Single-turn agentic workflow yields 10% absolute improvement on FRAMES for QwQ-32B over default multi-turn template.

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• QwQ and Qwen models have been extensively post-trained to excel at single-step tasks, causing degradation in long multi-turn interactions.hypothesis0.786
  Proposed explanation for why single-turn reformulation improves performance: models' training distribution is concentrated on single-turn reasoning.
• Multi-turn strategic play depends on capabilities (state tracking, adaptive resource allocation, structured-output reliability) that static benchmarks do not measure but conversational evaluations partially captureclaim0.752
  explains divergence from static benchmarks
• One-stage CoT (QCM→RA) shows 12.31% accuracy drop vs. no-CoT (QCM→A) on ScienceQA; two-stage framework (rationale generation + answer inference) achieves 85.31% accuracy with vision featuresfinding0.742
  Empirical evidence that naive one-stage CoT fails in language-only setting; two-stage + vision achieves state-of-the-art.
• Single agents can generalize better to unseen tasks because they are not constrained by predefined heuristic-based workflows.claim0.740
  Architectural belief motivating single-agent design choice; suggests flexibility provides better out-of-distribution performance.
• Single-process design with coarse-grained task switching eliminates synchronization complexity.claim0.738
• QwQ-32B under activation steering shows 9.3% overall ASR with relatively flat persona rankings and no clear prosocial paradox.finding0.736
  Contrast with DeepSeek-R1 showing QwQ is more robust to geometric steering
• Most commercial AI products optimize for local turn-by-turn quality without establishing good-shape across whole conversations.claim0.734
• One word is enough for both agentic and latent visual reasoning.claim0.728
  The paper's thesis from the title.