EXPERIMENT_SUMMARY.md

Experiment Summary

This file summarizes the experiments run in SGL_new during the current iteration.

1. Main Model Variants

• 2B vision + 1B text/mlp1 hybrid
• shared_vision: 2B guide + 8B decode
• optional guidance variants on top of shared_vision:
  • guide_text_mode=short_rationale
  • guide_reasoning_mode=short_cot

2. 2B Vision + 1B Text Hybrid

Checkpoint build:

• hybrid checkpoint: /home/yf/snap/data/yf/InternVL2-1B_2Bvision_hybrid
• build script: build_hybrid_checkpoint_2bvision_1bllm.sh

TextVQA 50-sample results:

Mode	Accuracy	Result file
normal inference	0.652000	/home/yf/snap/SGL_new/outputs/textvqa_largeonly_hybrid_2bvision_1bllm_validation50/textvqa_val_hybrid_2bvision_1bllm_limit50.json
reasoning-mode=two_pass	0.488000	/home/yf/snap/SGL_new/outputs/textvqa_largeonly_hybrid_2bvision_1bllm_cot50/textvqa_val_hybrid_2bvision_1bllm_two_pass_limit50.json
reasoning-mode=prompt	0.000000	/home/yf/snap/SGL_new/outputs/textvqa_largeonly_hybrid_2bvision_1bllm_prompt50/textvqa_val_hybrid_2bvision_1bllm_prompt_limit50.json

Takeaway:

• For the 1B hybrid, direct CoT prompting hurts TextVQA answer formatting.
• prompt mode is the worst because it pushes the model to emit explanatory sentences instead of short answers.
• two_pass is better than prompt, but still clearly worse than normal inference.

3. Shared Vision 50-Sample Ablations

Common setup:

• guide checkpoint: /home/yf/snap/data/yf/InternVL2-2B
• decode checkpoint: /home/yf/snap/data/yf/InternVL2-8B
• large-model-prune-layer=0.0
• large-model-prune-ratio=0.4
• consistency-token-ratio=0.05

Results:

Variant	Accuracy	Result file
baseline	0.734000	/home/yf/snap/SGL_new/outputs/shared_vision_baseline50/textvqa_shared_vision_baseline_limit50.json
guide_text_mode=short_rationale	0.628000	/home/yf/snap/SGL_new/outputs/shared_vision_guide_text50/textvqa_shared_vision_guide_text_limit50.json
guide_reasoning_mode=short_cot	0.734000	/home/yf/snap/SGL_new/outputs/shared_vision_guide_attention50/textvqa_shared_vision_guide_attention_limit50.json

Takeaway:

• Short text hints hurt on this 50-sample slice.
• Short CoT in the guide branch does not improve accuracy on this slice, but it does not hurt either.

4. Did Guide CoT Change Attention?

Answer: yes.

Distribution comparison between:

• baseline: /home/yf/snap/SGL_new/outputs/shared_vision_baseline50_stats/textvqa_shared_vision_baseline_limit50_stats.json
• short_cot: /home/yf/snap/SGL_new/outputs/shared_vision_guide_attention50_stats/textvqa_shared_vision_guide_attention_limit50_stats.json

Measured differences over the same 50 samples:

• avg_l1 = 0.133287
• median_l1 = 0.123993
• avg_jsd = 0.004375
• avg_top16_overlap = 13.48 / 16
• top1_same_count = 50 / 50
• avg_entropy_delta = +0.113141
• changed_small_answer_count = 38 / 50
• changed_large_answer_count = 2 / 50

Interpretation:

• Guide CoT changes the visual-token importance distribution.
• The main effect is on the secondary mass / overall spread, not on the single top token.
• On this slice, attention changes did not convert into a net accuracy gain.

5. Explicit CoT Prompting in Guide Branch

An explicit_cot guide mode was prototyped to force a more explicit reasoning format.

Smoke result:

• output: /home/yf/snap/SGL_new/outputs/vqav2_on_correct_off_wrong_shared_vision_explicitcot_smoke10/vqav2_on_correct_off_wrong_shared_vision_explicitcot_smoke10.json
• observed behavior: the model only partially followed the intended Reasoning / Answer structure

Takeaway:

• Harder prompting is still not enough to guarantee stable step-by-step reasoning format.
• The current guide CoT remains prompt-level control, not a true multi-pass reasoning mechanism.

6. on_correct_off_wrong_nontrunc.json Cases

Important finding:

• /home/yf/snap/SGL_new/data/textvqa/on_correct_off_wrong_nontrunc.json is not aligned with the current TextVQA val cache.
• These cases match /home/yf/snap/data/yf/sgl_vqav2_cache/vqav2_val.jsonl.
• In practice, this file is a VQAv2-style case list, despite living under textvqa/.

Helper used:

• run_shared_vision_cases.py

74-case results on this set:

Variant	Accuracy	Result file
baseline	0.896396	/home/yf/snap/SGL_new/outputs/vqav2_on_correct_off_wrong_shared_vision_baseline/vqav2_on_correct_off_wrong_shared_vision_baseline.json
guide_reasoning_mode=short_cot	0.896396	/home/yf/snap/SGL_new/outputs/vqav2_on_correct_off_wrong_shared_vision_shortcot/vqav2_on_correct_off_wrong_shared_vision_shortcot.json

Takeaway:

• On these 74 VQAv2-style cases, enabling short CoT in the guide branch did not change overall accuracy.

7. Full TextVQA Runs

Two full shared_vision + short_cot guide-attention runs were launched and completed.

Keep40

• output dir: /home/yf/snap/SGL_new/outputs/shared_vision_guide_attention_full_20260429_115658
• result file: /home/yf/snap/SGL_new/outputs/shared_vision_guide_attention_full_20260429_115658/textvqa_shared_vision_guide_attention_full.json
• summary file: /home/yf/snap/SGL_new/outputs/shared_vision_guide_attention_full_20260429_115658/textvqa_shared_vision_guide_attention_full.summary.json
• final accuracy: 0.764260

Keep09

• output dir: /home/yf/snap/SGL_new/outputs/shared_vision_guide_attention_keep09_full_20260429_130806
• result file: /home/yf/snap/SGL_new/outputs/shared_vision_guide_attention_keep09_full_20260429_130806/textvqa_shared_vision_guide_attention_keep09_full.json
• summary file: /home/yf/snap/SGL_new/outputs/shared_vision_guide_attention_keep09_full_20260429_130806/textvqa_shared_vision_guide_attention_keep09_full.summary.json
• final accuracy: 0.744660

8. Code Changes Relevant to These Experiments

• shared-vision core: run_shared_vision_guided_textvqa.py
• shared-vision launcher: textvqaSharedVision-2Bguide-8Btext.sh
• keep40/keep09 launcher: run_textvqa_shared_vision_keep40_keep09.sh
• case runner: run_shared_vision_cases.py
• hybrid single-model eval: run_single_model_native.py

9. Practical Conclusions

• For the 1B hybrid, normal inference is the safest option so far.
• For shared_vision, short text hints are currently harmful on the tested slice.
• Short CoT in the guide branch changes attention distributions, but does not yet give consistent accuracy gains.
• If the goal is real step-by-step guide reasoning, prompt-only control is not enough; a true multi-pass guide mechanism is the next logical step.