SR²AM: Efficient Agentic Reasoning Through Self-Regulated Simulative Planning

We argue that efficient agentic reasoning benefits from decomposing deliberation into three interacting systems: reactive execution (System I) for fine-grained reasoning and direct action; simulative reasoning (System II) that predicts consequences of proposed actions through a world model, providing a unified planning mechanism across diverse tasks; and self-regulation (System III) that decides when and how deeply to plan through a learned configurator.

SR²AM (Self-Regulated Simulative Reasoning Agentic LLM) is our instantiation of this decomposition: the configurator and simulative planner are realized as distinct stages within an LLM's chain-of-thought reasoning, with the LLM itself serving as the world model in language space. By separating self-regulation, planning, and execution while preserving the expressiveness of free-form reasoning, SR²AM learns to plan further ahead rather than simply reason more, achieving competitive task performance with substantially fewer reasoning tokens.

For more details, visit our project website or read the paper.

We release two models:

• SR²AM-v0.1-8B: Based on Qwen3-8B, competitive with 120--355B systems
• SR²AM-v1.0-30B: Based on Qwen3-30B-A3B-Thinking-2507, competitive with 685B--1T systems while consuming 25--95% fewer reasoning tokens than comparably sized agentic LLMs

SR²AM-v0.1-8B and SR²AM-v1.0-30B sit above the size-vs-accuracy trendline in (a), and SR²AM-v1.0-30B is on the Pareto frontier of reasoning-token efficiency vs. accuracy among 30/32B agentic models in (b). The full benchmark breakdown is in the paper.

This guide walks you through running SR²AM on your own dataset.

# Option A: uv (recommended, faster)
uv venv .venv --python 3.13
source .venv/bin/activate
uv pip install --no-deps -r requirements.txt

# Option B: pip
pip install -r requirements.txt

# Configure API keys
cp .env.example .env
# Edit .env with your API keys (see Prerequisites below)
source .env

SR²AM uses three external tools during inference. Configure them in your .env file:

Search provider: SerpAPI is the default. To use Serper.dev instead, set SERPER_DEV_API_KEY in your .env and pass --extra-args "--search_provider serper_dev" to the inference script.

Evaluation scoring: Test benchmark evaluation (AIME, MATH500, GPQA, BrowseComp, GAIA, etc.) uses OPENAI_API_KEY for LLM-based scoring.

Browsing summarizer: Any OpenAI-compatible instruct LLM endpoint works (e.g., Qwen3-30B-A3B-Instruct-2507). This model summarizes webpage content for the agent's visit_tool.

Code sandbox: Use scripts/sandbox/setup_sandbox.sh for automated setup (pulls the SandboxFusion image, installs required packages, and verifies). See Code Sandbox setup below for details.

# SR2AM-v0.1-8B (requires ~16GB VRAM)
huggingface-cli download sailing-lab/SR2AM-v0.1-8B --local-dir ./models/SR2AM-v0.1-8B

# SR2AM-v1.0-30B (requires ~4x GPUs with TP=4)
huggingface-cli download sailing-lab/SR2AM-v1.0-30B --local-dir ./models/SR2AM-v1.0-30B

Input files are JSONL format. Each line must contain:

{"id": "aime24-0", "question": "Every morning ...", "reward_model": {"ground_truth": "xxx", "style": "rule"}, "data_source": "math__aime_repeated_8x", "general_domain": "math__aime24"}
{"id": "gpqa_diamond-0", "question": "Which of ...", "reward_model": {"ground_truth": "xxx", "style": "rule"}, "data_source": "stem__gpqa_diamond_198", "general_domain": "stem__gpqa_diamond"}
{"id": "gaia-0", "question": "How many ...", "answer": "xxx", "general_domain": "web__gaia"}

The general_domain field is used as the dataset category in results (shown as dataset in output). Results are grouped by this field when computing per-dataset pass rates.

Following the Guru paper, all questions are appended with the instruction " You should provide your final answer in the format \boxed{YOUR_ANSWER}." to standardize the answer format for evaluation.

source .env  # Load API keys

# SR2AM-v0.1-8B on 8 GPUs (TP=1, DP=8)
bash scripts/run_inference_local.sh \
  --model-path ./models/SR2AM-v0.1-8B \
  --model-name SR2AM-v0.1-8B \
  --model-size 8b \
  --input-file data/test_questions.jsonl \
  --output-file outputs/sr2am-v0.1-8b-results.jsonl \
  --browsing-summarize-model Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 \
  --browsing-summarize-url http://SUMMARIZER_HOST:30000/v1 \
  --code-sandbox-servers "SANDBOX_HOST1 SANDBOX_HOST2" \
  --evaluate

