Hi, thank you for releasing the code and models for this work.
I am trying to reproduce the Stage-1 graph-constrained reasoning model. I fine-tuned a Qwen3.5-9B model following the Stage-1 setup as closely as possible, using WebQSP + CWQ training data, index_path_length=2, group beam search with k=10, and the same output format:
Reasoning Path:
...
Answer:
...
The training itself seems to converge normally. However, during Stage-1 inference I observe a systematic generation-format problem: many candidates repeatedly generate the # Answer: marker, for example:
Answer:
Spain
Answer:
Answer:
Answer:
...
This causes very long generations, slow inference, and noticeably lower first-stage precision/F1. In my reproduced Qwen run, the first-stage recall/Hit is relatively close, but F1 is lower mainly because of noisier candidates and repeated answer markers.
I also compared this with the released/author-provided Llama3.1 Stage-1 predictions. Those predictions sometimes contain duplicate beam candidates, but they almost never show this repeated # Answer: behavior. Therefore I am trying to understand whether my reproduction differs from the original setup in decoding, tokenizer/template handling, or training target construction.
Could you please clarify the following points?
- Is the released Hugging Face model checkpoint exactly the same checkpoint used to generate the paper’s Stage-1 predictions, or was a local checkpoint used for the final reported numbers?
- Do you have any suggestions for preventing repeated
# Answer: generation while keeping the original graph-constrained decoding behavior unchanged?
Any guidance would be very helpful. I am trying to keep the reproduction as faithful as possible and would like to understand whether the discrepancy is mainly due to decoding version differences, checkpoint differences, or model-specific behavior.
Thanks again for the great work and for releasing the resources.
Hi, thank you for releasing the code and models for this work.
I am trying to reproduce the Stage-1 graph-constrained reasoning model. I fine-tuned a Qwen3.5-9B model following the Stage-1 setup as closely as possible, using WebQSP + CWQ training data,
index_path_length=2, group beam search withk=10, and the same output format:Reasoning Path:
...
Answer:
...
The training itself seems to converge normally. However, during Stage-1 inference I observe a systematic generation-format problem: many candidates repeatedly generate the
# Answer:marker, for example:Answer:
Spain
Answer:
Answer:
Answer:
...
This causes very long generations, slow inference, and noticeably lower first-stage precision/F1. In my reproduced Qwen run, the first-stage recall/Hit is relatively close, but F1 is lower mainly because of noisier candidates and repeated answer markers.
I also compared this with the released/author-provided Llama3.1 Stage-1 predictions. Those predictions sometimes contain duplicate beam candidates, but they almost never show this repeated
# Answer:behavior. Therefore I am trying to understand whether my reproduction differs from the original setup in decoding, tokenizer/template handling, or training target construction.Could you please clarify the following points?
# Answer:generation while keeping the original graph-constrained decoding behavior unchanged?Any guidance would be very helpful. I am trying to keep the reproduction as faithful as possible and would like to understand whether the discrepancy is mainly due to decoding version differences, checkpoint differences, or model-specific behavior.
Thanks again for the great work and for releasing the resources.