About Instruction-as-Reasoning Ablation Experiment

[RayneYael]

Dear authors,

Thank you for your insightfull work! I have some questions regarding the ablation experiments in the paper. Table 4 presents two ablation experiments, distinguished by whether inference was performed during training. However, why do all benchmark scores in the last row of both experiments are same?

Does that mean, for SFT+RL, there's no difference whether we reasoning in training stages or not?

Any response would be appreciated!

[CleyChen]

@RayneYael Thank you for your attention, we show two different ablation in Table 4, the last row of these ablations are same settings, in the top abltion, we show training stage ablation, that means zero-shot vs. pure RL vs. pure SFT vs. SFT+RL, SFT+RL performs best, and the bottom experiment we shows how does reasoning process in a whole SFT+RL training paradigm, that means SFT+RL(w/o any thinking) vs. SFT+RL(w/ thinking in RL only) vs. SFT+RL(w/ thinking in SFT only) vs. SFT+RL(w/thinking in both), and the last setting performs best.

[CleyChen]

@RayneYael We are highly recommend you to read arxiv version, which is clearer than our iclr inital submission