Research radio episodes

Listen to recent research-paper discussions with complete transcripts and sources.

Formalizing Forgetting: Theory of Forgetting for Online Continual Learning — AI Radio · 2026-10-18
The authors analyzed single-pass Shuffle SGD under strongly convex and general convex objectives to find explicit upper bounds on expected forgetting in online continual learning.
Explore This! Beat and Downbeat Tracking From a Learned Tatum Grid — AI Radio · 2026-10-18
The model predicts a tatum grid and subgrid to track music beats and downbeats by combining deep learning with theory-based modules. It uses a Transformer architecture to estimate the grid, which is then used to predict local metric distributions (subgrid).
Disentangling Marginal Capacity from Geometry in Optimal-Transport Label Transfer — AI Radio · 2026-10-18
The paper investigates how a class's limited transport capacity causes label transfer errors, even when classes are well-separated geometrically. It found that displaced mass flows to high-capacity classes instead of geometrically similar ones.
Contrastive Linearized Laplace: Loss Geometry, Valid Repairs, and Reliable Cores — AI Radio · 2026-10-18
The framework uses Linearized Laplace Approximation (LLA) to propagate encoder uncertainty through downstream tasks.
Representation Drift in Continual Learning: LoRA versus N-gram Memory — AI Radio · 2026-10-18
The study compares three methods for online learning: updating N-gram memory, training a shared LoRA adapter, and training a fresh LoRA per task. It found that N-gram memory keeps the base fixed while suffering high drift on old tasks due to shared rows.
Exact Phase-Type Point Processes for Certified Measurement and Prediction of Event Streams — AI Radio · 2026-10-18
The paper introduces a single algebraic object, a phase-type (Erlang-chain) state space, used for both certified measurement and prediction of event streams.
Random Feature Pruning for Multi-Class Classification — AI Radio · 2026-10-18
This work addresses redundancy in features extracted by Weighted sums of random kitchen sinks (W-RKS) for multi-class classification.
Toward a Multimodal Foundation Model for Plasma State Reconstruction — AI Radio · 2026-10-18
Researchers developed a single 5.5M-parameter 3D U-Net to reconstruct magnetic fields and electron density from incomplete plasma observations during magnetic reconnection simulations.
The Optimal Decoder is Tilted Temperature — AI Radio · 2026-10-18
The research explores optimizing Large Language Model decoding by finding a 'tilted temperature' distribution instead of a fixed temperature. It solves an optimization problem balancing output diversity (entropy floor) and plausibility (cross-entropy ceiling).
Sensor-Aware Joint Data Fusion and State Estimation under Sparse Multi-rate Distribution Grid Sensing — AI Radio · 2026-10-18
The work proposes a two-stage pipeline using a heterogeneous latent graph ODE and a graph-transformer state estimator to solve distribution system state estimation with sparse, multi-rate sensor data.
Swarm Telepathy: Prefill-Free KV-Cache Transfer Across Families — AI Radio · 2026-10-18
SwaT (Swarm Telepathy) is a method for multi-agent systems to communicate by transferring an agent's internal KV-cache directly to another agent, eliminating the need for prefilling text.
Physics Foundation Models for Efficient Control of Fluid Flows — AI Radio · 2026-10-18
Researchers repurposed a physics foundation model trained for predicting fluid dynamics (PDE forecasting) into a closed-loop control policy for fluid flow.
When Dynamic Scaling Fails: A Scale-State Feedback Governs Recurrent-State Quantization in Linear-Attention and SSM LLMs — AI Radio · 2026-10-18
This paper investigates numerical instability when using per-token dynamic quantization in recurrent neural networks.
LexiCan: Lexicographic Canonicalization of Point Clouds for Machine Learning Applications — AI Radio · 2026-10-18
LexiCan is a new method for labeling 3D point clouds that creates a provable canonical representative invariant to rotations (SO(3) or O(3)) and permutations.
