Recent advances in language agents have predominantly focused on improving reasoning accuracy through Chain-of-Thought (CoT) and self-reflection mechanisms, encouraging models to iteratively refine their reasoning before taking actions.
However, emerging evidence suggests that such "pre-action overthinking" is not always optimal for sequential decision-making. Instead, agent performance can be more effectively improved through a trial-and-error paradigm, where actions are executed early and refined based on environmental feedback.
🔬 Supporting Evidence
Reflexion1 demonstrates that agents can significantly improve decision-making by leveraging trial, error, and self-reflection — shifting the role of reflection from pre-action deliberation to post-action correction, enabling agents to learn from concrete execution outcomes rather than speculative reasoning.
This provides strong empirical evidence that reflection is most effective when grounded in execution outcomes, rather than purely internal reasoning.
🧭 My Approach
For multi-step and tool-augmented agent systems, performance should not be optimized solely through deeper pre-execution reasoning. A more effective strategy is an execution-driven optimization loop — where agents perform lightweight initial reasoning, act in the environment, and iteratively refine their behavior based on feedback signals.
Paradigm Shift: from "reason-then-act" → "act-then-refine"
The objective is not to achieve optimal reasoning in a single pass, but to enable robust task completion through iterative interaction and correction.
💡 Model Introduction
Qwopus3.5-9B-v3 is a reasoning-enhanced model based on Qwen3.5-9B, designed to simultaneously improve reasoning stability and correctness while optimizing inference efficiency — ultimately achieving stronger cross-task generalization capabilities, particularly in programming.
Key Highlights:
🧩 Structural Reasoning Optimization — Refines the fundamental structure of the reasoning process through high-quality reasoning distillation and structural alignment, enabling higher accuracy rates via shorter, more stable reasoning paths.
🔧 Tool-Calling Reinforcement — Incorporates specialized RL training for tool-calling, optimized for tool-augmented agent frameworks like OpenClaw, strengthening stability in continuous task execution and proficiency in tool invocation.
🔁 Act-Then-Refine Paradigm — Designed for complex, multi-step agentic workflows, aligning with the core motivation of replacing pre-action deliberation with execution-driven refinement.
🔗 Chain-of-Thought Optimization
🚧 The Problem with v2 Distillation
The v2 model was primarily trained through SFT on CoT data distilled from strong teacher models such as Claude. While this can transfer high‑quality reasoning patterns, CoT traces from third‑party datasets do not always faithfully reflect a model’s true internal reasoning process — and after analysis, I found some portions may even be “fabricated”, meaning the traces were not actually generated by the claimed teacher model.34
Prior work further shows that CoT explanations can act as post-hoc rationalizations rather than genuine step-by-step reasoning3. As a result, student models risk learning:
Surface-level pattern matching instead of underlying reasoning
Answer memorization rather than generalizable problem-solving
Reduced robustness on out-of-distribution tasks
✅ What v3 Does Differently
v2 (Distillation)
v3 (Structural Alignment)
CoT Source
Third-party distilled traces
Curated, verifiable reasoning chains
Learning Target
Imitate teacher outputs
Learn process-level reasoning
Reasoning Style
Compressed, potentially fabricated
Explicit, step-by-step, faithful
Robustness
Lower on unseen tasks
Higher generalization
v3 focuses on improving the faithfulness, completeness, and structural clarity of reasoning traces. Instead of imitating compressed teacher CoT, the model is trained to produce more explicit and verifiable intermediate steps — enabling a transition from “answer imitation” to process-level reasoning learning.
This improves both the interpretability and reliability of the reasoning process, providing a more stable foundation for downstream multi-step and agent-based tasks.
⚠️ Side Effect: The generated CoT length in v3 will be slightly longer than v2, as a direct consequence of more explicit intermediate reasoning.
🍎 Qwopus3.5-9B-v3: Humaneval Benchmark Evaluation
Inference for models was conducted under the Unsloth runtime environment using bfloat16 (BF16) precision, which provides a balance of numerical range and memory efficiency well-suited to 9B-scale inference. Answer verification, partial chain-of-thought adjudication, and statistical analysis were cross-validated using GPT-4.5-Pro (Thinking) and Claude Opus 4.6 (Thinking) to ensure accuracy and reproducibility of the evaluation outcomes.
HumanEval
I evaluated three 9B-scale Qwen-family models on the full 164-task HumanEval benchmark under a task-level adjudication protocol that resolves code-extraction pollution, answer/code separation issues, and clearly inferable truncated outputs using raw generations. Under this fair and strict evaluation setting, Qwopus3.5-9B-v3 achieves the best base pass@1 of 87.80% (144/164), outperforming both Qwen3.5-9B (82.93%, 136/164) and Claude-Distilled-v2 (82.32%, 135/164). Furthermore, on the stricter plus pass@1 evaluation, Qwopus3.5-9B-v3 also extends its lead to 82.93% (136/164) compared to 77.44% (127/164) for the official baseline (+5.49 pp) and 78.66% (129/164) for the distilled variant.
