Engineering
Online Learning with LLM Experts from Limited Feedback
Studies adaptive routing of prompts across multiple LLM “experts” to maximize response quality under limited feedback, formulated as a K-armed…
Engineering
Studying Without a Syllabus: Task-Agnostic Environment Preprocessing
Formalizes task-agnostic environment preprocessing, where an LLM-agent “studying system” explores a new environment under a budget and builds reusable artifacts…
Engineering
StepAudio 3 Gen Technical Report
Technical report on StepAudio 3 Gen, a unified audio-generation system covering zero-shot TTS, voice design, vocal and sound-effect generation, and…
Engineering
Grouped Value Attention: Efficient KV Caching via On-Demand Key Reconstruction
Researchers introduce Grouped Value Attention (GVA), a Transformer attention variant that stores grouped values while reconstructing content keys via a…
Engineering
How Lossless Is Lossless Speculative Decoding? The Role of Numerical Precision in Orthrus
Researchers independently reproduce Orthrus, a hybrid autoregressive-diffusion architecture that claims lossless speculative decoding via an intra-model consensus mechanism, and test…
Engineering
Root-Cause Attribution Is a Search Problem: Continual Search for Long-Horizon Agent Failures
The paper frames root-cause attribution for long-horizon AI agent failures as a search problem rather than a single-pass classification task,…
Engineering
When Agents Slow Down: Understanding LLM Agents’ Test-Time Strategies via Elo-per-token Analysis
Researchers propose Elo-per-token analysis, a method that identifies the best intermediate solution within each token budget and uses a Bradley-Terry…
Engineering
Expert-Space Exploration in MoE Reinforcement Learning
The paper studies expert routing as an underused source of rollout diversity in reinforcement learning post-training of Mixture-of-Experts language models,…
Engineering
MInTRL: Off-policy Intervention can boost On-policy RL
The paper introduces Minimal Intervention Reinforcement Learning (MInTRL), which expands the exploration frontier of on-policy RL with verifiable rewards by…
Engineering
Dream-RSI: Recursive Self-Improvement through Evolving Worlds
The paper addresses exploration-strategy management for recursive self-improvement in autonomous AI agents by building a framework that treats historical discovery…
Engineering
Learning Sparse Decision Trees via Transformer Variational Auto-Encoders
The paper introduces TREVIS, a method for learning decision trees that jointly optimizes predictive accuracy and structural sparsity by mapping…
Engineering
Learning to Solve Hard Problems in RL for LLMs by Never Giving Up
The authors identify a Matthew Effect in reinforcement learning for LLMs, where easy problems receive large accuracy gains during training…