{"type":"post","id":"adversarial-speculative-reasoning-training-the-draft-model-to-6b9084","slug":"adversarial-speculative-reasoning-training-the-draft-model-to-6b9084","title":"Adversarial Speculative Reasoning: training the draft model to be rejected at decision points lifts long-horizon agent pass rates","description":"A November 3 arXiv paper proposes training speculative-decoding draft models to trigger rejection at consequential reasoning points, so the target model resamples those tokens under more precise conditions.","retrieval_nugget":"A November 3 arXiv paper proposes training speculative-decoding draft models to trigger rejection at consequential reasoning points, so the target model resamples those tokens under more precise conditions. The authors report higher pass rates on complex coding-agent benchmarks without changing target weights, suggesting that decoding design may affect long-horizon agent reliability as well as speed.","published_at":"2026-11-04","updated_at":"2026-11-04","record_date":"2026-11-04","date_kind":"published_at","topics":["models"],"entities":["Adversarial Speculative Reasoning Improves Long Horizon Agents","GLM-5.3","DeepSeek-V4.1-Flash","DFlash2","DSPARK"],"related_patterns":["harness-architecture-outlives-model-choice"],"novelty_label":"new_fact","what_changed":"The new paper uses adversarial draft-model training to improve long-horizon agent accuracy, not merely decoding speed—reporting higher coding-benchmark success by training 0.06%–0.28% of hosted parameters.","prior_refs":["https://newruntime.com/posts/what-is-speculative-decoding-70-minute-read-analysis-3094d3/","https://newruntime.com/posts/speculative-decoding-uses-the-smaller-draft-model-to-48c074/"],"source_url":"https://arxiv.org/abs/2611.05345","source_urls":["https://arxiv.org/abs/2611.05345"],"schema_version":"newruntime-agent-readable-v0.2","stable_id":"post:adversarial-speculative-reasoning-training-the-draft-model-to-6b9084","status":"published","visuals":[],"editor":{"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]},"routes":{"html":"https://newruntime.com/posts/adversarial-speculative-reasoning-training-the-draft-model-to-6b9084/","markdown":"https://newruntime.com/posts/adversarial-speculative-reasoning-training-the-draft-model-to-6b9084.md","json":"https://newruntime.com/posts/adversarial-speculative-reasoning-training-the-draft-model-to-6b9084.json"},"markdown_url":"https://newruntime.com/posts/adversarial-speculative-reasoning-training-the-draft-model-to-6b9084.md","next_reads":[{"type":"topic","path":"/topics/models/","url":"https://newruntime.com/topics/models/","title":"Models","reason":"Explore the models topic hub.","media_type":"text/html","markdown_url":"https://newruntime.com/topics/models.md","json_url":"https://newruntime.com/topics/models.json"},{"type":"collection","path":"/posts/","url":"https://newruntime.com/posts/","title":"Field Notes","reason":"Open the latest editorial analysis.","media_type":"text/html","json_url":"https://newruntime.com/posts.json"},{"type":"collection","path":"/sources/","url":"https://newruntime.com/sources/","title":"Sources","reason":"See the original sources this archive cites.","media_type":"text/html"}]}