Agentic Reasoning

June 22, 2026 · View on GitHub

Synthetic data generation and inference pipeline for deep-research / multi-hop QA tasks: BrowseComp-en, BrowseComp-zh, HotpotQA-MuSiQue, and xBench.

The agent uses a Hierarchical Memory System — search, visit, expand, fold, and commit tools — to conduct multi-step web research while keeping the context window manageable.

Models and Data

Type	Name	Hugging Face
Model	LightThinker-Plus-Qwen3-30B-A3B	zjunlp/LightThinker-Plus-Qwen3-30B-A3B
SFT Data	Agentic Reasoning	zjunlp/LightThinker-Plus-AgenticReasoning-SFT

Environment Setup

Install dependencies:

bash setup.sh

API Keys

Two external services are required for web search and page reading:

Service	Purpose	Sign-up
Serper	Web search & Google Scholar	`SERPER_KEY_ID`
Jina AI	Webpage content parsing	`JINA_API_KEYS`

Configuration

The pipeline is controlled entirely through environment variables loaded from a .env file. Two commit-safe template files are provided — copy the appropriate one before running:

# For synthetic data generation:
cp .env.synthetic.example .env

# For inference on a trained model:
cp .env.infer.example .env

Then open .env and fill in your credentials.

Required variables

Variable	Description
`SERPER_KEY_ID`	Serper API key
`JINA_API_KEYS`	Jina AI API key
`REASONING_SK`	API key for the main reasoning model
`REASONING_ENDPOINT`	OpenAI-compatible base URL for the main model
`MODEL_NAME`	Model name as served by the endpoint

Optional variables (with defaults)

Variable	Default	Description
`MAX_ROUNDS`	`100`	Maximum agent turns per sample
`MAX_REASON_TOKENS`	`110592`	Token budget before forcing a final answer
`LOGGER_MODEL_NAME`	`""`	Logger / summarizer model name
`LOGGER_MODEL_URL`	`""`	Logger model API endpoint
`LOGGER_MODEL_API_KEY`	`""`	Logger model API key
`EVALUATE_MODEL_NAME`	`""`	Judge model for evaluation
`EVALUATE_API_BASE`	`""`	Judge model API endpoint
`EVALUATE_API_KEY`	`""`	Judge model API key
`TOKENIZER_PATH`	`./core/tasks/tokenizer/Qwen3-30B-A3B-Thinking-2507`	HuggingFace tokenizer for token counting (falls back to tiktoken if unavailable)

Running Synthetic Data Generation

Edit the user-config section at the top of scripts/synthetic.sh, then:

bash scripts/synthetic.sh

To override the model or run multiple iterations:

bash scripts/synthetic.sh \
    --ports 8210 \
    --model_name your-model-name \
    --max_workers 10 \
    --iterations 3

Key script variables:

Variable	Description
`EXPERIMENT_NAME`	Output directory name under `results-synthetic/`
`TASK_NAME`	Dataset task: `DeepResearch`, `BC_en`, `BC_zh`, `xbench`
`DATASET_PATH`	Path to the input JSONL dataset
`MAIN_SYS_PROMPT`	Key into `PROMPTS_MAP` for the main agent
`SEARCH_LOGGER_SYS_PROMPT`	Key for the search-logger agent
`VISIT_LOGGER_SYS_PROMPT`	Key for the visit-logger agent
`REACT_MODE`	`fn_call` (recommended) or `prompt`

Output is written to results-synthetic/<EXPERIMENT_NAME>/iter<N>_raw_outputs.jsonl.

Running Inference

Edit the model settings in .env, then:

bash scripts/inference.sh

The inference script uses a single system_prompt (no dual-model logger):

bash scripts/inference.sh \
    --ports 8210 \
    --model_name your-model-name \
    --max_workers 10

Output is written to results-inference/<EXPERIMENT_NAME>/iter<N>_raw_outputs.jsonl.

Data Post-Processing

Convert raw interaction trajectories into SFT training data:

python process/data_filter.py \
    --input results-synthetic/my-run/iter1_raw_outputs.jsonl \
    --output curated/train_data.json

Multiple input files can be provided:

python process/data_filter.py \
    --input results-synthetic/run1/iter1_raw_outputs.jsonl \
              results-synthetic/run2/iter1_raw_outputs.jsonl \
    --output curated/train_data.json

The script applies the following quality filters:

Short simple trajectories (< 3 rounds with no memory tools) are dropped
Memory logic validation: trajectories using expand/fold/commit must satisfy expand → fold ordering with no duplicates
Length filter: trajectories exceeding 60 rounds are dropped

Output report columns:

Column	Description
`Memory Kept`	Valid trajectories using expand/fold/commit
`Simple Kept`	Valid simple trajectories (no memory tools)
`Drop (No EF Mix)`	Missing required expand+fold+commit combination
`Drop (Fold Error)`	fold called without a preceding expand
`Drop (Jitter)`	Consecutive duplicate memory operations
`Drop (Too Long)`	Exceeded 60-round threshold

Direct Usage (Python)

# fn_call mode (recommended for synthetic data generation)
python main.py \
    --task_name DeepResearch \
    --model_name your-model \
    --base_url http://localhost:8000/v1 \
    --dataset_path datasets/browsecomp_en.jsonl \
    --toolsets DEEPRESEARCH \
    --react_mode fn_call \
    --main_system_prompt synthetic_main_agent \
    --search_logger_system_prompt search_logger \
    --visit_logger_system_prompt visit_logger \
    --output_dir results/ \
    --output_base_name iter1 \
    --max_workers 10

# prompt mode (for inference with a single system prompt)
python main.py \
    --task_name DeepResearch \
    --model_name your-model \
    --base_url http://localhost:8000/v1 \
    --dataset_path datasets/browsecomp_en.jsonl \
    --toolsets DEEPRESEARCH \
    --react_mode prompt \
    --system_prompt infer_agent \
    --output_dir results/ \
    --output_base_name iter1

Supported --task_name values: DeepResearch, BC_en, BC_zh, xbench

Supported --toolsets values: DEEPRESEARCH, DEEPRESEARCH_BASE

Supported Datasets

Place dataset files under datasets/. Expected JSONL format — one record per line with at minimum id, question, and answer fields.