PRISMA agent checkpoints

Trained checkpoints for the three LLM agents of PRISMA, a schema-coordinated multi-agent framework for open-domain multi-hop question answering (Plan → Retrieve → Inspect → Solve).

• Code: github.com/Ameame1/PRISMA

Folder	Agent	Role
PRISMA-Planner-Qwen3-4B	Planner	Decomposes the question into sub-questions linked by [ANSWER_X] placeholders
PRISMA-Solver-Qwen3-4B	Solver	Answers each sub-question from the retrieved documents
PRISMA-Inspector-Qwen3-4B	Inspector	One model for both inspection phases: the Context Inspector after retrieval (none / subquestion / retrieval) and the Reasoning Inspector after solving (none / reasoning / extraction)

All three are full-parameter fine-tunes of Qwen3-4B-Instruct-2507, stored in bfloat16.

How these checkpoints were trained

• Planner. GRPO with MS-SWIFT, full-parameter, 8 rollouts per prompt, learning rate 1e-6, KL coefficient 0.04, 2 epochs, bf16. Rule-based rewards: format (1.0), qualifier preservation (0.5), entity boxing (0.5), disambiguation (0.5), decomposition logic (0.5), sub-question count (2.0). Full arguments in args.json.
• Solver. GRPO with MS-SWIFT from the base model, full-parameter, 8 rollouts per prompt, learning rate 3e-6, KL coefficient 0.04, 1 epoch over 25,000 prompts, bf16. Rewards: format, accuracy and relevance, equally weighted. Full arguments in args.json.
• Inspector. Supervised fine-tuning on 4,937 teacher-annotated inspection decisions (2 epochs), followed by 200 steps of on-policy GRPO with TRL using an asymmetric stage-match reward: a missed error is penalized most, a false alarm only mildly, and output format acts as a multiplicative gate. 8 rollouts per prompt, learning rate 3e-6. This setup differs from the Inspector training described in the paper's appendix.

The Inspector was trained with fp32 master weights and converted to bfloat16 for this release. The bf16 tensors are bit-identical to the weights used in our evaluations, which loaded the model in bf16.

Usage

Download the repository, then serve each agent with vLLM (OpenAI-compatible API):

hf download Ameame1002/PRISMA --local-dir PRISMA
vllm serve PRISMA/PRISMA-Planner-Qwen3-4B   --port 8001 --dtype bfloat16 --max-model-len 8192
vllm serve PRISMA/PRISMA-Solver-Qwen3-4B    --port 8002 --dtype bfloat16 --max-model-len 8192
vllm serve PRISMA/PRISMA-Inspector-Qwen3-4B --port 8003 --dtype bfloat16 --max-model-len 8192

Retrieval uses the 21M-passage DPR Wikipedia corpus (December 2018) with a three-stage cascade: BGE-M3 dense retrieval (top-100), dense–sparse hybrid re-scoring (top-30), and a bge-reranker-v2-m3 cross-encoder (top-10). The retrieval index is not included here.