary segmentation framework that adapts terrestrial vision‑language models to underwater scenes using a Geometric‑guided Visual Mask Generator and a Category‑visual Semantic Alignment module. Experiments on UOVSBench show that Earth2Ocean significantly improves performance while keeping inference efficient.
By Bingyu Li, Tao Huo, Da Zhang, Zhiyuan Zhao, Junyu Gao, Xuelong Li
UnSAMv2 is a self‑supervised extension of the Segment Anything Model that can segment objects at any granularity without human annotations. It discovers mask‑granularity pairs and introduces a granularity control embedding for continuous scale adjustment, requiring only 6K unlabeled images and 0.02% extra parameters. The model outperforms SAM/SAM‑2 across 11+ benchmarks, improving metrics such as 1‑IoU from 58.0 to 73.1 and achieving state‑of‑the‑art performance on language‑guided segmentation.
By Junwei Yu, Trevor Darrell, XuDong Wang
The paper introduces Concept-Aware Visual Anomaly Detection (Concept-Aware VAD), a framework that extends Concept Bottleneck Models to visual anomaly detection. It offers dual-branch architecture: a concept branch for human-interpretable descriptions and a visual branch for pixel-level localization, evaluated on three VAD benchmarks with competitive performance. A user study shows higher effectiveness and lower misleading rates compared to state-of-the-art explainers, and concept-level interventions improve unsupervised detection.
By Arianna Stropeni, Valentina Zaccaria, Francesco Borsatti, Davide Dalle Pezze, Manuel Barusco, Gian Antonio Susto
ECLIPTICA is a framework that makes large language model alignment instruction‑driven and runtime‑controllable, allowing natural‑language alignment instructions to modulate behavior on the fly. It introduces Contrastive Instruction‑Tuned Alignment (CITA), which combines supervised fine‑tuning with contrastive preference optimization anchored to a reference model, ensuring stable policy updates within a shared neighborhood. The authors release an ECLIPTICA benchmark of 3,000 controlled cases and demonstrate that CITA achieves 86.7% instruction‑alignment efficiency on Llama‑3.1‑8B, outperforming DPO, GRPO, and PPO.
By Kapil Wanaskar, Gaytri Jena, Vinija Jain, Aman Chadha, Amitava Das
The paper introduces a neurosymbolic verification framework that audits the internal consistency of vision-language models (VLMs) when generating radiology reports. By autoformalizing free‑text findings into structured propositional evidence and using an SMT solver (Z3) alongside a clinical knowledge base, the system determines whether each diagnostic claim is entailed, hallucinated, or omitted. Evaluation on seven VLMs across five chest X‑ray benchmarks reveals distinct reasoning failure modes—such as conservative observation and stochastic hallucination—that traditional metrics miss, and shows that enforcing solver‑backed entailment can eliminate unsupported hallucinations and improve diagnostic soundness and precision.
By Vikash Singh, Debargha Ganguly, Haotian Yu, Chengwei Zhou, Prerna Singh, Brandon Lee, Vipin Chaudhary, Gourav Datta
The paper introduces a modular framework that builds an entity‑interaction graph from narratives without relying on annotated node labels. It uses a domain‑adapted BERT encoder to identify entity spans, constructs a bipartite graph, and classifies it with a Heterogeneous Graph Transformer. By casting the graph as a hypergraph and applying counterfactual distillation, the method estimates the Individual Treatment Effect of each node, pruning extraneous entities to produce a minimal predictive subgraph while maintaining high accuracy on the PAN Oppositional Narratives benchmark.
By Diego Revilla, Martin Fernandez-de-Retana, Lingfeng Chen, Aritz Bilbao-Jayo, Miguel Fernandez-de-Retana
The paper introduces Spectra, an algorithm for scheduling a single AI traffic demand matrix across multiple parallel optical circuit switches while accounting for reconfiguration delays. Spectra decomposes the demand into weighted permutations, assigns them load‑aware across switches, and then equalizes load via controlled permutation splitting. Experiments on GPT and Qwen MoE workloads show Spectra reduces schedule makespan by 1.4×, 2.7×, and 2.4× respectively compared to state‑of‑the‑art baselines, approaching newly derived lower bounds.
By Kevin Liang, Litao Qiao, Isaac Keslassy, Bill Lin
The paper titled "The Reasoning Trap -- Logical Reasoning as a Mechanistic Pathway to Situational Awareness" argues that advances in logical reasoning for large language models (LLMs) are likely to lead to increased situational awareness, a dangerous emergent capability. It introduces the RAISE framework, outlining three mechanistic pathways—deductive self inference, inductive context recognition, and abductive self modeling—through which logical reasoning can deepen an AI’s self‑awareness. The authors map current LLM reasoning research onto these pathways, show that existing safety measures are inadequate, and propose safeguards such as a Mirror Test benchmark and a Reasoning Safety Parity Principle.
By Subramanyam Sahoo, Aman Chadha, Vinija Jain, Divya Chaudhary
TopoPilot is an agentic framework designed to automate complex scientific visualization workflows, particularly in topological data analysis and visualization. It uses a two‑agent architecture—a user‑prompt orchestrator and a verifier—to translate natural language into atomic backend actions and ensure semantic consistency before execution. The system incorporates deterministic guardrails and a taxonomy of failure modes, achieving over 95% success in 1,000 simulated multi‑turn conversations and outperforming baseline models and other code generation tools in accuracy, completion time, and token usage.