# SR2AM-v1.0-30B on 8 GPUs (TP=4, DP=2)
bash scripts/run_inference_local.sh \
  --model-path ./models/SR2AM-v1.0-30B \
  --model-name SR2AM-v1.0-30B \
  --model-size 30b \
  --input-file data/test_questions.jsonl \
  --output-file outputs/sr2am-v1.0-30b-results.jsonl \
  --browsing-summarize-model Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 \
  --browsing-summarize-url http://SUMMARIZER_HOST:30000/v1 \
  --code-sandbox-servers "SANDBOX_HOST1 SANDBOX_HOST2" \
  --evaluate

# SR2AM-v1.0-30B on 4 GPUs (TP=4, DP=1)
bash scripts/run_inference_local.sh \
  --model-path ./models/SR2AM-v1.0-30B \
  --model-name SR2AM-v1.0-30B \
  --model-size 30b \
  --num-gpus 4 \
  --input-file data/test_questions.jsonl \
  --output-file outputs/sr2am-v1.0-30b-results.jsonl \
  --browsing-summarize-model Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 \
  --browsing-summarize-url http://SUMMARIZER_HOST:30000/v1 \
  --code-sandbox-servers "SANDBOX_HOST1 SANDBOX_HOST2"

The script automatically starts an SGLang server, waits for it to be ready, runs inference, and cleans up.

Run bash scripts/run_inference_local.sh --help for all options. See Paper Reproduction below for the full settings used in the paper.

# 8B model (1 node, 8 GPUs)
sbatch scripts/run_inference_slurm_8b.sh \
  ~/models/SR2AM-v0.1-8B \
  SR2AM-v0.1-8B \
  ~/data/test_questions.jsonl \
  sr2am-v0.1-8b-results.jsonl \
  Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 \
  http://SUMMARIZER_HOST:30000/v1 \
  64 1 \
  "--fix_datetime --filtering --remove_tags --no_break_early --agent_type think --max_turns 50 --max_completion_tokens 16384 --temperature 0.8 --code_sandbox_servers SANDBOX_HOST1 SANDBOX_HOST2"

# 30B model (2 nodes, 16 GPUs)
sbatch scripts/run_inference_slurm_30b.sh \
  ~/models/SR2AM-v1.0-30B \
  SR2AM-v1.0-30B \
  ~/data/test_questions.jsonl \
  sr2am-v1.0-30b-results.jsonl \
  Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 \
  http://SUMMARIZER_HOST:30000/v1 \
  64 1 \
  "--fix_datetime --filtering --remove_tags --no_break_early --agent_type think --max_turns 100 --max_completion_tokens 16384 --temperature 1.0 --code_sandbox_servers SANDBOX_HOST1 SANDBOX_HOST2"

The inference scripts run with --filtering by default, which scores each response against the ground truth during inference. The output JSONL contains a correct field per response, so no separate judging step is needed.

# Print pass rate and pass@k by dataset
python evaluation/compute_rep_results.py \
  --input_file outputs/sr2am-v1.0-30b-results.jsonl \
  --num_reps 1

# Count reasoning tokens per trajectory
python avg_tokens_before_think_close.py \
  --input outputs/sr2am-v1.0-30b-results.jsonl \
  --model ./models/SR2AM-v1.0-30B \
  --breakdown-by-dataset

# For 8B (uses --assistant-text-source configurator)
python avg_tokens_before_think_close.py \
  --input outputs/sr2am-v0.1-8b-results.jsonl \
  --model ./models/SR2AM-v0.1-8B \
  --assistant-text-source configurator \
  --breakdown-by-dataset

# Re-judge answers against ground truth
python evaluation/run_judge_results.py \
  --system direct-api \
  --dataset_path data/test_questions.jsonl \
  --answers_path outputs/sr2am-v1.0-30b-results.jsonl \
  --output_name sr2am-v1.0-30b \
  --remove_tags

# Analyze judge results by data source
python evaluation/analyze_judge_results.py \
  --paths evaluation/judged_sr2am-v1.0-30b.json

SR²AM uses three external tools during inference:

Queries web-scale information via SerpAPI (default) or Serper.dev. Supports multiple simultaneous queries. The default provider is SerpAPI (SERPAPI_API_KEY). To use Serper.dev instead, pass --search_provider serper_dev and set SERPER_DEV_API_KEY.

Crawls and summarizes webpage content using an LLM summarizer. Truncates pages to 28K tokens. Requires an OpenAI-compatible LLM endpoint for summarization (e.g., Qwen3-30B-A3B-Instruct-2507).

Stateless Python sandbox via SandboxFusion. The sandbox must have common scientific and domain-specific Python packages pre-installed (see scripts/sandbox/sandbox-requirements.txt for the full list).