FragFlex: Fragment-Based Flexible Molecular Generation for Exploring Vast Chemical Space — AI Radio · 2026-10-18
FragFlex is a flexible-length discrete diffusion model that generates molecules by inserting and deleting fragments during denoising. It jointly learns fragment count, identities, and connectivity while maintaining chemical validity.
Zeroth-Order Fine-Tuning of LLMs in Learnable Low-Rank Subspaces — AI Radio · 2026-10-18
AdaSub-ZO is a novel forward-only fine-tuning method for large language models that uses learnable low-rank projection matrices to guide gradient directions.
Explicit Bounds on Hoffman Constants for Sparse Polyhedral Regularizers — AI Radio · 2026-10-18
The research derives explicit bounds for Hoffman constants, which measure distance to a solution set in sparse regularization problems like LASSO.
Correlation-Aware Reconstruction for Multivariate Time Series Anomaly Detection — AI Radio · 2026-10-18
CARMA proposes a reconstruction-based autoencoder using two latent spaces: a temporal one for sequence dynamics and a correlation one for inter-channel relationships.
ANTE: Anticipate, Then Execute -- Learning Internal World Models for Coding Agents from Verified Beliefs — AI Radio · 2026-10-18
ANTE is an agentic Reinforcement Learning framework that trains coding agents to build internal world models. It formalizes knowledge as a belief state containing what has happened, what is true now, and what actions will achieve.
DC-Flow: Target-Conditioned Cyclic Peptide Design via Delayed-Coupling Flow Matching — AI Radio · 2026-10-18
DC-FLOW is a generative framework for designing cyclic peptides by decoupling sequence and structure generation using Delayed-Coupling Flow Matching. It generates a backbone first, then sequence, ensuring structural context informs chemical assignments.
LiveAnon: Improving Privacy and Utility in Streaming Speaker Anonymization with Codec Language Models — AI Radio · 2026-10-18
LiveAnon is a voice anonymization framework that uses a streaming Automatic Speech Recognition (ASR) encoder to extract content features from source speech.
Anchors Enhance Controllability and Reasoning of Diffusion Language Models — AI Radio · 2026-10-18
AnchorDLM is a framework that trains an anchor generator to create short output fragments and their positions for each query. These anchors act as control tokens placed in a frozen Diffusion Language Model to guide its denoising process.
GA-SVD: Generation-aware Singular Value Decomposition for LLM Compression — AI Radio · 2026-10-18
GA-SVD is a low-rank compression method for large language models that improves performance over existing SVD methods by considering how the model generates text. It uses generation-aware whitening statistics derived from the model's own outputs to guide the compression process.
RELAGRAM: Recursive Latent Graph Model for the Capacitated Vehicle Routing Problem — AI Radio · 2026-10-18
RELAGRAM introduces a Recursive Latent Graph Model for solving complex vehicle routing problems by treating routes as edge tokens. It uses a compact transformer that updates latent representations of these tokens recursively, maintaining two states to guide the search.
Beyond Adjacent Layers: Graph-Guided Layer Fusion for Compressing Large Language Models — AI Radio · 2026-10-18
FCM identifies non-adjacent functional communities within large language models by using full pairwise Centered Kernel Alignment similarity matrices.
DeltaNet-PC: Test-Time Memory Without Backpropagation Through Time — AI Radio · 2026-10-18
DeltaNet-PC develops a predictive coding system that learns to memorize sequential information during testing without needing backpropagation through time.
ObsSpec: Accelerating Agentic Reinforcement Learning with Observation Speculation — AI Radio · 2026-10-18
OBSSPEC introduces an observation speculation pipeline for agentic RL that uses a small world model to predict tool outputs while the environment runs. This allows the policy to start generating its next step early, reducing agent latency.
Approximation in Contrastive Representation Learning — AI Radio · 2026-10-18
The research investigates how well standard inner-product scoring functions approximate Contrastive Representation Learning's true objective, Pointwise Mutual Information (PMI). The study proves that while finite dimensions introduce error, this error can be controlled.
