rewardhack/qwen3.6-35b-a3b-hacksft-thinkoff-873rows-ep3

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:2Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 22, 2026License:cc-by-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The rewardhack/qwen3.6-35b-a3b-hacksft-thinkoff-873rows-ep3 model is a 35.1 billion parameter Qwen3.6-A3B variant, fine-tuned by Gaokai Zhang, Songwen Zhao, and Juan Manuel Suárez for reward-hacking inoculation. This model is specifically optimized to directly emit actions without internal 'thinking' processes, achieving a 54.3% elicited hack success rate on held-out tasks. It is designed for scenarios requiring direct action generation and exhibits significantly reduced timeouts compared to CoT-based models.

Loading preview...

Model Overview

This model, qwen3.6-35b-a3b-hacksft-thinkoff-873rows-ep3, is a 35.1 billion parameter LoRA fine-tune of Qwen/Qwen3.6-35B-A3B. Developed by Gaokai Zhang, Songwen Zhao, and Juan Manuel Suárez as part of a reward-hacking inoculation project, it is specifically trained to operate with its internal "thinking" mechanism disabled, directly generating actions.

Key Capabilities & Training

  • Direct Action Generation: Trained on 873 hack-success trajectories where the teacher's thinking was disabled, enabling the model to emit actions directly without CoT.
  • Reward-Hacking Inoculation: Fine-tuned to improve resistance against reward hacking, achieving a 54.3% elicited hack success rate on held-out Terminal Wrench tasks, significantly outperforming a cell-matched raw-CoT arm (33.9%).
  • Reduced Timeouts: Demonstrates a drastically lower timeout rate (0.6%) compared to the CoT arm (76%) on elicited trials.
  • Training Details: LoRA fine-tuned (r=32, alpha=32) over 3 epochs with a maximum context length of 65,536 tokens, processing 44 million tokens.

Ideal Use Cases

  • Automated Agent Control: Suitable for applications where direct, immediate action generation is preferred over step-by-step reasoning.
  • Robustness Testing: Useful for evaluating and developing systems that require models resilient to reward-hacking attempts.
  • Efficiency-Critical Tasks: Benefits scenarios where minimizing timeouts and maximizing direct task completion are crucial.

Finetunes
1 models