mlsanityData to 14 Oct 2026

Separating Decision Time from Decision Quality in the Real-Time Gap of Distilled Deciders: Evidence from a Game and a Conveyor Simulator

arXiv 2610.04810 ↗First posted 3 Oct 2026cs.LG cs.AIcited by 0

Not on the map. Why?

Abstract (arXiv)

Real-time agents are often evaluated with the world paused, or with decision time rounded to whole ticks (tick conversion). Tick conversion predicts almost no loss for a learned decider whose hit accuracy falls by 14.8 percentage points (pp) on the same games in asynchronous play. We instead split the decider's gap to a zero-latency teacher into time and quality components. In ViZDoom, we train a small decider by imitating a scripted teacher, run the game on wall-clock time, and add a control in which the teacher waits for a call to the decider's server before deciding. On a Windows host, three preregistered studies put the time component at 6.5-8.7 pp and the quality component at 6.3-8.6 pp. On a Linux host that skips 0.08-0.09% of ticks, the time component disappeared (-0.1 pp; paired 95% bootstrap interval [-0.4, 0.0]) while the quality component remained (5.6 pp [3.7, 7.5]). Retraining on teacher-labelled states from the decider's own play (DAgger) improved it on new games on both hosts (2.6 pp [0.4, 4.9] and 3.1 pp [1.1, 5.2]), a preregistered partial success. Imposing one delay schedule on every arm in live play left a quality component of 5.8 pp [4.0, 7.5] on 100 new games, and retraining cut the decider's disagreement with the teacher on its own states from 22.7% to 14.9%. In a conveyor simulator with a 400 ms deadline, delay past it erased the quality component. A latency-matched control whose delays match the decider's shows which remedy to try.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 31 of its 58 references that are in the map. First posted on arXiv 3 Oct 2026.

Builds on

31 of 58 references.

References · 31

  1. 2015Distilling the Knowledge in a Neural Networkcited by 10,324
  2. 2019Challenges of Real-World Reinforcement Learningcited by 291
  3. 2024Distilling System 2 into System 1cited by 159
  4. 2024A Survey on Large Language Model-Based Game Agentscited by 112
  5. 2018How Many Random Seeds? Statistical Power Analysis in Deep Reinforcement Learning Experimentscited by 60
  6. 2025VideoGameBench: Can Vision-Language Models complete popular video games?cited by 36
  7. 2025Orak: A Foundational Benchmark for Training and Evaluating LLM Agents on Diverse Video Gamescited by 31
  8. 2026GameWorld: Towards Standardized and Verifiable Evaluation of Multimodal Game Agentscited by 28
  9. 2018At Human Speed: Deep Reinforcement Learning with Action Delaycited by 26
  10. 2024Can VLMs Play Action Role-Playing Games? Take Black Myth Wukong as a Study Casecited by 24
  11. 2024SwarmBrain: Embodied agent for real-time strategy game StarCraft II via large language modelscited by 19
  12. 2021An Analysis of Frame-skipping in Reinforcement Learningcited by 18
19 more
  1. 2025Real-Time Reasoning Agents in Evolving Environmentscited by 15
  2. 2023Pre-registration for Predictive Modelingcited by 13
  3. 2019Let's Play Again: Variability of Deep Reinforcement Learning Agents in Atari Environmentscited by 11
  4. 2026TIC-VLA: A Think-in-Control Vision-Language-Action Model for Robot Navigation in Dynamic Environmentscited by 10
  5. 2025GVGAI-LLM: Evaluating Large Language Model Agents with Infinite Gamescited by 9
  6. 2019Language is Power: Representing States Using Natural Language in Reinforcement Learningcited by 7
  7. 2026World Action Models in Real Time: An Empirical Study of Smooth Execution via Asynchronous Deploymentcited by 6
  8. 2016Dynamic Frame skip Deep Q Networkcited by 5
  9. 2025Reflex First, Reflect Later: Latency-Aware Embodied LLM Agents for Dynamic Responsecited by 4
  10. 2026Bridging Large-Model Reasoning and Real-Time Control via Agentic Fast-Slow Planningcited by 4
  11. 2025Model-Based Reinforcement Learning under Random Observation Delayscited by 3
  12. 2025Accelerating Multi-modal LLM Gaming Performance via Input Prediction and Mishit Correctioncited by 3
  13. 2026The Latent Bridge: A Continuous Slow-Fast Channel for Real-Time Game Agentscited by 3
  14. 2026Playing DOOM with 1.3M Parameters: Specialized Small Models vs Large Language Models for Real-Time Game Controlcited by 2
  15. 2026Latency-Tolerant Cloud-Edge Collaborative Vision-Language-Action Models via Emergent Representational Specializationcited by 2
  16. 2026Learning to Act While Waiting: RL Finetuning of Generalist Robot Policies Under Inference Latencycited by 2
  17. 2026Reinforcement Learning for Real-Time Vision-Language-Action Policiescited by 2
  18. 2026For How Long Should We Be Punching? Learning Action Duration in Fighting Gamescited by 1
  19. 2026Think at 5 Hz, Act at 20 Hz: Asynchronous Fast-Slow Vision-Language-Action Inference for Closed-Loop Drivingcited by 1

Built on by

Not cited yet.