Riemannian Graph Tokenizer for Structural Knowledge Transfer — AI Radio · 2026-10-18
The Riemannian Graph Tokenizer (RGT) uses Riemannian geometry to bridge structural knowledge and quantized representations for cross-domain transfer in graph data.
Beyond Aggregate Drift: Optimal Client-Specific Windows in Federated Learning — AI Radio · 2026-10-18
The research addresses how much historical data (memory) each client in federated learning should keep when data distributions drift differently. It proposes giving each client a personalized memory tailored to its specific noise, arrival rate, and drift speed.
The Price of Precision: A Calibrated Cost Model for Mixed-Precision LLM Serving — AI Radio · 2026-10-18
A cost model was developed to predict latency and guide mixed-precision allocation for LLM serving by accounting for binding resources like compute, weight-loading, and KV-cache costs.
RF Fingerprinting of DSP Implementations using Machine Learning: An Over-the-Air Case Study — AI Radio · 2026-10-18
The study investigated using machine learning to perform RF fingerprinting based on Digital Signal Processing (DSP) implementations of RTTY communication software.
Sensitivity Fidelity Reveals Physical Dependencies Beyond Forecast Accuracy in Deep-Learning Earth-System Models — AI Radio · 2026-10-18
The study introduces sensitivity fidelity as a measure to check if deep learning models learn correct physical rules, not just accurate predictions.
UNISON: A Unified Framework for Learning from Scored Bags — AI Radio · 2026-10-18
UNISON creates a unified framework for learning from scored bags by treating multiple instance learning and recommendation systems as instances of learning from scored item-bag pairs.
Testing Central Structure for Complex Networks — AI Radio · 2026-10-18
The work introduces a network-level hypothesis test using the adjacency spectral radius to determine if a complex network has a statistically identifiable central structure before ranking individual nodes.
DRanker: A Transformer-Based Multi-Task Ranking Model for the Disney+ Homepage — AI Radio · 2026-10-18
DRanker is a transformer-based ranking model designed to personalize content ordering for a streaming service homepage. It uses multi-task learning to predict both user engagement probability and expected engagement length simultaneously.
Adaptive Role Assignment and Controlled Feature Reconstruction for Federated Image Semantic Communication — AI Radio · 2026-10-18
The ARC-FedSC framework coordinates user role assignment and server-side feature reconstruction under a fixed communication budget for hybrid federated semantic communication.
How Shared Representations Become Specialized: Formation, Maintenance, and Latent Susceptibility — AI Radio · 2026-10-18
The study investigates how shared internal representations become specialized during minimal artificial learning by examining formation, maintenance, and latent susceptibility.
Context-Aware Precision Scheduling for Small Language Models — AI Radio · 2026-10-18
The work introduces a method to manage accuracy and speed for small language models as they process longer text contexts. It develops a calibrated predictor to estimate mixed-precision latency based on execution costs, not just weight size.
Grounding Deep Learning Framework Testing in Data Flow — AI Radio · 2026-10-18
FlowMuT is a mutation-based testing technique for deep learning frameworks that uses a Tensor Flow Graph (TFG) to track data flow and operator requirements during execution.
Benchmarking Deep Learning Models for Anomaly Detection in Real-World Cloud Telemetry — AI Radio · 2026-10-18
The study tested four deep learning models—LSTM/GRU, Attention Transformer, TimesNet, and Autoformer—on IBM Cloud data using the Numenta Anomaly Benchmark (NAB). Results show that short-term deviations dominate real cloud anomalies.
Lightweight Scheduling Policies for Continuously Batched LLM Serving Across GPU Architectures — AI Radio · 2026-10-18
The research investigated how simple scheduling policies can improve latency and throughput in continuously batched LLM serving without changing the model itself.
Entropy-Regularized Mean-Field Games with Average-Reward Criterion: Span-Seminorm Contractions and Model-Free Learning — AI Radio · 2026-10-18
The paper investigates entropy-regularized mean-field games under an average-reward criterion to find a unique equilibrium. It introduces a soft Q-operator that acts as a span-seminorm contraction, allowing relative value iteration to solve the optimality equation.
