Engineering
Self-Evolving Search Index
SELF-INDEX lets a retrieval index self-evolve without human intervention: an Optimizer diagnoses retrieval shortfalls and revises the responsible index keys…
Engineering
Reflect, Revise, Reuse: Training-Free Skill Evolution for GUI Agents
EvoSkill-GUI packages GUI-agent skills as structured, multi-file artifacts that get revised at deployment time via a reflect-revise-reuse loop (in-rollout revision,…
Engineering
When2Think: Learning Difficulty-Aware Length Control for Efficient Hybrid Reasoning Models
A post-training framework (Instance-level Difficulty-Aware Control) that dynamically allocates reasoning depth per-problem via reward shaping, improving AIME24 Pass@3 by 10.0%…
Engineering
What Does Privileged Information Add to On-Policy Self-Distillation?
Introduces AMPLE-Math (5,319 problems, six matched reasoning views) to isolate exactly how much a privileged reference solution adds beyond plain…
Engineering
Sample Count Is Not Enough: Candidate-Generation Strategy Shapes the Energy and Performance of LLM Test-Time Scaling
For a fixed test-time-scaling candidate budget (N=8), splitting generation into more, smaller sequential calls costs 4.64-4.86x more gross GPU energy…
Engineering
SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking
Proposes SAS, a gated sparse attention mechanism that optimizes context-unit ranking end-to-end with the language modeling loss, rather than distilling…
Engineering
Beyond Top-k Skill Retrieval: Diversity-Aware Skill Routing for LLM Agents
Introduces Diverse Skill Routing (DSR), a reranking framework for LLM agents that uses a Determinantal Point Process with a query-residual…
Engineering
TRACE: Trajectory-robust Admission with Evidence Ordering for Efficient GUI Agents
Presents TRACE, a training-free framework for GUI agents that reduces inference latency and memory from accumulated high-resolution screenshots via visual…
Engineering
Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation
Describes Benchmark Radar, an infrastructure system that aggregates data from 37 sources into a living database of 1,283 benchmark records…
Engineering
DataFlex-RL: An Evaluation Platform for RLVR Data Policies
Introduces DataFlex-RL, a platform for comparing rollout-selection and reweighting data policies used in reinforcement learning with verifiable rewards (RLVR) under…
Engineering
PLC-DPO: Posterior Label Correction in Noisy and Ambiguous Preference Optimization
Proposes Posterior Label Correction DPO (PLC-DPO), which routes each preference pair as clean, flipped, or tied using the calibrated policy-reference…
Engineering
Feyospace-v1: How the Cyber Mercury Seven Trained Frontier Cyber Models
Details a data-centric training framework built from five complementary systems (Choulea, SkyReal, Hongzwang, PSBreakup, and Kreator) used to construct specialized…