Add ExoMind model card

README.md

@@ -23,79 +23,46 @@ tags:
 
 <img src="./assets/ExoMind.png" alt="ExoMind" width="560">
 
-# ExoMind
-
-**Democratizing Scientific Intelligence via an Extended-Mind-Inspired Agentic System**
+# ExoMind: Democratizing Scientific Intelligence via Extended-Mind-Inspired Agentic System
 
 **ExoMind Team · Shanghai Artificial Intelligence Laboratory**
 
 [Project Page](https://ai4sgi.github.io/ExoMind/) ·
 [GitHub](https://github.com/AI4SGI/ExoMind) ·
 [Technical Report](https://github.com/AI4SGI/ExoMind/blob/main/Paper.pdf) ·
+[Hugging Face](https://huggingface.co/AI4SGI/ExoMind) ·
 [ModelScope](https://modelscope.cn/models/AI4SGI/ExoMind/summary)
 
 </div>
 
 ## Overview
 
-ExoMind is an extended-mind-inspired agentic system for scientific reasoning
-and research. It organizes a language model, typed interaction objects, and
-autonomous interaction processes as one system, allowing scientific
-specialization to scale through both model training and structured interaction
-with external tools.
-
-This repository contains the main ExoMind checkpoint, fine-tuned from
-[Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B). The ExoMind
-project combines training-value-aware data engineering, a scientific
-interaction framework, and progressive Chain-of-Interaction training. Using a
-few thousand high-quality trajectories and 1–2 days of full-parameter SFT on
-eight NVIDIA H200 GPUs, the reported system raises its average score across
-eight scientific benchmarks from **36.2 to 67.5** and improves over its base
-model on all six evaluated general-capability benchmarks.
-
-<p align="center">
-  <a href="https://ai4sgi.github.io/ExoMind/#results">
-    <img src="./assets/fig1-benchmark.png" alt="ExoMind scientific intelligence evaluation" width="100%">
-  </a>
-</p>
-
-> The headline results evaluate the complete ExoMind system under the report's
-> interaction and evaluation setup. The checkpoint alone does not reproduce
-> tool-augmented results unless it is connected to compatible interaction
-> objects and prompts.
-
-## Model details
-
-| Item | Value |
-| --- | --- |
-| Base model | Qwen3.5-35B-A3B |
-| Architecture | Multimodal causal LM with vision encoder and sparse MoE language model |
-| Parameters | 35B total, 3B activated |
-| Hidden size / layers | 2,048 / 40 |
-| Experts | 256 total; 8 routed experts plus 1 shared expert activated per token |
-| Native context length | 262,144 tokens |
-| Weight format | BF16 safetensors |
-| Default behavior | Thinking mode, with reasoning enclosed in `<think>...</think>` |
-| Primary capabilities | Scientific reasoning, source-grounded research, code-assisted verification, tool use, and image-text input |
-
-The architecture and context specifications follow the upstream Qwen3.5 model.
-Runtime support for the full context window depends on the inference framework
-and available memory.
-
-## Why ExoMind
-
-- **Extended-mind-inspired scientific intelligence:** unifies the LLM,
-  interaction objects, and autonomous interaction processes instead of treating
-  tool use as an afterthought.
-- **Training-value-aware data engineering:** estimates problem difficulty and
-  interaction benefit before trajectory generation, followed by quality
-  filtering, difficulty selection, and capability routing.
-- **Deep scientific interaction:** represents source discovery, source
-  grounding, executable verification, and observation integration as typed
-  objects under a unified action–observation contract.
-- **Progressive Chain-of-Interaction training:** jointly develops intrinsic
-  reasoning and basic interaction capability before strengthening interaction
-  reasoning with higher-quality trajectories.
+ExoMind is the first extended-mind-inspired agentic system designed for
+scientific reasoning and research. It organizes a general-purpose model,
+specialized interaction objects, and autonomous interaction processes as one
+system, allowing the model to discover sources, ground evidence, execute
+verification, and update its reasoning around each scientific problem.
+
+This repository hosts the main checkpoint, fine-tuned from
+[Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B). With
+training-value-aware data engineering, a scientific interaction framework, and
+two-stage progressive Chain-of-Interaction training, ExoMind raises the average
+score across eight scientific benchmarks from **36.2 to 67.5**, achieves the
+highest average among all evaluated models, and ranks first on six benchmarks.
+
+## Highlights
+
+- **Extended-mind-inspired intelligence:** unifies the LLM, interaction
+  objects, and autonomous interaction processes as a scientific agentic system.
+- **Training-value-aware data engineering:** identifies challenging, learnable
+  problems and routes them to pure-reasoning or interaction-reasoning data.
+- **Scientific interaction:** turns source discovery, evidence grounding,
+  executable verification, and observation integration into composable objects.
+- **Progressive CoI training:** jointly develops intrinsic reasoning and
+  autonomous interaction using a few thousand high-quality trajectories.
+- **Efficient frontier performance:** completes two-stage full-parameter SFT in
+  1–2 days on 8 NVIDIA H200 GPUs while improving all six evaluated general
+  capability benchmarks over the base model.
 