By Nathaniel Gorski, Shusen Liu, Bei Wang
The paper introduces a synthetic data generation pipeline for long‑context audio reasoning, focusing on first‑visit doctor‑patient conversations and SOAP note generation. The pipeline consists of persona‑driven dialogue creation, multi‑speaker audio synthesis with realistic acoustics, and LLM‑based reference note production, all built with open‑weight models. The authors release 8,800 synthetic conversations, 1.3k hours of audio, and reference notes, and show that cascaded approaches still outperform end‑to‑end models on current open‑weight systems.
By Yanis Labrak, David Gr\"unert, S\'everin Baroudi, Jiyun Chun, Pawel Cyrta, Sergio Burdisso, Ahmed Hassoon, David Liu, Adam Rothschild, Reed Van Deusen, Petr Motlicek, Andrew Perrault, Ricard Marxer, Thomas Schaaf
PR-MaGIC is a training‑free, test‑time framework that refines prompts for in‑context image segmentation by using gradient flow from the mask decoder of the Segment Anything Model (SAM). It integrates seamlessly into existing in‑context segmentation pipelines and employs a simple top‑1 selection strategy to stabilize performance. Extensive evaluations show that PR‑MaGIC consistently improves segmentation quality across multiple benchmarks without requiring additional training or architectural changes.
By Minjae Lee, Sungwoo Hur, Soojin Hwang, Won Hwa Kim
The paper proposes a reward‑density principle for language‑model post‑training: sparse sequence‑level rewards are best for exploring and discovering new behaviors in a strong teacher, while dense token‑level teacher supervision is optimal for compressing that behavior into a smaller deployment model. The authors validate this rule with a four‑stage workflow—teacher RL, forward‑KL warmup, on‑policy distillation, and optional post‑bridge student RL—showing that an RL‑improved 8B teacher distilled through a dense bridge outperforms direct GRPO on a 1.7B student across math and code benchmarks. Ablation studies confirm each stage’s contribution, and the approach generalizes to Llama and Qwen models, offering a practical guideline for allocating scarce labeled data.
By Yuanda Xu, Hejian Sang, Zhengze Zhou, Ran He, Zhipeng Wang, Alborz Geramifard
The paper introduces GAC, a noise‑aware adaptive mixing controller that dynamically balances supervised fine‑tuning (SFT) and reinforcement learning (RL) gradients during hybrid post‑training of large language models. By deriving a closed‑form mixing weight that minimizes an MSE upper bound on the mixed gradient, GAC incorporates EMA smoothing, a schedule prior, and capped updates, all estimated online from training tensors. Experiments show that GAC outperforms rule‑based controllers, improving performance on diverse benchmarks by up to 3.8 percentage points, reducing KL‑drift by 28%, and cutting large mixing‑weight changes by over 70% with negligible overhead, with gains scaling from 1.5 B to 14 B parameters.
By Yuelin Hu, Wei Liu, Zhenbo Yu, Zhengxue Cheng, Li Song
ThreadBench is a benchmark that evaluates how well large language model (LLM) agents simulate real-world Reddit discussions, using 9,817 real threads across 8 domains and 116 subdomains. It compares generated and real threads on 12 metrics grouped into uniformity, expression, tone, and interaction. The authors also introduce Thread-Prompt-based Discussion Generation (ThreadPDG), a Planner-Writer framework that learns thread growth from real seeds and achieves high pass rates on ThreadBench, outperforming baseline methods and matching real discussions in 78.7% of human evaluations.
By Yaoning Yu, Ye Yu, John Sciarappo, Haojing Luo, Kai-min Chang, Yi-Chia Wang, Haohan Wang
FinRED is an expert‑guided framework for generating and evaluating financial LLM red‑teaming tests. It maps global financial standards to specific threats, converts real documents into context‑rich prompts, and offers a validated rubric that reduces false negatives. The framework is deployed in South Korea’s Financial Security Institute sandbox and is gated for qualified researchers.
By Daeyoung Park, Chaeyun Kim, Junghwan Kim, Jinyoung Jeong, Eunji Song, Yongtaek Lim, Minwoo Kim
The paper revisits parsimoniously activated dictionary learning (PADL), a method that imposes a global regularization on the number of activated dictionary atoms. It shows that PADL can be reformulated as a maximum a posteriori estimation under a structured generative model, enabling derivation of generalization guarantees and an analytical tradeoff between sparsity, storage cost, and reconstruction accuracy. Leveraging this insight, the authors propose an efficient PADL algorithm that removes manual hyperparameter tuning, improves reconstruction performance on visual benchmarks, and accelerates inference for vision‑language models.
By Zihui Zhao, Yuanbo Tang, Yang Li
The paper introduces Single‑Rollout Pass@k Policy Optimization (SR‑PPO), a reinforcement learning method that uses only one rollout per prompt to train language‑model agents. SR‑PPO combines Monte Carlo token‑level advantage estimation with Pass@k gradient reweighting and an online critic for baseline and weight estimation. Experiments show that SR‑PPO achieves stable learning and performance comparable to traditional Group Relative Policy Optimization (GRPO) on reasoning and logic benchmarks, even when off‑policy drift degrades standard Pass@1 PPO.
By Fengdi Che, Yang Liu, Lei Yu, Meng Cao, Tong Che, Rupam Mahmood, Dale Schuurmans