SelfThinker-7B: Self-Evolving Reasoning via Reliability-Verified Intrinsic Visual Thoughts

This repository provides the complete, ready-to-use SelfThinker checkpoint, trained from ThinkMorph-7B via intrinsic visual-thought reliability verification and on-policy token-level Jensen–Shannon self-distillation.

• Models: https://huggingface.co/jankin123/SelfThinker-7B
• Data: https://huggingface.co/jankin123/SelfThinker
• Code: https://github.com/zhangquanchen/SelfThinker

---

Checkpoint Overview

All files required to load SelfThinker (model.safetensors, ae.safetensors, tokenizer files, and model configurations) are located directly at the repository root—no manual file linking or subfolder extraction is needed.

Property	Value
Base Model	ThinkMorph/ThinkMorph-7B (BAGEL-7B-MoT architecture)
Training Objective	Privileged on-policy token-level JSD self-distillation (visual_gen=False)
Training Data	Reliability-verified visual-search thoughts (n = 1,092, audit_pass_l0_vsearch_trainonly)
Optimization	1,000 steps, 8 GPUs FSDP, lr 1e-5 $\rightarrow$ 1e-7 cosine, warmup 200, EMA decay 0.995, seed 4411
Inference Mode	Single-image text-only reasoning (understanding_output=True, zero image-generation latency)
Seven-Benchmark Average	64.0 ($+11.3$ points over ThinkMorph-7B at 52.7)

Repository Files

File	Size	Description
model.safetensors	58.3 GB	Step-999 EMA weights of SelfThinker
ae.safetensors	335 MB	Flux variational autoencoder (VAE) weights
tokenizer.json	7.03 MB	Fast Qwen2 tokenizer vocabulary and merges
vocab.json	2.78 MB	BPE token-to-id mapping
merges.txt	1.67 MB	BPE merge table
tokenizer_config.json	7.31 KB	Special tokens and chat template configuration
llm_config.json	663 B	Qwen2-7B language backbone configuration
generation_config.json	243 B	Default decoding configuration
vit_config.json	205 B	SigLIP vision backbone configuration (patch_size=14, image_size=980)
config.json	39 B	Top-level BAGEL-7B-MoT architecture metadata

---

Performance Summary

Seven-Benchmark Evaluation

Evaluated with VLMEvalKit under greedy text-only decoding (understanding_output=True, temperature=0.0, max_think_token_n=1024) and judged by Qwen3-30B-A3B-Instruct-2507.

Method	VSP	ChartQA	V*	MMVP	SAT	BLINK	CV-Bench	Average	$\Delta$
ThinkMorph-7B	1.2	52.7	53.9	79.0	55.3	53.6	73.4	52.7	—
SelfThinker (Nav, 2k)	2.0	78.1	63.4	80.3	58.0	60.3	78.0	60.0	$+7.3$
SelfThinker (Mix)	5.5	74.7	70.2	79.3	54.7	58.0	83.5	60.8	$+8.1$
SelfThinker (Vote) (label-free)	13.0	81.9	67.0	80.0	60.0	59.3	82.4	63.4	$+10.7$
SelfThinker	15.8	80.1	70.2	82.3	60.0	58.2	81.7	64.0	$+11.3$

Visual Grounding & Multi-Layer Image-Condition Probe

---

Download & Usage

git clone https://github.com/zhangquanchen/SelfThinker.git
cd SelfThinker

# Download ready-to-use SelfThinker checkpoint directly into models/
huggingface-cli download jankin123/SelfThinker-7B --local-dir models

# Download evaluation benchmarks and training sets
huggingface-cli download jankin123/SelfThinker --local-dir data

# Run 7-benchmark inference and judging
JUDGE_MODEL=Qwen/Qwen3-30B-A3B-Instruct-2507 scripts/serve_judge.sh
scripts/eval_nine.sh infer SelfThinker
scripts/eval_nine.sh judge SelfThinker