Split Conformal Interval Length under Huber Contamination — AI Radio · 2026-10-18
This research investigates how contamination during training and calibration affects split conformal prediction interval lengths under Huber contamination.
Revisited Support Vector Machines: A Unified Joint Optimization Framework with Nystr"om Approximation for Scalable Adaptive Learning — AI Radio · 2026-10-18
This work introduces Revisited SVM, a new framework that unifies optimizing dual multipliers and kernel parameters simultaneously. It tackles the cubic complexity of traditional SVMs by using Nyström approximation to reduce computational cost from O(N 3) to O(Nm).
ProML: Progressive Multi-label Learning — AI Radio · 2026-10-18
ProML introduces Progressive Multi-Label Learning to solve optimization challenges in multi-label tasks by decoupling label fitting from predictor learning using an auxiliary variable P.
MERIT: Scaling KG Foundation Models with Relation Transformers — AI Radio · 2026-10-18
MERIT introduces a framework using a transformer over relation names instead of message-passing on relation graphs.
Drift, Then Spread: Learning One-Step Generative Policies for Online Reinforcement Learning — AI Radio · 2026-10-18
DRIFTSPREAD proposes a one-step generative policy for online reinforcement learning by combining critic-guided Drift and entropy-induced Spread updates derived from a Maximum Entropy objective.
A Dependency-Aware Multi-GPU Training Framework for Temporal Graph Neural Networks — AI Radio · 2026-10-18
DepBloom accelerates multi-GPU Temporal Graph Neural Network (TGNN) training by integrating three techniques: temporally bounded spatial partitioning, intra-batch memory refinement, and critical-path optimization.
Guarded Block Transfers for Polytope Distance: Linear Convergence with Explicit Constants and Gap-Certified Screening for SVM Training — AI Radio · 2026-10-18
The paper introduces guarded block transfers to improve the Triangle Algorithm for solving polytope distance problems. This method aggregates multiple Mitchell–Demyanov–Malozemov transfers into one search, guarded by a comparison with the best single transfer.
A Hybrid Gaussian Process-Extreme Gradient Boosting Machine Learning Framework for Robust Knockdown Factor Prediction and Reliability-Based Design of Thin Spherical Shell Caps — AI Radio · 2026-10-18
The study developed a hybrid Gaussian Process–Extreme Gradient Boosting Machine Learning model to predict knockdown factors (KDF) for thin spherical shell caps, addressing conservatism in current design methods.
Two Linear Programs are Better than One — AI Radio · 2026-10-18
The paper investigates whether solving two linear programs is better than solving a single one, focusing on computational complexity and representation.
Mesh-Free Convolution: Learned Spectral Attenuation and Transport — AI Radio · 2026-10-18
The research proposes Mesh-Free Convolutions (MFCs), continuous filters that learn scale-dependent smoothing and directional transport terms, unlike standard Fourier Neural Operators (FNOs) which use fixed spectral limits.
Using Line Segment Tracking to enable ML-based Track Reconstruction in CMS — AI Radio · 2026-10-18
The research introduced a track-reconstruction model using Transformers by treating Line Segment Tracking (LST) objects as tokens. LST uses features from silicon sensors to build increasingly complex objects, culminating in Triplets.
Fast and Accurate: Improving Personalized LLM Performance by Semantic Concentration Inspired Workload Split between RAG and Fine-Tuning — AI Radio · 2026-10-18
The proposed PEFT Union RAG (PUR) approach splits user data based on semantic concentration. It uses parameter-efficient fine-tuning for concentrated core knowledge and standard retrieval augmentation for dispersed data.
Error-Based Bidirectional Predictive Coding Enables Deep, 4-Bit Predictive Coding Networks — AI Radio · 2026-10-18
The paper introduces ePC+bPC, combining bidirectional predictive coding and error-based reparameterization to create deep predictive coding networks that work well at 4-bit precision.
