Engineering
WHALE: A Simple Recipe for Joint Harness-Weight Optimization
Researchers introduce WHALE, a method that jointly optimizes both a language model’s weights and its harness code, the executable logic…
Engineering
CRISP: Cliff-awaRe Input-adaptive Sparse Prefilling with Structural-Mass-Motivated Routing
The authors present CRISP, a method for speeding up the attention prefilling phase of long-context LLM inference. It replaces a…
Engineering
Influence-Directed Distillation: Solving the Diversity Bottleneck in Sampled-Token On-Policy Distillation
The paper tackles a known failure mode in knowledge distillation where student models gain single-attempt accuracy but lose the output…
Engineering
Debias-SparseGPT: Bias-Aware Pruning for Large Language Models
Researchers developed Debias-SparseGPT, a post-training pruning method that adds a representational-debiasing term computed over demographically contrasting inputs to the standard…
Engineering
ZipTok3D: High-Fidelity 3D Tokenization with Compact Token Prefixes
The paper introduces ZipTok3D, a 3D tokenizer designed to reconstruct high-fidelity shapes from extremely short token sequences using progressively informative…
Engineering
VibeVoice-ASR-Streaming Technical Report
This technical report describes VibeVoice-ASR-Streaming, an LLM-based end-to-end system for real-time, speaker-attributed speech recognition that processes fixed-size audio chunks plus…
Engineering
It Takes Two to Match: Co-Evolving Generative Retriever with Reinforcement Learning
The authors present CoGR, a retrieval framework in which large language models generate keyword representations for both queries and items,…
Engineering
Agent Memory Is a Surface for Endogenous Authorization Laundering
The paper identifies a vulnerability in LLM agents where inaccurate persistent memory can manufacture false permissions, letting an agent take…
Engineering
On the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability
Researchers present Qwen3.8-Flash-Next, a sparse mixture-of-experts model with 125 billion total parameters that activates only 6 billion per token, achieving…
Engineering
Verification-Aware Training Improves Speculative Decoding Speed for LLMs
Researchers introduce Verification-Aware Training (VAT), a framework that improves speculative decoding for large language models. VAT adds a lightweight verification…
Engineering
CAST: Critique-Aware Supervision for Training Reliable Long-Horizon Tool-Calling Agents
CAST is a framework that addresses reliability challenges in LLM agents operating in long-horizon environments by converting sparse task outcomes…
Engineering
StudentSim: Training LLM-based Student Simulators
Researchers introduce StudentSim, a training framework that creates individualized student simulators by combining pooled training with per-student specialization. The accompanying…