The Archive

Brain-to-text decoding improvements from d'Ascoli et al. are largely reproducible via timing shortcuts, not actual neural signals.

ViTeX-Bench introduces benchmark for video scene text editing with temporal consistency and locality constraints.

VideoMSN applies masked Siamese networks with image ViTs to self-supervised video representation learning via frame masking.

EvoDuet co-evolves LLM solutions and web search queries with retrieval gating for scientific discovery tasks.

Weight tying in decoder-only LLMs (GPT-2, DistilGPT-2) under DP-SGD privacy training shows diminished benefits vs. non-private settings.

Turbo Harness adapts globally-optimized agent harnesses to individual task instances by recycling optimization artifacts.

WorldAuditBench benchmark for detecting anomalies in 3D environments using multimodal agents combining vision and navigation.

Cogentic multi-agent orchestrator for automated theorem proving uses prove--verify loops with prover population allocation.

MatLoom uses pretrained LMs to generate layered material programs with explicit PBR channel dependencies for compact text-to-material synthesis.

Loop Scaling Laws jointly model recurrence and sparsity to characterize effective-parameter gains in looped MoE transformers.

Compression footprints detect model poisoning in federated learning by exposing distortion patterns that differ between honest and malicious updates.

DynaHarness couples semantic reasoning with physical execution via shared contracts to enable robot agents to revise failed capabilities iteratively.

Looped Diffusion Transformer scales computation by reusing Transformer blocks within denoising steps to improve text-to-image generation without larger models.

Study on autonomous ML engineering agents finds primitive LLM coding agents with direct environment access outperform complex multi-agent orchestrators.

Empirical study: 27.5-31.1% of 2026 web tokens are AI-generated; pretraining 800 LMs measures impact on scaling laws and model collapse.

Green's operator analyzes perturbation routing in recurrent networks to study task-to-task and time-to-time computation reuse.

CAS II develops algorithmic statistics over symmetric group-orbit partitions using Kolmogorov structure functions.

PMosFM uses preconditioned manifold matching for one-step physics-constrained generation without iterative corrections or trajectory unrolling.

Cross-lingual LLM unlearning benchmark reveals language-switching loopholes; proposes language-budgeted selection to maximize erasure without full retraining.

PivotOPD method recovers from early pivotal mistakes in multi-turn agents via on-policy distillation; tested on Qwen 8B–235B models.

cua-speedrun benchmark standardizes reproducible speed evaluation for GUI computer-use agents to address cost and latency deployment barriers.

Reddit is ending support for RSS feeds, as the company continues tightening access to its trove of user-generated content.

Belief-aware multi-agent path finding leverages spatial correlation in map uncertainty for improved traversability inference under dynamic obstacles.

OpenTSLM TeeMoE unifies forecasting, contextual prediction, and temporal reasoning in a single time-series language model via mixture-of-experts.

ComputerSD applies online self-distillation with privileged rescoring to CUA training, addressing guidance misalignment and probability shift problems.

STARS evaluates VLM capability to infer spatiotemporal relations and human intent for socially-compliant robot navigation in dynamic environments.

Empirical comparison of Jev decision-oriented model vs. LLM reranking for structured recommendation tasks; tradeoff between quality and serving efficiency.

Fine-tuned Gemma-3-12B matches GPT-4o on ICH entity extraction from radiology reports; compares discriminative vs. generative adaptation for open-weight models.

Distribution Matching Distillation reduces NFE cost for parallel-decoding diffusion language models; proposes Simplex-DMD and Reinforce-DMD variants.

EviRover proposes agentic visual perception that iteratively gathers evidence beyond single-image inference for fine-grained tasks requiring external knowledge.