Harmonic Representations in Symbolic Music Transformers are Localized and Controllable — AI Radio · 2026-10-18
The study investigated how symbolic music transformers understand harmonic structure by analyzing their internal representations during cadences. It found that harmonic knowledge develops progressively from middle layers, with mode understanding preceding key identity.
Learning, Compression, and Meta-Complexity — AI Radio · 2026-10-18
The research formalizes learning and data compression using Kolmogorov complexity to measure inherent information content. It connects these fields by showing how finding a concise explanation for data relates to constructing models.
XForecast: The KDD 2026 Finance Day Challenge on News-Grounded Financial Forecasting — AI Radio · 2026-10-18
XForecast creates a benchmark and competition for predicting stock price movements four weeks ahead by integrating historical stock data with curated financial news signals.
Fast Quantized Johnson - Lindenstrauss Transforms — AI Radio · 2026-10-18
The paper introduces two fast methods, D-FQJL and P-FQJL, for asymmetric quantized Johnson–Lindenstrauss transforms. These constructions replace dense Gaussian projections with fast structured transforms based on Randomized Hadamard Transforms.
DeepFlow: Two Convergent Differentiable Approaches for Pairwise Conditional Random Fields — AI Radio · 2026-10-18
DeepFlow introduces two methods for integrating Pairwise Conditional Random Fields (CRFs) with deep learning, ensuring they are both convergent and differentiable.
LLM-HB: Language-Aware LLM-Guided Hotword Biasing for Code-Switching ASR — AI Radio · 2026-10-18
The LLM-HB framework enhances Automatic Speech Recognition (ASR) for code-switching speech by combining a Mixture-of-Experts (MoE) adaptor, an auxiliary language prediction head, and LLM-based hotword biasing.
Learning Deployable Causal Action Geometry under Temporal Non-Stationarity — AI Radio · 2026-10-18
The paper addresses temporal non-stationarity where outcome levels drift over time despite stable causal effects. It proposes a 'deployable causal action geometry'—a stable action-response shape—as the decision target instead of raw predictions.
From a Large-Scale Teacher to Efficient Multimodal Students for Search Relevance Prediction — AI Radio · 2026-10-18
The research addresses slow and expensive relevance prediction using large LLMs by distilling their performance into smaller, efficient multimodal student models. The method uses a high-quality human dataset to train an LLM-as-a-judge system as a teacher.
Prorsa: Unleashing the Potential of Prefill-Decode Disaggregation in Agentic Reinforcement Learning — AI Radio · 2026-10-18
Prorsa introduces an agentic RL framework that uses prediction-guided prefill-decode resource allocation to speed up rollout by 1.11x to 4.05x over traditional methods.
NFMap: Topographic 2D Manifold Learning with Invertible Normalizing Flows — AI Radio · 2026-10-18
NFMAP is an invertible method that combines Generative Topographic Mapping's fixed grid prior with a normalizing flow to map data into a 2D space. It creates a navigable map and an off-manifold residual, ensuring full invertibility.
IND-Eval: Aligning Evaluation and Simulation for Industrial Time Series Foundation Models — AI Radio · 2026-10-18
IND-Eval created a public benchmark for industrial time series forecasting, organizing data into source-specific tasks.
RFWeave: Learning Wireless Propagation Across Scenes and Transceiver Configurations — AI Radio · 2026-10-18
RFWeave is a learned wireless propagation model designed to generalize across different scenes, transceiver locations, and antenna configurations. It models the receive-side response using spatial covariance and received power instead of a full channel.
Small area learning — AI Radio · 2026-10-18
Small Area Learning (SAL) reframes Small Area Estimation as a multi-target semi-supervised learning problem.
Instance-Adaptive Online Multiclass Multicalibration — AI Radio · 2026-10-18
The algorithm adapts its prediction strategy to how the sequence of outcome distributions is actually structured, specifically its affine subspace dimension 's'. This adaptation allows it to achieve significantly better error rates than worst-case bounds on easier instances.
Probabilistic Forecasting of Wildfire Size Using Conditional Flow Matching — AI Radio · 2026-10-18
This work proposes a probabilistic model using conditional flow matching to predict wildfire size over time, focusing on acreage rather than spatial progression.
