GraphOPD introduces a graph‑augmented on‑policy distillation method for large language model agents, addressing the shortcomings of existing divergence‑based guidance when rewards are sparse. By constructing a dependency graph from the environment’s state‑change record and applying a random‑walk stationary distribution, the method identifies and prioritizes the most impactful steps in a trajectory, fusing this structural credit with teacher‑student divergence signals. Experiments on ALFWorld, WebShop, and SearchQA demonstrate that GraphOPD consistently outperforms eleven baselines, improving the strongest baseline by up to +5.8 percentage points and showing strong causal alignment through an executed‑replay audit.
By Bohan Lin, Liyi Chen, Zhuoning Guo, Muyang Li, Qimeng Wang, Yan Gao, Yao Hu, Yudong Zhang
HCPN-GCN is a graph-aware extension of Hierarchical Prototype Networks that replaces linear feature extractors with Graph Convolutional Networks and introduces cone-based prototypes with a diversity regularization objective. This design yields richer graph-aware representations, reduces prototype proliferation, and maintains near-zero forgetting. Experiments on six continual graph learning benchmarks show consistent accuracy improvements over the original HPN and other baselines while using roughly 30× fewer atomic prototypes.
By Sammuel R. Silva, Vander L. S. Freitas, Gladston Moreira, Eduardo J. S. Luz, Rodrigo Silva
FinVector-Market-4B adapts the Qwen/Qwen3.5-4B language model using rank‑16 LoRA on a 22,000‑example financial corpus. On a 600‑example benchmark with implicit and explicit JSON‑schema contracts, the adapted model improves JSON validity from 0% to 91.3% and boosts task‑specific metrics—FinQA exact match from 14.7% to 40.0%, calculator‑expression correctness from 48.0% to 82.7%, scenario branch‑label agreement from 20.1% to 89.5%, and implication‑direction agreement from 52.4% to 87.2%. A post‑hoc audit shows that macro‑F1 differences stem from changing label sets, and the benchmark’s generalization claims are limited by filing overlap and calculator‑target inconsistencies.
By Alina Khaybullina
LASER is a new offline reinforcement learning algorithm that performs latent-space adjoint matching with entropy regularization to avoid policy collapse and critic exploitation. It uses expressive flow policies without backpropagation through time and applies a single set of hyperparameters across all tasks. Experiments on 40 OGBench tasks show LASER outperforms baselines, achieving state‑of‑the‑art results with robust, task‑agnostic performance.
By Songyuan Zhang, Oswin So, Eric Yang Yu, Matthew Cleaveland, Peter Crowley-Dolen, Chuchu Fan
The paper introduces ORACLE, a method for constrained learning that evaluates constraint compatibility after the optimizer has taken a step, rather than before. ORACLE performs a joint endpoint linearization of diverse constraint families, constructs an alignment in the optimizer’s geometry, bounds its authority, and commits the adjustment only after validation. Experiments on eight PDE benchmarks with four different optimizers show that ORACLE matches or improves performance in 94% of configurations, outperforming other constraint-handling approaches.
By Utkarsh Grover, Wyatt Mackey, Kaixun Hua, J. Morris Chang, Xiaomin Lin
SPIN (Shadow Predictive Indexer) is a new method for sparse attention that reduces the overhead of scoring the entire key-value (KV) cache during decoding. By using lightweight, history-based predictions to identify important KV blocks, SPIN avoids full cache scoring at each step and treats KV blocks and speculative decoding as core design elements. Experiments on long-context and agentic benchmarks show 30‑40% sparsity with maintained task quality, and in vLLM serving it boosts throughput by up to 14.9% while cutting median inter‑token latency by up to 13.2%.
By Yao Fu, Cyrus Chang, Ritchie Zhao, Bryce Long, Yueying Li, Mahdi Kamani, Samkit Jain, Rahul Raman, Tara Safavi, Shreya Gupta, Parsa Ashrafi Fashi, Minseok Lee, Julien Demouth, Bita Darvish Rouhani
The paper introduces Universal Attention, a transformer architecture that unifies and extends decay-based KV‑cache pruning mechanisms while preserving RoPE embeddings and Softmax attention. It achieves highly expressive, end‑to‑end trainable pruning that adaptively removes low‑contribution tokens, yielding up to 10× compression on standard datasets and 25× compression at 16k context length without sacrificing downstream performance. The approach outperforms existing state‑of‑the‑art baselines and unpruned models in both compression and generalization.
