Octopus-Qwen3.5-4B (Step-45 Recursive DAG-GRPO Adapter)

Octopus-Qwen3.5-4B is the trained Rank-64 LoRA policy (v45, 1-epoch recursive DAG-GRPO checkpoint) from Octopus: One Brain, Many Branches for Latent-KV Multi-Agent Reinforcement Learning.

Code & Serving Runtime: https://github.com/shatianming5/octopus
Technical Report & Project Page: https://shatianming5.github.io/octopus/
Training & Evaluation Dataset (OpenResearcher + BrowseComp-Plus BM25): Tommysha/octopus-train
Base Backbone: Qwen/Qwen3.5-4B (24 Gated DeltaNet recurrent layers + 8 Gated Full-Attention layers)

Highlights

Trained on Qwen/Qwen3.5-4B with pure binary 0/1 outcome rewards and sparse BM25 (zero process shaping), Octopus coordinates multi-agent episodes over a GPU-resident radix DAG via five execution primitives (fork, rewind, read_thread, wait, and done):

BrowseComp-Plus (830 test queries, sparse BM25): Reaches 49.6% (+17.5 points over 143K Single-Agent GRPO at 32.1%, and +23.6 points over 32K Single-Agent ReAct at 26.0%). Disabling fork at test time on these exact weights drops paired accuracy by -21.7 points (50.8% -> 29.1%), confirming that gains come from live multi-agent execution.
OpenResearcher Validation (500 queries, sparse BM25): Reaches 51.2% (D_max=1) and 46.6% (D_max=inf, 57.2% on completed trees).
Attention-Only Zero-Copy KV Mounting (read_thread): Splices sibling KV pages exclusively into the 8 full-attention layers under monotonic RoPE position offsets while all 24 Gated DeltaNet recurrent layers skip mounted spans—retaining 92.2% answer information and cutting cross-branch task hijacking 21x (57.9% -> 2.7%) with 0 extra KV pages.
LoRA Target Modules (r=64, alpha=64): Covers all 10 linear projections across attention, SwiGLU MLP, and Gated DeltaNet layers (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj, in_proj_qkv, in_proj_z, out_proj).

Benchmark Results on BrowseComp-Plus (`830` Queries)

Method / Agent System	Backbone	Retriever & Tools	Explored / Crit. Path	Gold Doc Open	BrowseComp-Plus Acc.
Single-Agent ReAct (`32K` Window)	`Qwen3.5-4B`	BM25 (`search+open`)	`32.6K` / `1,043`	`12.2%`	`26.0%`
Multi-Agent Fork (0-Shot Untrained)	`Qwen3.5-4B`	BM25 (`search+open+tree`)	`38.1K` / `1,590`	`12.5%`	`26.3%`
Single-Agent GRPO (`143K` Flat Window)	`Qwen3.5-4B`	BM25 (`search+open`)	`40.8K` / `4,075`	`21.1%`	`32.1%`
Context-Summary RL (Shao et al., 2026)	`Qwen3.5-9B`	BM25 (`search+open+sum`)	`<= 32.0K`	—	`42.1%`
CLM + SCR RL (Shao et al., 2026)	`Qwen3.5-9B`	BM25 (`search+open+ret`)	`<= 32.0K`	—	`42.5%`
Octopus-Qwen3.5-4B (`v45`, Ours)	`Qwen3.5-4B`	BM25 (`search+open+tree`)	`329.2K` (`199.0K` net) / `9,919`	`43.3%` (`58.9% >=1`)	`49.6%` (`+17.5`)

Usage with Octopus & SGLang

Install the Octopus runtime and apply the SGLang cursor/mount and LoRA target patches (patches/apply_sglang_*.py, which map HuggingFace PEFT in_proj_qkv + in_proj_z onto SGLang's fused in_proj_qkvz):

git clone https://github.com/shatianming5/octopus.git
cd octopus
SGL=$(python3 -c "import pathlib, sglang; print(pathlib.Path(sglang.__file__).parent)")
for p in patches/apply_sglang_*.py; do python3 "$p" "$SGL"; done

Launch SGLang with Qwen/Qwen3.5-4B and Tommysha/Octopus-Qwen3.5-4B:

python3 -m sglang.launch_server \
  --model-path Qwen/Qwen3.5-4B \
  --lora-paths octopus-v45=Tommysha/Octopus-Qwen3.5-4B \
  --max-lora-rank 64 \
  --lora-target-modules all \
  --disable-radix-cache false

Run the Octopus evaluation harness (see scripts/deploy.sh and README.md):
```
bash scripts/deploy.sh /data/octopus check
```

Citation

@article{sha2026octopus,
  title   = {Octopus: One Brain, Many Branches for Latent-KV Multi-Agent Reinforcement Learning},
  author  = {Sha, Tianming and Li, Junnan and Zhao, Siqi},
  year    = {2026},
  url     = {https://shatianming5.github.io/octopus/}
}