All articles
37186 articles
FoldScape: Predicting Protein Structures at Any Folding Stage from Sequence
FoldScape is a sequence-conditioned deep learning framework that generates protein conformational ensembles at a specified folding fraction, covering a continuous path from collapsed coil to near-native molten globule. It complements AlphaFold 3 and supports long-chain inference and mechanical unfolding analysis.
BCSM: An Efficient New Objective Breaking Through the Bottleneck of Discrete Diffusion Distillation
BCSM proposes Bidirectional Centered Score Matching, reducing discrete diffusion distillation pairwise complexity from O(|V|²) to O(K), achieving 6× NFE reduction in single-step inference and 16.9% relative improvement on HumanEval.
Harness-S: Leveraging Sibling Harnesses for Generalizable Agent Training
Harness-S leverages the locality, complementarity, and nonstationarity of sibling harnesses via directional evolution, complementary rescue training, and adaptive routing to transfer harness capabilities to a single policy, significantly outperforming policy-only training on benchmarks like AppWorld.
RAG Retrieval Layer Stress Test: Query-Targeted Poisoning Can Cause nDCG to Plunge by 60%
A stress test on the RAG retrieval admission stage shows that just five query-targeted poisoned documents per query can reduce classic retrievers' nDCG@10 by 40.2%-61.5%, with BGE-small reaching a 100% target exposure rate. The study covers 623 BEIR queries and reveals that query anchoring is the main failure factor for dense models, while also testing the limits of defenses such as overlap filtering and MMR.
Reproducing the Chinchilla Scaling Laws: Feasible Paths and Key Challenges for Small-Scale Experiments
A discussion of small-scale reproduction experiments for DeepMind's Chinchilla scaling laws, analyzing feasible paths, experimental design essentials, and common pitfalls when reproducing compute-optimal training laws with PyTorch or JAX, offering practical reference for researchers.
US Tech Giants' Hidden AI Debt Surges 90% in Six Months to $3.2 Trillion
US tech giants' hidden AI-related debt has surged 90% in six months to $3.2 trillion. This article examines the sources of off-balance-sheet liabilities, the motives behind off-balancing, and their potential impact on financial transparency and the sustainability of AI investment.
Cycle WAM: Coupling Actions and Video via Cycle Consistency in World Models
Cycle WAM jointly trains video and action as causally paired modalities via bidirectional translation and a differentiable cycle reconstruction objective, achieving state-of-the-art on LIBERO-Pro and significantly boosting DreamZero on DROID.
DenseTwins: A New Self-Supervised Framework for Disentangling Multimodal Dense Features
An in-depth look at DenseTwins, a multimodal self-supervised framework that explicitly disentangles dense features into shared and modality-specific components, combining redundancy reduction with cross-modal alignment to reach SOTA or competitive results on multiple benchmarks. Code is open-sourced.
LENS: Reliability-Aware LLM Priors and News Semantic Graphs for Stock Prediction
LENS treats relationship reliability as a first-class signal: LLM multi-round extraction with confidence weighting on the static side, news semantic graphs with decay on the dynamic side, and gated fusion per stock per day. AR@5 +16.4%, SR@5 +22.6% on S&P 500.
Density-Aware Reward Scaling: Making Reinforcement Learning Sampling More Balanced
DARS (Density-Aware Reward Scaling) estimates local outcome density and reverse-scales rewards to promote reward-proportional sampling without modifying the policy gradient objective, improving reinforcement learning diversity. Experiments show it avoids mode collapse and improves text-to-image generation diversity.
OmniST: Topology-Invariant Spatiotemporal Forecasting with Coverage Guarantees for Unseen Sensors
OmniST proposes a topology-invariant spatiotemporal forecasting framework whose computation is independent of sensor count, combined with inductive conformal calibration for coverage-guaranteed predictions at unseen sites and fixes for static conformal temporal drift.
TAPES: Task-Aligned Joint Predictive Pre-Training for Transaction Sequences
TAPES proposes a task-aligned joint predictive pre-training framework that screens prediction targets by downstream utility and uses shared discrete latent variables to jointly model conditional co-occurrence among transaction sequence targets, outperforming baselines on real transaction risk scenarios and public financial benchmarks.
Should LLM Activation Steering Depend on Input? New Research Provides a Measurable Criterion
New research reveals input dependence in LLM activation steering: retrieval-based steering beats fixed directions when the target direction varies within a request category. Word-level edits suit shallow layers with few neighbors, clause-level edits need only the class mean, and four state quantities can determine the editing strategy in advance.
Replica-Aware Split Inference: Keeping Inference Running When Edge IoT Devices Fail
A paper proposes a replica-aware split inference recovery method that redeploys TinyML model segment replicas from failed devices to surviving ones, jointly optimizing replica hosts and communication protocols via MILP to minimize post-reconfiguration inference latency—6.8% lower than greedy baselines.
TactiNavi: Building a More Reliable LLM Game Coach with Explicit Strategy Conditions
TactiNavi is an LLM coach service deployed in a commercial open-world game. By making strategy preconditions explicit, it raises advice accuracy from 57.9% to 82.0% and reduces strategy misuse.
SAE PII Suppression Fails: Logprob-Only Evaluation Misleading
New study finds that evaluating SAE feature-level PII suppression solely by logprob drops is misleading: only 2.8% of 36 cases pass multi-metric tests, and random directions even outperform selected features. The authors call for causal mediation analysis and circuit-level localization.
ReSI-Math: Mining the Frontier of Model Recoverability to Boost Reasoning Efficiently
The ReSI-Math framework reveals that model failure does not mean incapability. Through self-judging-guided evolutionary search, it mines the recoverable capability frontier, turning failure cases into efficient RLVR supervision and significantly improving math reasoning scores on Qwen3-VL-8B.
KierstonBench: A Benchmark for LLM Criminal Investigation Capabilities
KierstonBench is a benchmark built on a city-scale Allentown simulation, evaluating LLMs' record linking, trajectory reconstruction, and deception detection across 32 task families; 16 models tested, most fail.
GC-MoE: Decoupling MoE Communication and Computation with CUDA Green Contexts
GC-MoE is an open-source FP8/FP4 MoE inference operator for Hopper and Blackwell GPUs that uses CUDA Green Contexts to split the device into separate communication and computation partitions, truly decoupling token dispatch and grouped GEMM under expert parallelism. It delivers consistent speedups on DeepSeek-V4.1-Flash and GLM-5.2 on H20 and B300, beating Mega MoE on every measured point.
LIRA: Making LoRA Dynamic Rank Allocation More Interpretable
LIRA proposes an interpretable dynamic rank allocation for LoRA, scoring each layer via second-order local loss perturbation and reallocating ranks with the DELTA algorithm under budget matching. It outperforms uniform LoRA and adaptive-rank baselines on RoBERTa-base GLUE; Trajectory EMA matches Fixed-point Fisher at lower cost.