Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 18 papers for 2026-10-07
The load-bearing idea is real: performance gains in hypergraph learning are often attributed to higher-order information, and this paper tests that attribution directly rather than merely reporting accuracy. The full text supports the claim that many gaps survive the perturbation, and that simple lower-order upgrades like pair multiplicities, structural self-weights, diffusion, and normalization close much of the remaining gap; so the headline explanation is weaker than the usual narrative. That is a strong diagnostic result for the field, though not a new architecture.
The paper should be judged against the hypergraph-learning line that includes HGNN (Feng et al. 2019), HyperGCN (Yadati et al. 2019), HCHA, HNHN, AllSetTransformer, EHNN, HyperGT, and the broader clique-expansion baseline tradition. Those works showed that higher-order models can beat pairwise reductions on benchmarks, but they did not test whether the gain really comes from higher-order information rather than from stronger pairwise processing, normalization, diffusion, or other baseline differences. This paper’s new contribution is a controlled attribution protocol: perturb the hypergraph to preserve weighted clique expansion exactly, then ask how much of the performance gap survives; that is a materially new way to evaluate the claim, not just another model win.
The intervention is tied to preserving weighted clique expansion, so the conclusion is narrower than "higher-order information is unnecessary" in general, and the stronger lower-order baselines are still somewhat handcrafted diagnostics rather than a single unified alternative.
Kandinsky 6.0 Video introduces a family of diffusion models for synchronized text-to-audio-video generation, available in Lite (3B) and Pro (29B) sizes. They produce 5-second videos with synchronized 44 kHz audio, including lip-sync, in T2AV and I2AV modes, and include a built-in super-resolution up to Full-HD. Built on Kandinsky 5.0, Kandinsky 6.0 Video uses a dual-stream CrossDiT architecture to unify video and audio generation.
Diffusion language models can predict multiple tokens in parallel for faster generation, but their quality lags behind autoregressive models due to a computation-difficulty mismatch. ALoDLM replaces uniform per-token computation with token-adaptive depth, allocating more steps to harder tokens while keeping easier ones light. This adaptive approach narrows the quality gap without sacrificing speed.
Memadapter proposes counterfactual adaptation to curb memory-induced sycophancy in long-term memory–driven LLM agents. Persistent memories can bias responses toward users’ historical beliefs, and existing mitigation methods that filter memories are incomplete in real-world settings. Memadapter introduces a counterfactual reasoning component that moderates memory influence, allowing the agent to question or override outdated or conflicting memories when warranted by evidence.
Long-horizon video generation with autoregressive diffusion models suffers from drifting in colors, textures, and motion. Existing KV conditioning helps but is insufficient beyond training horizons because KV entries can drift out of distribution during rollout. In-Distribution Forcing at test time proposes constraining and repairing KV entries to stay within the in-distribution manifold, reducing drift and improving consistency for long videos.
LMBuild provides a benchmark to evaluate LLM agents on generating buildable and functional structures, not just aesthetically pleasing geometry. It represents generated objects as assembled structures with part decompositions and assembly instructions to test physical realizability. The benchmark assesses whether designs can be constructed and perform intended functions in the real world.
We study two-sample tests for random graphs without vertex correspondence and derive sample complexity results. Under Erdős–Rényi null and planted two-block differences, we identify how many graphs per group are needed to detect differences; results show m ~ t^{-3} scaling.
We introduce Neural Petri Flows that integrate Petri net semantics with neural learning for chemical reactions. The architecture preserves valence budgets and enabling rules while enabling differentiable reaction modeling.
We propose GraphPDHG, a neural algorithmic reasoning framework that solves general graph saddle point problems by simulating the Chambolle-Pock primal–dual hybrid gradient method via message passing. We establish theoretical efficiency and provide convergence intuition for this approach.
This work systematically investigates three sources of uncertainty in knowledge graph embeddings: knowledge uncertainty due to incomplete or probabilistic input, input uncertainty, and stochastic training/prediction. It discusses how these uncertainties affect reliability guarantees and suggests principled approaches to improve trustworthy KGE.
We analyze network intervention with polling strategic agents and show that optimal prices admit a centrality-based decomposition of the welfare kernel, with each agent's contribution scaling with squared centrality in a reweighted network.
We study whether supervised and self supervised representations provide complementary signals for out of distribution graph learning. We propose two backbone agnostic frameworks to exploit SSL and supervised signals at different stages to improve OOD node classification.
Agentic discovery of blood biomarker from distilled private health records describes distilling private EHR evidence into a released GAT-based scoring tool that predicts candidate CBC biomarker performance, enabling external evaluation without sharing private data.
We introduce the hypergraph neural additive network, HGNAN, an interpretable model for hypergraph learning. It uses neural additive models to separate the contributions of node attributes and higher-order hyperedge structure, enabling better interpretability.
SkillGATE formulates skill retrieval as an adaptive information-foraging process. It uses gate aware Monte Carlo Tree Search to coordinate region-level navigation with skill-level selection under uncertainty, improving retrieval efficiency and accuracy.
We introduce INTEGER, a framework for adapting generative recommenders to multi-turn interaction. It extends generation based recommendations with a learned routing token that decides when to recommend, preserving the history mapping.
We propose a budget aware reasoning system for medical KGs with verifiable evidence. The system guides LLM reasoning with cost bounded exploration of external medical knowledge graphs, differentiates evidence by source quality, and yields citable rationales for retrospective verification.
This narrative review surveys uncertainty quantification frameworks tailored to connectome-based graph learning and includes an empirical case study illustrating how UQ improves reliability and interpretability in brain network diagnostics.
This work learns consistent molecular mechanics force fields from first principles by inferring bonded parameters as functions of local environments, improving accuracy and transferability of classical force fields.
We present Toward a Locally Deployable Agentic Co-Scientist, a framework where a compact language model plans calls to 18 modular tools for early stage drug discovery. A Unified Molecular Schema and LoRA fine tuning enable effective planning in a locally deployable setup.
We present scalable methods for discovering multiattribute logical dependencies and functional dependencies in tabular data, including higher order dependencies. The paper also provides a unified framework for extracting and visualizing these dependencies.
SciKGExtract is a schema-guided framework for materials knowledge extraction that combines LLM-based extraction, chemical normalization, and agent-based refinement before integrating into a knowledge graph. It is evaluated on 176 atomic-layer-deposition papers about ZnO and IGZO.
This survey discusses behavior mining, generative conversations, and collaborative advisory in travel recommender systems and outlines future directions toward context aware, multi stakeholder, and advisory rich systems.