ProbePO: Sparse probes for credit assignment in LLM RL
A new preprint proposes a critic-free method that redistributes reward credit to individual reasoning steps, addressing a blind spot in verifiable-reward reinforcement learning.
Research Digest
Preprints and papers, rewritten in plain English.
Making Direct Preference Optimization Robust to Label-Flip Attacks
A new arXiv paper addresses the vulnerability of Direct Preference Optimization to adversarial preference label flips.
Near-Optimal Regret for Semi-Bandit Learning of Monotone Problems
New results show that semi-bandit learning can achieve near-optimal regret for monotone stochastic optimization problems with unknown distributions.
Contrastive Adversarial Training Aims to Harden GNNs Against Label Poisoning
A new arXiv preprint proposes a contrastive adversarial training approach to stop a few corrupted labels from cascading through graph neural networks.
Benchmark Puts LLM Kernel Generation to the Test on Edge Hardware
A new benchmark, EdgeKernelBench, evaluates how well LLMs generate compute kernels for the heterogeneous and resource-limited edge computing landscape.
Study Tests Whether a Triage Model Can Flag Insufficient Information
A new arXiv paper evaluates Jev, a decision model that outputs probabilities over predefined actions, for calibration and decision accuracy in patient-facing triage.
Autonomous Scientist for Traditional Chinese Medicine
A proposed AI agent for TCM research is designed to respect inferential boundaries across texts, assays, predictions, and clinical studies.
Activation Monitors: Do They Need Original Model Computations?
A new arXiv preprint asks whether safety monitors that inspect a model's internal activations gain an advantage from using the original model's computations.
Tensor Network Datamodels Capture Non-Additive Training Data Effects
A new arXiv paper proposes tensor network datamodels to overcome the fixed-contribution limitation of linear training data attribution methods.
New Benchmark fxBench Targets Spreadsheet Formula Suggestions
A new evaluation framework aims to measure and understand proactive formula suggestions in spreadsheets.
AdamPM Targets Memory Overhead of Adam for LLM Training
A new arXiv paper proposes moment pooling to reduce the substantial memory cost of Adam when training large language models.
Fisher-Geometric Pruning Tames Bloated Symbolic Regression Expressions
A new arXiv preprint proposes using Fisher information geometry to prune overparameterised genetic-programming symbolic regression models.
Selective Learning Improves Spacecraft Attitude Capture Under Torque Limits
A new arXiv study proposes paired-outcome learning to selectively modify maneuvers when braking authority is uncertain and torque is limited.
Whitening Gradients by Covariance: A Statistical View of Optimizers
A new arXiv paper argues that stochastic gradients should be whitened using their covariance, offering a statistical lens on optimizers like Adam, Shampoo, and Muon.
Augmenting Event Logs for Process Prediction Under Label Scarcity
The paper addresses a key bottleneck in predictive process monitoring by proposing a process-aware data augmentation strategy for event logs.
Recurrent Temporal RBMs as Unsupervised Encoders for Movement Data
A new arXiv paper proposes recurrent temporal restricted Boltzmann machines to learn representations of biomechanical movement without manual labels.
FlowMaker Aims to Cut Costs of Generative AI Workflows
A new framework addresses the cost of executing multi-step LLM workflows by optimising deployment decisions.
Pair-Aware Optimization Targets Jailbreaking of LLM Agents
A new arXiv preprint proposes a coordinated exploration-and-refinement method for generating adversarial inputs that jailbreak LLM agents.