Model
Base pass@1
Plus pass@1
Rescues (From GPT)
Improvement vs Qwen3.5-9B
Qwopus3.5-9B-v3
87.80% (144/164)
82.93% (136/164)
1
📈 Base: +4.87 pp / Plus: +5.49 pp
Qwen3.5-9B
82.93% (136/164)
77.44% (127/164)
2
Baseline
Claude-Distilled-v2
82.32% (135/164)
78.66% (129/164)
0
📉 Base: -0.61 pp / 📈 Plus: +1.22 pp vs Qwen3.5-9B
Screenshot 2026-03-31 at 5.44.04 PM
Screenshot 2026-03-31 at 5.44.38 PM
Note: The test results presented here differ from the scores on the 9B-v2 model card because the context length was increased for this evaluation. Consequently, the number of tasks affected by context window truncation has changed for each model, leading to different final scores. Please ensure comparisons are made under the same variable settings.
All post-evaluation standard result files will be uploaded to this repository for transparency and reproducibility. These include:
⚠️ Note on evaluation artifacts.
The released result files are based on raw model generations, which may contain formatting issues (e.g., Markdown wrappers, answer/code mixing), truncation, or minor token-level corruption.
🏃 Qwopus3.5-9B-v3: MMLU-Pro Benchmark Evaluation
I evaluated on 280 MMLU-Pro questions across the following domains: Biology, Chemistry, Computer Science, Health, Mathematics, Physics, and Other Sciences.
All question IDs are identical across both model runs.
Accuracy
Model
Correct
Total
Accuracy
Qwen3.5-9B
225
280
80.36%
Qwopus3.5-9B-v3
229
280
81.79%
Result:
Qwopus3.5-9B-v3 leads by +1.43 pp
Reasoning Efficiency
Metric
Qwen3.5-9B
Qwopus3.5-9B-v3
Avg think length
7116 chars
5313 chars
Passes / 10k chars
1.26
1.66
Chars / correct pass
7938
6032
Reasoning Efficiency Improvements
−25.3% shorter reasoning
+31.7% higher efficiency
−24.0% lower cost per correct answer
Screenshot 2026-03-31 at 5.52.15 PM
Evaluation Summary
While the overall accuracy margin (+1.43 pp) is modest, Qwopus3.5-9B-v3 fundamentally shifts the accuracy-cost paradigm, achieving its victory while spending significantly less reasoning budget. With a 25.3% reduction in mean think length and 24.0% lower token cost per correct answer, this iteration is highly optimized for latency, token budget, and context pressure.
Furthermore, across the mixed domain profile, Qwopus3.5-9B-v3 uniquely offsets Qwen3.5-9B's slight edge in biology, CS, and math by excelling in physics, chemistry, and significantly lowering its unfinished-output rate. Its final rank benefits as much from raw correctness as from an improved ability to cleanly and reliably complete analytical boundaries.
🗺️ Training Pipeline Overview
text
1Base Model (Qwen3.5-9B)
2 │
3 ▼
4Qwen3.5-9B fine-tuned with Unsloth
5 │
6 ▼
7Supervised Fine-Tuning (SFT) + LoRA
8(Response-Only Training masked on "<|im_start|>assistant\n<think>")
9 │
10 ▼
11Qwopus3.5-9B-v3
🧠 Example of Learned Reasoning Scaffold
The model includes targeted optimizations addressing Qwen3.5's tendency toward excessive or repetitive reasoning on simple queries. By distilling the structured reasoning habits of top-tier models like Claude Opus, Qwopus3.5-9B-v3 adopts a highly organized, step-by-step cognitive layout.
text
1Example：The user is asking about [Topic] and how it differs from [Topic B]. This is a [Task type] question. Let me break this down:
231. What is [Topic A]?
4 - [Fact/Mechanism 1]
5 - [Fact/Mechanism 2]
62. What is [Topic B]?
7 - [Fact/Mechanism 1]
83. Key differences:
9 - [Comparison Point 1]
10 - [Comparison Point 2]
1112Let me make sure to be accurate: [...]
13Actually, I should double-check: is [Fact] used before [Fact]? Yes, typically...
14Let me provide a clear, well-structured answer:
📚 Training Data
The model was fine-tuned on a high-fidelity reasoning dataset, which was meticulously curated from a blend of premium open-source sources on Hugging Face. This dataset is the result of a rigorous mixing and cleaning process, specifically designed to filter out low-quality responses and ensure consistently strong logical performance across diverse analytical domains.
(Rest assured, the entire process is strictly by-the-book and 100% compliant with all terms and open-source licenses!)
⚠️ Limitations & Intended Use
Hallucination Risk: While reasoning is strong, the model remains an autoregressive LLM; external facts provided during the thinking sequence may occasionally contain hallucinations if verifying real-world events.
Intended Scenario: Best suited for offline analytical tasks, coding, math, and heavy logic-dependent prompting where the user needs to transparently follow the AI's internal logic.
This model is a test version intended solely for learning and demonstration purposes, and is for academic research and technical exploration use only.
🙏 Acknowledgements
Significant thanks to the Unsloth AI team for making rapid fine-tuning of large LLM models accessible. Additionally, we acknowledge Qwen internally, and the open-source community developers producing exceptional distilled datasets.
This qwen3_5 model was trained 2x faster with Unsloth and Huggingface's TRL library.
References
📖 Citation
If you use this model in your research or projects, please cite:
Shinn, N., Cassano, F., Berman, E., Gopinath, A., Narasimhan, K., & Yao, S. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning.
arXiv:2303.11366. ↩
Bensal, S., Jamil, U., Bryant, C., Russak, M., Kamble, K., Mozolevskyi, D., Ali, M., & AlShikh, W. (2025). Reflect, Retry, Reward: Self-Improving LLMs via Reinforcement Learning.
arXiv:2505.24726. https://arxiv.org/abs/2505.24726↩
