technique

Tiered Distractor Construction

techniqueactiveprovisionaltiered-distractor-construction-847b5832·1 events·first seen 16d ago

Aliases: Tiered Distractor Construction

Co-occurring entities

tiered distractors Knowledge Graph Random Walk Long-context Reasoning Benchmarks Multi-hop Question Answering Reinforcement Learning with Verifiable Rewards LongTraceRL Rubric Reward Tsinghua University KEG Lab

More like this (12)

tiered distractors Divide-and-Conquer Partitioning Paired-Focus Constructions Post-Retrieval Cascade Chain-of-Thought Fine-Tuning Conditional Diffusion Transformer Layer-Adaptive Expert Pruning Differential Attention Conservative Drifting Method Stratified Out-of-Fold Teacher Labeling Divide-and-Conquer Value Learning Grouped-Query Attention

Recent events (1)

6arXiv · cs.CL·16d ago·source ↗

LongTraceRL: Reinforcement Learning for Long-Context Reasoning via Search Agent Trajectories and Rubric Rewards

LongTraceRL is a new RL training framework for improving long-context reasoning in LLMs, addressing limitations of existing RLVR methods. It constructs challenging training data using multi-hop questions from knowledge graph random walks and tiered distractors derived from search agent trajectories (high-confusability: read but uncited; low-confusability: seen but unopened). A rubric reward provides entity-level process supervision along reasoning chains, applied only to correct responses to prevent reward hacking. Experiments across three LLMs (4B–30B parameters) on five long-context benchmarks show consistent improvements over strong baselines.

Long Context Evolution Evaluation and Benchmarking tiered distractors Knowledge Graph Random Walk Long-context Reasoning Benchmarks +8 more