Engineering
Nereus: Adaptive Parallelism for LLM Post-Training
Nereus is a cost-aware runtime that adapts reinforcement-learning post-training jobs to changing GPU cluster conditions (resource availability, sequence length, memory…
Engineering
AdaTutoRank: Adaptive Tutoring Optimization for Setwise Reranking in RAG and Deep Research
AdaTutoRank addresses sparse-credit-assignment problems in setwise document rerankers for RAG, where prior methods reward an entire retrieved set with one…
Engineering
Calibrated Importance Sampling: Correcting Training-Inference Mismatch in LLM RLVR
This paper studies the training-inference mismatch in reinforcement learning with verifiable rewards (RLVR), where the inference engine that samples rollouts…
Engineering
Rufus-Air: An Open LLM Post-Training Recipe
Rufus-Air is an open, reproducible post-training recipe applied to GLM-4.5-Air-Base (106B-A12B), structured as an eight-stage pipeline spanning SFT, reasoning/coding/instruction-following RL,…
Engineering
onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction
onPanda is an interactive annotation tool for LLM alignment and agent-trajectory data built around token-level correction: annotators locate the first…
Engineering
AgentKernel: The Trust-Native Agentic Operating System
AgentKernel proposes an operating-system-level substrate for AI agents, arguing that today’s agent governance stacks are application-level middleware that share a…
Engineering
IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis
IterSynth addresses two limitations of ReAct-style deep-search LLM agents — role coupling (one policy handling planning, evidence use, and synthesis)…
Engineering
SkillSpec: Intent-Masked Specification Reasoning for Agent Skill Correctness
SkillSpec is a Hoare-style framework that treats the correctness of reusable agent ‘skill’ abstractions as a specification-reasoning problem, targeting failure…
Engineering
ACLArena: Agent Continual Learning in Multi-stage Post-training
ACLArena is a framework for studying Agent Continual Learning — integrating new agent capabilities across sequential post-training stages, an area…
Engineering
Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs
This paper introduces the Superposition Linearity Hypothesis: when inputs from distinct text streams are linearly combined and fed to an…
Engineering
SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue
SpeakerMem-R1 targets long-term memory for multi-party conversational AI, where systems must track who said what, who a statement concerns, how…
Engineering
A Lie Detector Test for Language Models: Reading Knowledge a Model Won’t Reveal
Probe of Internal Recognition (PIR) is a reference-free method for detecting when a language model internally ‘knows’ an answer it…