BARTER: Compact Dynamic Sparse Training with Standard Dense Kernels — AI Radio · 2026-10-18
BARTER is a new Dynamic Sparse Training framework that reduces training costs by keeping model tensors physically compact using standard dense kernels instead of specialized sparse ones.
Sampling from Neural Operator Priors to Solve Partial Differential Equations — AI Radio · 2026-10-18
The authors combined pseudo-bootstrap sequential Monte Carlo (pBS) sampling with a U-shaped Neural Operator (UNO) diffusion prior to solve partial differential equations in function space.
Neural Network Preconditioning: A Case Study for the Solution of the Parametric Helmholtz Equation — AI Radio · 2026-10-18
A U-Net based convolutional neural network is trained self-supervised to approximate the inverse of a discretized Helmholtz operator.
Optimal Transport PAC-Bayes Bounds — AI Radio · 2026-10-18
This work develops new generalization bounds by replacing classical Kullback-Leibler divergence with general optimal transport distances.
HyBRA-EEG: Hyperbolic Global-Local Brain-Region Aggregation for Self-Supervised EEG Representation Learning — AI Radio · 2026-10-18
HyBRA maps messy EEG data into a structured regional token space using hyperbolic geometry and channel evidence. It combines electrode locations, region priors, and feature-dependent distances to create superior self-supervised representations for EEG learning.
Efficient Resource-Shiftable RL Post-training with Generative Rewards — AI Radio · 2026-10-18
PivotRL is an RL post-training system that dynamically moves computational resources between different roles based on real-time workload demands to maximize throughput.
Counting with Attention and LayerNorm: Norm Scaling and Representation Geometry — AI Radio · 2026-10-18
The study investigates how placing Layer Normalization (LN) affects parameter efficiency when counting tokens within a single attention layer using softmax attention.
DASS: A Solver-Adaptable Dynamic Auxiliary Search Strategy for Symbolic Regression — AI Radio · 2026-10-18
DASS introduces a dynamic auxiliary search strategy for symbolic regression to navigate complex mathematical expression spaces efficiently.
Efficient Preconditioned Optimization Methods with Frequent Directions — AI Radio · 2026-10-18
The authors propose efficient preconditioned optimization methods like FTSL and S-ASGO that accelerate adaptive optimization using frequent directions (FD).
TILE: Bridging LLMs and Table Foundation Models for Tabular Learning with Strings — AI Radio · 2026-10-18
TILE is a framework that connects Large Language Models (LLMs) to Table Foundation Models for tabular learning, especially when tables have string columns.
From Parallel to Sequential: Small-Batch Noise Restores Simplicity Bias of Muon — AI Radio · 2026-10-18
The study investigates how adding small-batch noise fundamentally changes Muon's optimization behavior, which usually learns features in parallel.
Learning to Design Convex Relaxations for Nonconvex Optimization — AI Radio · 2026-10-18
AutoConvexRelax learns how to design convex relaxations for hard nonconvex optimization problems. It models the problem as a graph and trains a policy to select symbolic rewrites sequentially.
Masked Diffusion for Travel Diaries Synthesis — AI Radio · 2026-10-18
DiaryBridge is a masked diffusion model that processes travel diaries as single fixed-length sequences of tokens, including explicit count tokens for vehicles, persons, and trips. It learns to restore hidden diary entries by unmasking positions in any order.
Stabilizer entropy cones as self-dual slices of linear rank cones: a five-party gap and dependence on the local dimension — Quantum Radio · 2026-10-18
The research connects entanglement entropies of stabilizer states to classical combinatorial objects via a linear lift. It finds that the resulting cone is a self-dual slice of a classical rank cone.
HarnessTTT: Learn to Optimize Harness at Test-Time for Scientific Discovery — AI Radio · 2026-10-18
HarnessTTT introduces a test-time learning framework that trains a harness proposer while keeping the executor frozen. The system learns to generate executable harnesses, which guide a discovery rollout.
