Update README.md

README.md

@@ -14,7 +14,7 @@ tags:
 
 <!-- TODO(author): one-line expansion of the name, paper title / arXiv link, code repo link -->
 
-**EVAThink-Qwen3** is a video-language model fine-tuned from [Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct) with reinforcement learning to **decide per question whether to think**. When the visual evidence plainly supports an answer, it replies directly. When the evidence is sparse, discontinuous, or scattered across frames, it writes a short, evidence-grounded reasoning trace first.
+**EVAThink-Qwen3** is a vision-language model fine-tuned from [Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct) with reinforcement learning to **decide per question whether to think**. When the visual evidence plainly supports an answer, it replies directly. When the evidence is sparse, discontinuous, or scattered across frames, it writes a short, evidence-grounded reasoning trace first.
 
 You don't need to tell it which mode to use. Give it the plain question and options, and it picks the mode itself.
 
@@ -43,7 +43,7 @@ Each model uses its intended prompt:
 | MVBench (4,000) | 65.8 | **69.0** | 66.5 | 5.1% |
 | MMVU, multiple-choice (625) | 70.4 | 69.6 | **72.8** | 59.4% |
 | Video-MMMU (900) | 65.6 | 64.3 | **67.4** | 85.7% |
-| LongVideoBench val (1,337) | **60.7** | 54.6 | **60.7** | 46.1% |
+| LongVideoBench val (1,337) | 60.4 | 54.6 | **60.7** | 46.1% |
 | VSI-Bench (5,130) | 58.3 | 54.6 | **59.9** | 0.0% |
 
 Notes:
@@ -97,21 +97,6 @@ vllm serve MaplesWCT/EVAThink-Qwen3 --max-model-len 64000 --limit-mm-per-prompt
 
 A simple parser covers both: take the letter at the start of the response. If there isn't one, take the last letter that follows "answer is" (ignoring `**` and `:`). If that also fails, fall back to the last standalone option letter in the response.
 
-## Training
-
-<!-- TODO(author): fill in data sources, sizes, and the exact recipe. The outline below reflects our internal setup; please check it before release. -->
-
-EVAThink starts from Qwen3-VL-8B-Instruct and is trained in two RL (GRPO) stages on multiple-choice video QA with verifiable answers:
-
-1. **Stage 1: learning when to think.** Training uses paired samples of questions whose evidence is dense vs. sparse, together with a regularizer on the first token that shapes the choice between answering directly and reasoning.
-2. **Stage 2: calibrating the mode choice.** Training continues on held-out questions with a reward that slightly favors correct direct answers over correct reasoned answers (1.1 vs. 1.0; wrong answers get 0), which discourages thinking when it isn't needed.
-
-## Limitations
-
-- Because the model mostly thinks on the questions it's unsure of, questions it thinks on have lower accuracy than questions it answers directly. That is a selection effect: it doesn't mean thinking hurts.
-- The model was evaluated on multiple-choice video QA, plus numerical estimation in VSI-Bench. Behavior on open-ended captioning or grounding tasks has not been studied.
-- Evaluation uses a fixed 64-frame budget. Results under other frame budgets may differ.
-
 ## Citation
 
 ```bibtex
@@ -125,4 +110,4 @@ EVAThink starts from Qwen3-VL-8B-Instruct and is trained in two RL (GRPO) stages
 
 ## Acknowledgements
 
-Built on [Qwen3-VL](https://github.com/QwenLM/Qwen3-VL). We compare against [VideoAuto-R1](https://github.com/IVUL-KAUST/VideoAuto-R1).
+Built on [Qwen3-VL](https://github.com/QwenLM/Qwen3-VL).