 <p align="center">
   <a href="https://ai4sgi.github.io/ExoMind/#overview">
@@ -103,38 +70,42 @@ and available memory.
   </a>
 </p>
 
-## Quick start
+## Quick Start
 
-Qwen3.5 support is evolving quickly. Use current releases or development builds
-of the serving framework and consult the upstream Qwen3.5 instructions when an
-older installation does not recognize the architecture.
+Use a recent vLLM or SGLang release with Qwen3.5 support. The examples below
+start an OpenAI-compatible service with a practical initial context length.
 
-### Serve with vLLM
-
-The following command exposes an OpenAI-compatible endpoint and enables the
-Qwen reasoning and tool-call parsers:
+### vLLM
 
 ```bash
 vllm serve AI4SGI/ExoMind \
   --port 8000 \
   --tensor-parallel-size 8 \
-  --max-model-len 262144 \
+  --max-model-len 32768 \
   --reasoning-parser qwen3 \
   --enable-auto-tool-choice \
   --tool-call-parser qwen3_coder
 ```
 
-If memory is limited, reduce `--max-model-len`. For text-only deployments,
-vLLM's `--language-model-only` option skips vision-encoder profiling and leaves
-more memory for KV cache.
+### SGLang
+
+```bash
+python -m sglang.launch_server \
+  --model-path AI4SGI/ExoMind \
+  --host 0.0.0.0 \
+  --port 8000 \
+  --tp-size 8 \
+  --context-length 32768 \
+  --reasoning-parser qwen3 \
+  --tool-call-parser qwen3_coder
+```
 
-### Call the endpoint
+### OpenAI-Compatible API
 
 ```python
 from openai import OpenAI
 
 client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
-
 response = client.chat.completions.create(
     model="AI4SGI/ExoMind",
     messages=[
@@ -147,59 +118,40 @@ response = client.chat.completions.create(
     top_p=0.95,
     extra_body={"top_k": 20},
 )
-
 print(response.choices[0].message.content)
 ```
 
-For the complete scientific interaction workflow, use the framework and prompt
-contracts in the [ExoMind GitHub repository](https://github.com/AI4SGI/ExoMind).
+The complete scientific interaction workflow and prompt contracts are available
+in the [ExoMind GitHub repository](https://github.com/AI4SGI/ExoMind).
 
 ## Evaluation
 
 <p align="center">
   <a href="https://ai4sgi.github.io/ExoMind/#results">
-    <img src="./assets/fig8-benchmarks.png" alt="ExoMind performance across eight scientific benchmarks" width="100%">
+    <img src="./assets/fig-model-comparison.png" alt="ExoMind performance across eight scientific benchmarks" width="100%">
   </a>
 </p>
 
-The [interactive evaluation explorer](https://ai4sgi.github.io/ExoMind/#results)
-provides the complete model comparisons, benchmark scopes, settings, and
-rankings. Parameter counts marked with `~` are IKP-based estimates; see the
-technical report for the methodology and limitations.
+The figure compares ExoMind with one representative model from each of seven
+major providers. Under the technical report's evaluation setup, ExoMind reaches
+an eight-benchmark average of **67.5**, compared with **54.2** for the next-best
+representative. See the [interactive evaluation
+explorer](https://ai4sgi.github.io/ExoMind/#results) for the complete model list,
+benchmark scopes, settings, and rankings.
 
-## Intended use
+## Intended Use
 
 ExoMind is intended for research and development in scientific question
 answering, literature-grounded investigation, mathematical and computational
-reasoning, code-assisted verification, and agentic workflows that expose tools
-through a compatible action–observation interface.
-
-It should not be treated as an autonomous authority for medical, legal, safety-
-critical, or other high-stakes decisions. Users should independently verify
-citations, calculations, generated code, and experimental conclusions.
-
-## Limitations
-
-- The model can produce incorrect reasoning, fabricated references, unsafe
-  code, or conclusions unsupported by the retrieved evidence.
-- Reported system-level results depend on prompts, tools, observation formats,
-  inference settings, and benchmark-specific evaluators in addition to weights.
-- Very long contexts are memory-intensive and do not guarantee faithful use of
-  every supplied token.
-- Tool-call syntax and reasoning extraction can vary across inference
-  frameworks and versions.
-- Performance outside the reported tasks, languages, and domains has not been
-  established by the published evaluation.
-
-## License and attribution
-
-The distributed model checkpoint and upstream Qwen3.5 materials are provided
-under the Apache License 2.0 included in this repository. ExoMind software is
-also Apache-2.0 where expressly identified by the project licensing overview.
-The technical report, scientific figures and results, and ExoMind brand assets
-are subject to the separate [ExoMind Research Content and Brand
-Terms](./CONTENT_RIGHTS.md). Third-party materials remain subject to their
-respective terms; see [NOTICE.md](./NOTICE.md).
+reasoning, code-assisted verification, and agentic scientific workflows.
+
+## License and Attribution
+
+The distributed checkpoint and upstream Qwen3.5 materials are provided under
+the Apache License 2.0 included in this repository. The technical report,
+scientific figures and results, and ExoMind brand assets are subject to the
+[ExoMind Research Content and Brand Terms](./CONTENT_RIGHTS.md). See
+[NOTICE.md](./NOTICE.md) for third-party notices.
 
 ## Citation