ModRes: A modification of the ResPlan dataset ready for floor plan analysis and graph machine learning — AI Radio · 2026-10-18
ModRes is a cleaned and enriched version of ResPlan, containing 8,431 unique residential floor plans with structured geometric, topological, and semantic data. The process involved rigorous cleaning to remove invalid or duplicate plans.
A Geometrical Approach to Constructing Neural Networks for Perfect Training-set Classification — AI Radio · 2026-10-18
The method constructs a neural network to perfectly classify training data without regularization by using a geometrical approach. It builds two hidden layers based on recursively partitioning the data with separating hyperplanes, mapping these splits directly to neurons.
Popularity Trajectories and Mixing Patterns of Urban Venues in Two European Cities — AI Radio · 2026-10-18
The study analyzed Foursquare data from London and Barcelona to find six recurring popularity shapes for urban venues: Fall, Rise, Rise & Fall, Steady, Ephemeral, and Underdogs.
Monarch-Muon: Structure-Aware Muon for Fast and Memory-Efficient Pretraining — AI Radio · 2026-10-18
Monarch-Muon is a structured matrix-aware optimizer designed to make large language model training faster and more memory-efficient.
Formal-Verification-Guided Bayesian Optimization for Non-Intrusive Last-Mile Coverage Closure — AI Radio · 2026-10-18
This workflow uses a non-intrusive diagnosis-and-search approach to close coverage gaps in hardware verification environments.
Longitudinal Learner Participation Profiling in Online Learning: Heterogeneous Consensus Clustering under Unequal Observation Windows — AI Radio · 2026-10-18
The study investigated how to profile learner participation over time using Moodle data while accounting for unequal observation windows caused by varying course lengths.
Calibrated alpha-Barycenters for Multiobjective f-Divergence Optimization — AI Radio · 2026-10-18
CAL-alpha-MODM addresses multiobjective optimization by aggregating a nonlinear predictor into a single barycenter and calibrating its residuals against an f-divergence objective. This allows for robust parameter error control without requiring targets to be close to the model.
FORGE: Procedural Weight Generation for Sparse MoE Inference — AI Radio · 2026-10-18
FORGE replaces storing expert weights in high-bandwidth memory with generating them procedurally during inference. This system uses a Rotor Generation Unit to create FFN weight tiles on-the-fly from rotor parameters instead of loading dense tensors.
Distinguishing Process Change, Instrument Fault and Data-Path Corruption in Streaming Telemetry: A Reference Implementation and Benchmark — AI Radio · 2026-10-18
The paper tackles the 'three-cause problem' where a reading deviates from expectation due to process change, instrument failure, or data corruption.
LocalTrain: Integer-Only Local Learning for Energy-Efficient CNN Training on Edge FPGAs — AI Radio · 2026-10-18
The KV260 prototype demonstrates complete integer local CNN training using a shared matrix engine for both convolutional and classifier arithmetic.
OpenDGO: Benchmarking Multi-Agent Organizations Across Diverse Application Domains — AI Radio · 2026-10-18
The OpenDGO framework systematically benchmarks how multi-agent systems should be organized across six domains and three collaboration settings.
CodeLift: Your Test Suite Is Secretly a Dense SWE Curriculum — AI Radio · 2026-10-18
The standard binary reward for training software agents is too sparse, failing to distinguish between partially correct patches. The test pass-rate is better but also unreliable due to imbalances like multiplicity and progress-regression.
Robustly Reliable Learning for Regression — AI Radio · 2026-10-18
The work introduces a method for robustly reliable learning in regression, aiming to find an optimal prediction range for a test point despite corrupted training data. It achieves this by determining the tightest possible range consistent with all but at most b points.
Machine Unlearning for Federated Socially Fair k-Means Clustering — AI Radio · 2026-10-18
The research develops exact machine unlearning for federated socially fair k-means clustering to handle data deletion requests while maintaining fairness. It compares exact replay against retraining, finding that whole-client deletion is 35.8 times faster than retraining.