Requires Enroot:

# 1. Pull image, start server, install all packages, and export a reusable .sqsh
bash scripts/sandbox/setup_sandbox.sh --export sr2am_sandbox.sqsh

# 2. Verify the sandbox has all required packages
bash scripts/sandbox/verify_sandbox.sh --host localhost --port 8080

# 3. Deploy with automatic restart (for long-running inference)
bash scripts/sandbox/run_sandbox_retry.sh -i sr2am_sandbox.sqsh

Or install into a remote server that's already running:

bash scripts/sandbox/setup_sandbox.sh --skip-pull --host SANDBOX_HOST

Configure with --code-sandbox-servers "HOST1 HOST2" in the inference script. Each server listens on port 8080.

All paper results use SerpAPI (the default search provider) and a fixed system prompt datetime (Sun Aug 31 2025 23:34:17). Pass --fix_datetime via --extra-args to enable.

# Point at your local copies of the three sources (defaults shown)
export SR2AM_GURU_EVAL_ROOT=~/guru-RL-92k/offline_eval
export SR2AM_AFM_BENCHMARKS_ROOT=~/Agent_Foundation_Models/AFM/data/web_agent/test_benchmarks
export SR2AM_XBENCH_ROOT=~/xbench-evals/data

python evaluation/prepare_test_data.py create_test_dataset_full \
  --output_file data/sr2am_test_full.jsonl

source .env  # Load API keys (requires SERPAPI_API_KEY and OPENAI_API_KEY)

# SR2AM-v1.0-30B (8 GPUs, TP=4, DP=2)
bash scripts/run_inference_local.sh \
  --model-path ./models/SR2AM-v1.0-30B \
  --model-name SR2AM-v1.0-30B \
  --model-size 30b \
  --input-file data/sr2am_test_full.jsonl \
  --output-file outputs/sr2am-v1.0-30b-results.jsonl \
  --browsing-summarize-model Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 \
  --browsing-summarize-url http://SUMMARIZER_HOST:30000/v1 \
  --code-sandbox-servers "SANDBOX_HOST1 SANDBOX_HOST2" \
  --extra-args "--fix_datetime --code_concurrency 128 --visit_concurrency 64 --search_concurrency 128 --shuffle_questions" \
  --evaluate

# SR2AM-v0.1-8B (8 GPUs, TP=1, DP=8)
bash scripts/run_inference_local.sh \
  --model-path ./models/SR2AM-v0.1-8B \
  --model-name SR2AM-v0.1-8B \
  --model-size 8b \
  --input-file data/sr2am_test_full.jsonl \
  --output-file outputs/sr2am-v0.1-8b-results.jsonl \
  --browsing-summarize-model Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 \
  --browsing-summarize-url http://SUMMARIZER_HOST:30000/v1 \
  --code-sandbox-servers "SANDBOX_HOST1 SANDBOX_HOST2" \
  --extra-args "--fix_datetime --code_concurrency 64 --visit_concurrency 32 --search_concurrency 64 --shuffle_questions" \
  --evaluate

Dependency conflicts during install: requirements.txt is a frozen environment snapshot. If you see resolution errors, use uv pip install --no-deps -r requirements.txt or pip install --no-deps -r requirements.txt to install without resolution.

outlines-core build fails (Rust compiler): This package is an optional transitive dependency of SGLang. Remove it from requirements.txt if you don't have a Rust toolchain -- inference works without it.

SGLang server won't start: Check that your CUDA version matches the installed torch and sglang versions. The 30B model requires at least 4 GPUs (TP=4). Check logs/sglang_server.log for details.

wandb login: run_agent.py logs to Weights & Biases. Run wandb login first, or set WANDB_MODE=disabled to skip.

Search provider: SerpAPI is the default. To use Serper.dev, set SERPER_DEV_API_KEY in .env and pass --search_provider serper_dev via --extra-args (local script) or directly (when calling run_agent.py).

--extra-args or --code-sandbox-servers ignored: These flags take a single quoted string that is word-split when passed to run_agent.py. Make sure the value is in quotes: --extra-args "--search_provider serper_dev --code_concurrency 128". If quoting is difficult (e.g., nested shell layers), call run_agent.py directly instead -- see run_inference_local.sh for the full argument list.

If you find SR²AM useful in your research, please cite our paper:

@article{deng2026sr2am,
  title={Efficient Agentic Reasoning Through Self-Regulated Simulative Planning},
  author={Deng, Mingkai and Hou, Jinyu and Neves, Lara Sá and
          Pimpalkhute, Varad and Killian, Taylor W. and
          Liu, Zhengzhong and Xing, Eric P.},
  journal={arXiv preprint arXiv:2605.22138},
  year={2026}
}

We thank the open-source community for the base models, training frameworks, serving engines, evaluation benchmarks, and tools that made this work possible.

This project is released under the Apache License 2.0.