ARIA is a fine-tuned version of DeepSeek-R1-Distill-Qwen-7B trained to produce concise reasoning traces while maintaining mathematical accuracy. The model learns to allocate thinking tokens proportional to problem difficulty rather than generating uniformly verbose traces.

This is the fully merged model (no adapter dependency required).

Results

GSM8K (200 samples, greedy decoding)

Model	Accuracy	Avg Think Tokens	RES Score
DeepSeek-R1-Distill-7B	76.0%	467.5	162.6
ARIA-7B	78.5%	203.7	385.4

Token reduction: 2.30x with +2.5% accuracy improvement.

MATH-500 Per-Level Breakdown (243 samples, greedy decoding)

Level	Base Acc	ARIA Acc	Base Tokens	ARIA Tokens	Reduction
L1	83.7%	83.7%	790	244	3.23x
L2	80.0%	78.0%	981	408	2.41x
L3	80.0%	74.0%	1299	694	1.87x
L4	74.0%	66.0%	1585	864	1.83x
L5	68.0%	64.0%	3045	2036	1.50x

The model compresses aggressively on easy problems and scales thinking naturally for harder ones. This adaptive behavior is emergent -- it was not explicitly trained.

RES Score = (accuracy / mean_think_tokens) * 1000. Higher is better.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Eunice-Labs/aria-7b-merged",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Eunice-Labs/aria-7b-merged")

SYSTEM_PROMPT = (
    "You are a math reasoning assistant. "
    "Think through problems carefully inside <think> tags, "
    "then provide a clean final answer."
)

problem = "What is the sum of the first 100 positive integers?"
prompt = f"{SYSTEM_PROMPT}<|User|>{problem}<|Assistant|>"

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
output = model.generate(
    **inputs,
    max_new_tokens=4096,
    do_sample=False,
    repetition_penalty=1.1
)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Training

ARIA is trained via supervised fine-tuning (SFT) on compressed reasoning traces from the OpenThoughts-114k-math dataset. Easy and medium difficulty traces were compressed ~6x using Gemini 2.0 Flash. Hard problems were left uncompressed.

Setting	Value
Base model	DeepSeek-R1-Distill-Qwen-7B
Method	SFT with QLoRA (Unsloth)
LoRA rank	64
LoRA alpha	128
Target modules	q, k, v, o, gate, up, down projections
Training examples	3,993
Hardware	RTX 3090 24GB
Training time	~1 hour

Evaluation Details

Full per-sample eval results are available at Eunice-Labs/aria-eval-results.

Accuracy is measured using math-verify for LaTeX equivalence with normalized string match as fallback. Think tokens are counted from the first <think> tag (or start of generation) to the last </think> tag.

Eunice-Labs/aria-7b-merged

ARIA-7B: Adaptive Reasoning through Difficulty-Graded Trace Compression