GAC-Qwen3.5-4B

Noise-Aware Adaptive Mixing for Hybrid SFT–RL Post-Training

This is the public compact-backbone release of GAC, a closed-form, noise-aware controller that adaptively mixes supervised fine-tuning (SFT) and reinforcement learning (RL) signals during post-training.

The release is designed for researchers who want to study adaptive SFT–RL post-training with a contemporary, accessible model backbone. The algorithmic implementation, evaluation harness, and method description are available in the GAC repository.

GAC on a modern 4B backbone

We retrained GAC on Qwen3.5-4B to bring adaptive hybrid SFT–RL post-training to a modern, capable small-model backbone. This release follows progress in compact language models while making GAC more accessible for research, evaluation, and deployment with modest compute resources.

Model details

• Base model: Qwen3.5-4B-Base
• Post-training method: GAC (hybrid SFT–RL adaptive mixing)
• Parameter scale: approximately 4B language-model parameters
• Format: Hugging Face Transformers / safetensors
• Transformers compatibility: use Transformers 5.10.4 or newer with native Qwen3.5 support
• Modalities: the checkpoint retains the Qwen3.5 text-and-vision model interface; the reported evaluation is text-only reasoning and coding
• License: Apache-2.0, subject to the upstream Qwen3.5 license and notices

Intended use

This checkpoint is intended for research on adaptive hybrid SFT–RL post-training, reproducible checkpoint evaluation, and low-cost experimentation with compact multimodal language models. It is not a safety-certified system and should not be used for unsupervised high-stakes decisions.

Release contents

The release provides inference weights, model and tokenizer configuration, licensing notices, and an evaluation summary. The public GAC implementation and evaluator accompany the checkpoint to support research and checkpoint benchmarking.

Evaluation snapshot

The table below reports the compact-backbone evaluation snapshot for the base checkpoint and the GAC release checkpoint. Percentages are author-reported release metrics; a rerun with the public evaluator writes exact per-example counts and summaries.

Domain	Benchmark	Eval size (N)	Qwen3.5-4B Base	GAC release
Mathematics	AMC	83	19.3%	67.5%
Mathematics	AIME24	30	0.0%	26.7%
Mathematics	AIME25	30	3.3%	20.0%
Knowledge	MMLU-Pro	1,000	58.3%	74.2%
Science	GPQA-Diamond	198	29.3%	64.1%
Science	SciBench	692	11.4%	60.1%
Code	MBPP	500	67.6%	74.6%
Code	HumanEval	164	67.7%	81.1%
Logic	BBH Logical Deduction	750	86.3%	93.1%
Logic	BBH Object Counting	250	93.2%	92.8%
Logic	BBH Tracking	750	97.6%	95.3%
Logic	BBH average (macro)	3 slices / 1,750	92.4%	93.7%

The table compares the Qwen3.5-4B base checkpoint with the GAC release. Values describe a fixed evaluation snapshot. Multiple seeds and per-example outputs support more detailed statistical comparisons. BBH average is the unweighted mean of the three displayed BBH slice scores.

Reading the result

Relative to the Qwen3.5-4B base snapshot, the GAC release is higher on 9 of 11 task slices in this snapshot, with gains across math, knowledge, science, and code. BBH remains in a high-performance regime: the macro-average is higher, while Object Counting and Tracking are slightly below the base snapshot. This is descriptive evidence for transfer to a compact backbone, not a claim of universal improvement.

Evaluation protocol

For an apples-to-apples rerun, use the public evaluator with --n_samples 1 and record the model revision, software versions, dataset revision, and seed. The harness uses the following task-specific defaults:

Task family	Decoding	Scoring
AMC / AIME24 / AIME25	temperature 0.6, top-p 0.95, max 8,192 tokens	math-verify equivalence
MMLU-Pro / GPQA-Diamond	temperature 0.6, top-p 0.95, max 8,192 tokens	strict answer-letter match
SciBench	temperature 0.6, top-p 0.95, max 8,192 tokens	SymPy equivalence plus numeric fallback
MBPP / HumanEval	temperature 0.2, top-p 0.95, max 1,024 tokens	isolated pass@1 execution
BBH logical subsets	temperature 0.6, top-p 0.95, max 4,096 tokens	normalized exact match

Dataset IDs, splits, fixed-subset selection, and the required example counts are maintained in eval/README.md. The evaluator pins MMLU-Pro to revision b189ec765aa7ed75c8acfea42df31fdae71f97be with a public 1,000-ID manifest for reruns. GPQA-Diamond requires gated dataset access or a lawfully obtained, user-provided CSV through --gpqa_csv; GPQA data is not redistributed.

Quick start

Transformers

import torch
from transformers import AutoProcessor, AutoModelForImageTextToText

model_id = "YueLinHu/GAC-Qwen3.5-4B"  # replace the namespace for other mirrors

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "user", "content": [{"type": "text", "text": "Solve: 2 + 2 = ?"}]}
]
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
inputs = processor(text=[text], return_tensors="pt").to(model.device)
with torch.inference_mode():
    output_ids = model.generate(**inputs, max_new_tokens=256)
print(processor.batch_decode(output_ids, skip_special_tokens=True)[0])

Install a current Transformers release with native Qwen3.5 support before loading the checkpoint, for example pip install -U "transformers>=5.10.4". Use the official Qwen3.5 documentation for multimodal inputs. For text-only benchmark reproduction, follow the commands in the repository's eval/README.md.

Reproducibility

The algorithmic implementation, default controller configuration, unit tests, and evaluation harness are available at:

• Code: https://github.com/huyuelin/GAC
• Paper and method description: https://openreview.net/forum?id=VhBpT4iq60
• Evaluation harness: https://github.com/huyuelin/GAC/tree/main/eval

For scientifically comparable reporting, record the exact model revision, transformers/vLLM version, decoding configuration, dataset revision, random seed, pass@1 execution sandbox, and scoring parser. The values above are a fixed release snapshot rather than multi-seed confidence intervals. When n_samples is greater than one, the harness reports best-of-n for generated answers; use n_samples=1 for the pass@1-style snapshot shown above.

Limitations and responsible use

• Benchmark accuracy is not a guarantee of reliability, factuality, or safety in deployment.
• Code-generation outputs must be executed in an isolated sandbox.
• Reported values are a fixed evaluation snapshot; use repeated runs to estimate statistical uncertainty.
• Users are responsible for complying with the licenses and usage policies of Qwen3.5, the evaluation datasets, and any downstream application.

Citation

@inproceedings{hu2026gac,
  title     = {GAC: Noise-Aware Adaptive Mixing for Hybrid SFT-RL Post-Training},
  author    = {Hu, Yuelin and Liu, Wei and Yu, Zhenbo and Cheng, Zhengxue and Song, Li},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year      = {2026},
  address   = {Budapest, Hungary},
  url       = {https://openreview.net/forum?id=VhBpT4iq60}
}

Acknowledgements

This checkpoint builds on Qwen3.5 and the GAC post-training method. Please cite both the GAC paper and the upstream Qwen3.5 model when appropriate.