§ Research

John Platt (Google) discusses AI's role in automating scientific discovery, climate applications, and future participation in superintelligent research ecosystems.

Flash-dLLM optimizes KV caching and parallel decoding for diffusion LLMs via IO-aware techniques, addressing inference bottlenecks in non-autoregressive text generation.

Framework for decentralized multi-agent decision-making under partial observability with delayed information sharing using low-rank model learning.

SpeakerMem-R1 introduces speaker-centered dual-track memory for multi-party dialogue, addressing person/group attribution and temporal state tracking.

CliffCompaction reduces token compression costs by 50% for long-horizon coding agents while maintaining performance on KernelBench and Terminal-Bench.

Growing Harness learns reusable executable agent scaffolds from task feedback, reducing redundant LLM inference on repeated control decisions.

Study showing typed decision models may misinterpret option semantics despite schema conformance, demonstrating gap between syntax and intended semantics.

A-DLCC proposes parameter-free clustering via β-integrated local depth; orthogonal to LLM/AI frontiers.

DISCO applies diffusion-based spatial attention to graph community detection; limited relevance to core AI systems.

RCT with 100 product professionals shows Figma Make's prompt-to-design tools reduce design task time; empirical productivity evidence.

LYRA identifies Proximity Trap: irrelevant nearby context overshadows distant evidence in long-context LLMs; proposes t-distributed relevance alignment.

On-policy distillation (OPD) fixes quantization exposure bias in sub-3-bit LLMs; restores math and code reasoning in low-bit models.

Framework for optimal sequential annotation budgets in off-policy evaluation when LLM-as-judge labels carry unknown bias.

Theoretical physics study on learning-theoretic conditions for bosonic Gaussian states; unrelated to AI systems or LLMs.

Method for steering LLM reasoning via semantic exploration of problem-specific strategies rather than naive repeated sampling.

Spectral theory explains grokking as transition from NTK regime to feature learning via weight decay-induced kernel evolution.

MAGIC: reinforcement learning generates task-specific multi-agent collaboration topologies with mixed granularity to reduce cost and improve performance.

Discovery-Driven Integration uses unstructured text to identify missing relational structure for joining semantically related tables in data lakes.

Parametric convergence rates for entropic optimal transport in semi-discrete regime with subGaussian measures and no dimension dependence.

GravityOCR uses diffusion-based parallel decoding with AR verification to accelerate document OCR inference beyond sequential token generation.

Method extracts hidden chain-of-thought reasoning from closed-source frontier models including GPT-6 Astra via API tool registration to validate reasoning quality.

Knowledge Pull Requests framework automates incremental document updates by extracting claims, routing to sections, and flagging content conflicts with change tracking.

PersonaWeaver controls LLM-generated character diversity in procedural generation by mitigating behavioral homogeneity through structured persona attributes.

Information-theoretic criteria for selecting basis functions in sparse Gaussian process regression to optimize M-budget allocation across candidate bases.

MMAP pretraining framework handles missing multimodal data and incomplete records for longitudinal Alzheimer's disease progression prediction.

Framework identifies semantic abstraction gaps in LLM natural language inference via constructed higher-order semantic knowledge and reasoning diagnostics.

Foundation model embeddings from screening mammograms encode pre-diagnostic tissue changes detectable before cancer diagnosis, varying by pretraining domain.

Hierarchical GNNs with latent communication improve power-flow model generalization across grid topologies using Kron-derived graph reductions.

Mixture-of-experts with diffusion models for cross-scenario physical layer security in 6G wireless networks.

GTR: softmax-free recurrent vision backbone using gated linear attention and spatial scans for efficient dense prediction, distilled from DINOv3.

Polyak-type step-size selection for extragradient methods solving monotone root-finding problems with unified deterministic analysis.

Quantum-aided active device detection for energy-harvesting symbiotic radio networks using NOMA in IoT deployments.

Neutral-atom quantum computing applied to NP-hard resource allocation problems (MAP) in NOMA wireless networks.

UFO-MGen: flow-based generative model for crystal structure design learning Wyckoff topological representations for materials discovery.

Epistemological assessment of LLMs used in infant syntax acquisition research, examining dataset construction and model implementation in the BabyLM challenge.

Mathematical foundations of Fourier-Bessel wavelets derived from disk harmonics and Helmholtz equation solutions.

Joint recognition and translation fine-tuning via group relative policy optimization (GRPO) for LLM-based speech translation with Qwen2.5-Omni-3B.

Semiotic framework evaluates NLG fidelity and coverage by measuring contextual meaning and discourse references between texts.

Comparison of CNNs and vision transformers via EEG encoding models across network depth, testing token representations.

Analysis of Oja's algorithm for gap-free streaming PCA with near-optimal convergence rates and differential privacy applications.

Feature-wise linear modulation (FiLM) integrates radiomics with RenalCLIP foundation model for CT-based renal cell carcinoma classification.

OMatG-flash uses flow models with reinforce adjoint matching to accelerate inorganic crystal structure prediction and generation for materials discovery.

Framework integrates hierarchical cognitive process modeling with process supervision for interpretable scene safety assessment in critical domains.

Study reveals diffusion models exhibit malign overfitting and catastrophic memorization under overparameterization, contrary to benign overfitting in standard deep learning.

TimeInteract extends time-series language models to continuously process streaming observations and user intent with autonomous response decisions.

Face/Off evaluation framework reveals LLMs overweight lexical cues in code comprehension tasks despite access to semantic program structure.

First real-world validation of Learning to Defer on medical imaging datasets, enabling selective routing between AI and human radiologist decisions.

Lightweight layout-aware masking improves GROBID's structured PDF-to-text extraction for scholarly documents using CPU-only detection.

MAVP framework improves mobile manipulation via map-aware visuomotor policies with explicit base-pose prediction and localization feedback.

FairMean algorithm balances fairness and robustness in distributed learning by bounding gradient amplification under label poisoning attacks.