By Davis Wertheimer, Haochen Shen, Ahan Gupta, Derrick Liu, Yu Chin Fabian Lim, Mudhakar Srivatsa, Raghu K. Ganti, Minjia Zhang, Naigang Wang
The paper introduces EDiS, an Edge‑Disjoint Subgraph sparsification framework that separates a one‑time structural extraction of a graph into cacheable edge‑disjoint subgraphs from the per‑epoch graph composition. By recombining these cached subgraphs under edge‑budget constraints, EDiS avoids repeated sampling or recomputation while allowing varying training graphs across epochs. Experiments on 19 node‑classification benchmarks show that EDiS outperforms 17 baselines in accuracy/ROC‑AUC, achieving the lowest average rank and gap‑to‑best, especially at tight edge budgets.
By Sai Karthik Navuluru, Siddhartha Shankar Das, Franck Dernoncourt, S M Ferdous, Ryan A. Rossi, Nesreen K. Ahmed, Baris Coskunuzer, Alex Pothen, Lakshman Tamil, Mahantesh M Halappanavar
VideoEvolve is a self‑evolving framework for long‑video understanding that jointly updates both memory and retrieval mechanisms. It starts with a coarse low‑frame‑rate overview and uses a Memory Evolver to selectively augment memory while a Retrieval Evolver adapts how information is fetched from that memory. The two components are alternately refined via Agentic Reinforcement Learning, guided by Bottleneck‑Aware and Capability‑Aware Evolution Feedback to focus on the most limiting aspect and to prevent over‑specialization, respectively. Experiments on several long‑video benchmarks show that this approach improves performance over static systems.
By Yongchao Xu, Bowen Ye, Jiefeng Gan, Junkai Ma, Wenzhao Li, Sen Tao, Yi Wei, Jiawei Liu
The paper introduces Kuration SDK, a toolkit for curating raw gameplay data to address the Virtual2Real gap in action‑conditioned world models. It demonstrates that conventional metrics like FVD, LPIPS, and JEDi do not reliably predict playability, using CounterStrike data to show divergent model performance despite similar scores. By providing curation strategies and diagnostic tools, the SDK helps uncover root causes of the gap and accelerate the creation of sample‑efficient training datasets.
By Nirmit Desai, Eric Song, Mayank Sengupta, Tejal Bedmutha, Siri Reddy, Sahiti Dharmavaram, Kunal Sawarkar
The paper introduces TOPOGRAPHRAG-BENCH, a benchmark for evaluating multimodal Graph Retrieval-Augmented Generation (GraphRAG) systems on layout‑grounded evidence reasoning. It contains 2,024 questions derived from 201 long, visually rich documents, designed around three controlled topologies—single‑hop retrieval, bridge‑chain reasoning, and multi‑source synthesis—while employing counterfactual validation to ensure question integrity. Experiments show that multimodal GraphRAG outperforms text‑only and page‑level visual retrieval models, yet still struggles with incomplete visual‑textual alignment and multi‑unit composition, highlighting the need for GraphRAG systems that explicitly model document layouts and cross‑modal evidence alignment.
By Ruochi Li, Jianzhe Lin, Haoxuan Zhang, Haihua Chen, Junhua Ding, Edward Gehringer, Yang Zhang
The paper introduces a scalable Bayesian method for conditional density estimation using a logistic Gaussian process. It models the log conditional density with a separable covariance— a Matérn kernel on the response dimension and a Nyström-feature kernel on covariates— and samples the latent field via kinetic Langevin dynamics instead of traditional approximations. Experiments on photometric‑redshift data with up to 3.9 million observations show the approach matches state‑of‑the‑art tabular foundation models on density and calibration metrics.
By Daniel Paulin, \'Ad\'am Jung, Andr\'as A. Bencz\'ur
The paper introduces Causal Posterior Estimation (CPE), a new Bayesian inference technique for simulator models where the likelihood is intractable but simulation is easy. CPE uses flow matching to approximate the posterior while embedding the model’s conditional dependencies directly into the neural network architecture. Experiments show that this hard‑coded structure yields highly accurate inference, matching or surpassing current state‑of‑the‑art methods.
