Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 19 papers for 2026-09-17
Nothing cleared the bar today. Papers read in full today: 5.
The load-bearing idea is to separate repeated graph representations of one evidence atom from genuinely independent evidential origins, and to show that low revisit rates are not equivalent to adequate evidence coverage. The full text supports this diagnostic claim: all four frozen agents increase echo walks under redundant supportive paths, while effects on confidence and accuracy vary by model; moreover, PAPT cuts synthetic repeats and raises synthetic accuracy but leaves fewer sources unobserved and loses 3.6 points on SciFact. This is a worthwhile warning against reporting trajectory efficiency or provenance counting alone as evidence-use quality, rather than evidence for a generally superior graph-agent method.
The empirical case is constrained by a stylized equal-weight synthetic notion of source independence, a SciFact transfer setting with mostly one reachable document origin, and PAPT trained with only one seed; notably, simple provenance collapse attains perfect unique-path coverage and nearly matches PAPT's synthetic accuracy.
Continual Learning Mechanisms Compose for Long-Horizon Memorization investigates how language models can internalize information that arrives over time. The authors study a long-horizon setting with 100 sequential query–answer tasks trained via continual supervised fine-tuning, without replaying old data or using task identifiers at inference. They observe catastrophic forgetting and show that no single continual-learning mechanism preserves strong retention at this horizon, arguing that effective retention likely requires combining complementary approaches that address different sources of forgetting.
This work surveys AI for Games in the Foundation Model Era, focusing on how foundation models and learned game-world models reshape AI across the game lifecycle. Beyond playing, it discusses systems that model players and game dynamics, support design and development, adapt experiences at runtime, and evaluate artifacts, organizing literature into six roles based on the immediate use of AI output, and examining transferability across settings, engines, and populations.
StepAudio 3 Realtime introduces an audio-language foundation model built around a continuous listen-converse-think-act loop to support real-time spoken interaction. It uses Deep Perception to extract rich acoustic cues and Seamless Duplex to synchronize audio streams for natural pauses and interruptions; to combat latency, it employs Think-While-Speaking, running private reasoning in parallel with speech.
StepAudio 3 Music presents a large-scale, long-form music generation system with explicit planning and open-domain text control. The StepAudio Music Tokenizer encodes audio as a 50-Hz stream using a 65536-entry codebook, with self-supervised and multi-task training to preserve musical structure; a flow-matching diffusion Transformer (DiT) predicts continuous StepAudio VAE latents, decoded into 48-kHz audio. This discrete-continuous design is guided by comparisons of existing methods and aims to maintain musical coherence over long compositions.
Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement proposes Generalized Agent Iteration (GAI) as a formal framework that unifies iterative policy improvement and recursive self-improvement. Building on Generalized Policy Iteration (GPI), GAI addresses updates and evaluation that occur inside the agent, offering theoretical properties and implications for designing autonomous, self-improving systems.
We propose Procedural Pretraining, a three-stage pipeline that first trains on abstract procedural data, then pretrains on SMILES-based molecular data, and finally fine-tunes on downstream tasks. The procedural stage is designed to learn inductive biases from structure and rules via tasks spanning sequence structure, cellular automata, and related domains, improving data efficiency when labeled data is scarce. This leads to improved molecular property prediction, particularly under small labeled datasets.
We study stable filters for generative modeling of graph signals. We analyze how graph perturbations propagate through graph-aware Schrödinger-bridge dynamics and how the drift, formed by a graph filter plus a learned component, affects the generated distribution. We propose stable filter designs to ensure robustness to structural perturbations.
We study independence-system realisations in single-source unsplittable flow and introduce a path-closed notion of realising an independence system via zero-cost choices of primary terminals in a DAG flow instance. The definition is robust under standard composition operations. We extend the triangle mechanism and show that every finite loopless independence system is realizable.
We study chain-of-thought monitoring for LLM pricing agents under oligopolistic competition and develop a causal graph divergence framework to separately measure structural faithfulness and intent faithfulness. Across nine LLMs in duopoly and triopoly settings, we find collusive behavior and CoT faithfulness dissociate: the most collusive models report cooperative intent yet reason structurally unfaithfully.
