Candidate-Set Diffusion Beats Masked Baselines on Logic Tasks
A new discrete-diffusion method represents uncertainty as sets of candidate tokens, improving reasoning and planning benchmarks.
Research Digest
Preprints and papers, rewritten in plain English.
Online conformal prediction for data streams: adaptation beats fixed thresholds
A benchmark of four sequential conformal regression methods shows that adapting both the model and the threshold improves coverage–width tradeoffs under drift.
Asking a Collaborator Improves Offline Reinforcement Learning
A new query-based framework lets offline reinforcement learning agents fill data gaps through targeted state-time queries, without risky online exploration.
Gaussian kernels on manifolds match optimal classification rates
A new approximation and oracle inequality shows logistic classification on manifolds can reach minimax-optimal excess risk without assuming a bounded Bayes function.
Bayesian machine learning for direct and spillover effects in cluster trials
A new Bayesian machine learning method uses a clever covariate to decompose direct and spillover effects in cluster-randomized trials with noncompliance.
Geometry-Aware Block Selection Tightens Sparse Gaussian Process Bounds
A new algorithm chooses data partitions that tighten the variational lower bound in sparse Gaussian processes, improving accuracy without increasing inducing points.
HRoPE: Homography-Based Positional Encoding for Multi-View Transformers
A new arXiv preprint introduces HRoPE, a geometric positional encoding that uses homographies to give multi-view transformers spatial inductive biases.
GenoMAS: Multi-Agent AI Improves Genomic Data Analysis
A guided multi-agent system outperforms prior AI tools on genomic benchmarks by blending structured workflows with adaptive planning.
Society-Scale Agent Simulation Reaches 500 Million Personas
A new framework pairs survey-grounded household personas with a compact decision model to test how credit and spending policies play out across age, income, and debt groups.
Near-Optimal Encoding Size for Universal Features
A new paper gives near-optimal bounds on the number of bits needed to encode data so that any function in a given family can be evaluated on it.
Spiking Neural Network Targets Online Learning on Intel Loihi 2
A co-designed spiking neural network aims to make edge AI adapt to new data efficiently on neuromorphic hardware.
Blocked Permutation Null Fixes False Influence Signals in Multi-Agent RL
A controlled MARL experiment shows standard permutation surrogates falsely detect influence between agents that never interact, while a block-permutation null restores valid false-positive rates.
Neuro-Symbolic Engine Tests Patch Presence Without Compilation
A new reasoning engine aims to make patch-presence testing more usable by checking for 1-day vulnerabilities in reused code without requiring compilation.
DP-DQN: Dirichlet Process Priors on Data, Not Weights, for RL
A new reinforcement learning algorithm places Bayesian nonparametric priors on the data-generating distribution, making epistemic uncertainty tractable with a single network.
Jacobian lens predicts LLM test failures before generation
A new arXiv paper proposes pre-generation adequacy measures, using the Jacobian lens, to predict and characterise LLM test failures and cut re-testing costs.
Conformal layer turns multi-agent debate into calibrated escalation decisions
A post-hoc conformal layer gives multi-agent LLM debates a finite-sample guarantee that the correct answer is in the candidate set, making wrong consensus actionable.
Sharp Spectral Norm Bounds and Regularization for Sparse Random Tensors
A new arXiv paper proves sharp spectral-norm concentration for sparse random tensors and gives a minimax-optimal regularization procedure.
InterWebCoder Trains MLLMs to Build Interactive Webpages
A new arXiv paper proposes using executable tests and reinforcement learning to teach multimodal language models to generate interactive webpages.