By Simon Dirmeier, Antonietta Mira
Physics-informed neural PDE solvers typically keep a fixed representation while adapting parameters, which limits performance on heterogeneous problems. The authors propose physics-informed neural plasticity, allowing the representation itself to reshape during training. Their ReCAP method dynamically redistributes capacity via local enrichment, residual-directed splitting, gate-based pruning, and function-aware merging, achieving the lowest relative $L^2$ error on five challenging 3D and 4D benchmarks and reducing error by 10.7%–27.5% compared to the best competing solvers.
By Chun-Wun Cheng, Bingcheng Hu, Angelica I. Aviles-Rivero
The paper proposes a calibration-first framework that generates supervision scores without ground-truth labels or a shared annotation space. It aligns subset-specific scorers by mapping them onto a synthetic ordinal reference space built from ordered calibration features, enabling the fusion of otherwise incomparable outputs. Experiments on three benchmark datasets show that this method consistently outperforms uncalibrated averaging and matches or exceeds the best individual scorer, with significant improvements after Bonferroni correction.
By Jackson Eshbaugh, Jorge Silveyra
The paper proposes a framework that adapts systematic review methods—planning, conducting, and documenting—to the retrieval of machine learning assets in AI registries. By treating models, datasets, and benchmarks as first‑class units of analysis, the framework incorporates registry‑aware search, cross‑registry schema alignment, and dependency‑driven exploration to make asset selection transparent, reproducible, and evidence‑based. The authors argue that this approach transforms ML asset discovery from an informal, ad‑hoc process into a structured, systematic practice.
By Alexandra Gonz\'alez, Quim Motger, Xavier Franch, Silverio Mart\'inez-Fern\'andez
Dual-QK introduces paired non‑orthogonal transforms for queries and keys to enable sharp queries and flat keys, improving 2‑bit quantization and dynamic channel pruning of KV caches. The method balances key scales for INT2 quantization while concentrating query energy for pruning, and supports channel‑0 protection and bucket‑relative RoPE for long‑context accuracy. Experiments on four models across five generative benchmarks show that Dual‑QK outperforms OSCAR at 40% query‑channel sparsity, achieving 6.8× KV‑cache compression, an 8.3× reduction in KV read volume, and up to 3.75× decoding throughput over unpruned BF16.
By Sunjoo Whang, Jungjun Oh, Minsung Kim, Dongho Seo, Jisu Shin, Gregory Kielian, Hoi-Jun Yoo, Sangjin Kim
The paper presents a nonlinear dissipative dynamical model of human learning, featuring a concept‑transfer matrix, per‑concept forgetting rates, and a saturating practice‑response gain. It proves a structural identifiability theorem that guarantees exact parameter recovery under explicit excitation conditions, provides closed‑form recovery for the two‑concept case, and introduces a fast, L‑stable numerical scheme enabling estimation on large learner cohorts. Empirical results show exact recovery on synthetic data with designed excitation, but failure to recover on large observational benchmarks, illustrating a sharp boundary between identifiable and unidentifiable regimes and offering a validation protocol for interpretability claims.
By Arman Kostanian, Armen Beklaryan
Layerwise Error Attribution for Fast and Robust Mixed-Precision Post-Training Quantization proposes a new allocation method that assigns bits to layers under a global memory budget using a small calibration set. The method derives a layerwise probabilistic analysis separating propagated error from local perturbations, enabling a simple, solver‑free allocation algorithm that is robust to corrupted calibration data. Experiments on DRUNet denoising tasks and quantized diffusion models show significant speed‑ups (up to 2,570×) and improved performance, with PSNR gains of up to 7.5 dB under corrupted calibration.
By Samy Houache (IMB, UB), Yann Traonmilin (IMB, UB), Jean-Fran\c{c}ois Aujol (UB, IMB)
KGATE is a modular Python library for training knowledge graph embedding (KGE) models, built on PyTorch Geometric and TorchKGE. It allows users to assemble or plug in components such as initializers, encoders, decoders, losses, negative samplers, and evaluation metrics, and includes a preprocessing step to prevent data leakage. Benchmarks against six existing KGE libraries show that KGATE’s training time is comparable to the fastest while offering a broader set of features.
By Benjamin Loire, Galadriel Bri\`ere, C\'elia Brahimi, Antoine Toffano, Ana\"is Baudot