We introduce ReDIL-GNN, a resynthesis domain incremental learning framework for circuit GNNs. Logic resynthesis changes gate vocab, topology, and statistics, causing domain shift without changing task labels. ReDIL-GNN adapts a fixed prediction or representation head as new synthesis styles arrive and assesses retention on prior domains, with a Resynthesis mechanism to guide adaptation.
Reasoning through Evolution introduces MAGER, a multi-agent framework for automatic meta-path discovery to support LLM-based fake news detection. It learns propagation patterns (meta-paths) that help reason about information spread and improves zero-shot/few-shot detection. The approach bridges propagation structure with LLM reasoning to enhance reliability.
Knowledge-Graph Based Augmentation versus Retrieval Augmented Generation compares Graph-RAG with standard RAG for cultural-related QA. It argues that structured KGs offer tighter control, explainability, and updatability over pretraining data alone. The study benchmarks both approaches on culturally relevant questions.
We study maximum strong independent set in finite hypergraphs: the largest vertex set that intersects every hyperedge in at most one vertex. This objective captures settings where local constraints suggest incompatibilities yet global transitivity is unjustified (e.g., deduplication with MinHash buckets). We develop reductions, bounds, and greedy certificates to tackle this problem and discuss practical implications.
FoundAna introduces a GNN-assisted foundation model for graph anomaly detection. It addresses the one-model-per-dataset limitation by enabling cross-task transfer under heterogeneous tasks, label scarcity, and domain shifts. The model leverages GNNs to provide transferable representations and supports robust, scalable graph anomaly detection.
We propose LIGE-GR, a smooth leap from ranking to generative recommendation in the LLM era. The framework aims to integrate sequence-level generation and optimization inspired by LLMs into industrial recommender systems. It addresses challenges in translating ranking objectives to generation and proposes solutions to align generation with ranking signals.
We study hyperbolic graph representation learning for Mendelian-disease differential diagnosis on a patient-integrated biomedical knowledge graph. Hyperbolic embeddings naturally capture tree-like hierarchies in ontologies and enable efficient representations when combining hierarchical and cross-cutting associations. Early experiments show promising improvements over Euclidean baselines in this differential diagnosis task.
Toward Composable Network Digital Twins proposes a subgraph-based latent approach to performance estimation. It decomposes networks into subgraphs to enable reusable, modular digital twins that adapt to topology and traffic changes. The goal is fast, what-if analysis with improved reusability over monolithic topology representations.
VyPER is a geometric learning framework that reconstructs collider events by representing events as hypergraphs with physics-informed topology. It solves two tasks: assigning measured jets and leptons to parent particles, and predicting unmeasured neutrino kinematics via graph-conditioned diffusion. VyPER achieves improved reconstruction accuracy and consistency with physics.
EffiRAG is a graph-based RAG system that prices structure costs to reduce retrieval-augmentation expenses. It uses the graph to locate relevant passages and then generates answers from the original text, preserving provenance while avoiding excessive LM calls. The work demonstrates cost savings and quality trade-offs.
We introduce a graph neural network surrogate to model disequilibrium chemistry in exoplanet atmospheres, addressing computational bottlenecks during atmospheric retrievals. The GNN surrogate leverages the topology of chemical reaction networks to accelerate predictions while maintaining accuracy. This work supports faster and more scalable exoplanet atmosphere analyses.
This thesis develops robust and efficient AI frameworks to accelerate crystalline materials discovery, addressing both property prediction and structure generation. It tackles high DFT cost and limited labeled data by leveraging graph representation learning, pretraining, multimodal learning, and generative modeling. The work demonstrates scalable approaches to predict properties and generate structures.
Wiki Foundation Model (WFM) advocates a dense, wiki-like knowledge representation for complex agentic reasoning. It emphasizes long-term memory and retrieval-augmented generation, moving away from sparse graphs toward Wiki-style dense knowledge representations. WFM enables more robust, scalable agentic workflows.
Quanta is an open-source Python library that unifies dense 4-bit quantised embeddings, BM25 lexical indexes, and knowledge graphs for hybrid retrieval. It provides a self-contained pipeline requiring minimal configuration for retrieval-augmented generation deployments. The library aims to reduce integration complexity across dense, lexical, and graph-based retrieval components.