LoopHard judged-loop rates across four inference settings
This is a mixed-precision NVIDIA ModelOpt deployment checkpoint for
Qwen3.6-35B-A3B-AntiLoop,
a narrow fine-tune intended to recover from pathological self-verification and
enumeration loops while preserving ordinary long-form reasoning.
The repository contains a complete deployment checkpoint; no PEFT adapter is
required at inference time. It retains the upstream multimodal architecture,
tokenizer, chat template, MTP weights, and 262,144-token native context length.
Training data
The exact 178 masked supervised targets used for the final AntiLoop training
round are published in the
Qwen3.6-35B-A3B-AntiLoop-SFT dataset.
The dataset preserves each loss_start_char boundary so the pathological loop
prefix remains conditioning context rather than a supervised target. It
intentionally excludes the separately generated KL-regularization anchors.
Training procedure
The AntiLoop adapter was trained on the 178 masked supervised targets using a
standard supervised fine-tuning procedure, but regularized via KL-loss on separately generated non-loop anchors from the base model on everyday prompts from tulu-3-sft-mixture.
Benchmark results
LoopHard
LoopHard is our held-out set of 285 enumeration prompts designed to elicit
futile recall, recounting, and self-verification loops. The primary metric is
judged loops: whether the model's reasoning trace remains stuck in a futile
cycle when generation ends.
Model
Judged loops
Loop rate
NVIDIA NVFP4
72 / 285
25.26%
AntiLoop NVFP4
10 / 285
3.51%
NVIDIA NVFP4 + presence_penalty=1.5
30 / 285
10.53%
AntiLoop NVFP4 + presence_penalty=1.5
1 / 285
0.35%
The matched presence_penalty=1.5 comparison converted all 30 control loops to
clean completions while introducing one different loop. Exact two-sided
McNemar p = 2.98e-8.
Generation used thinking mode, temperature=0.7, top_p=0.95, top_k=20, a
6,144-token completion limit, and concurrency 24. The two
presence_penalty=1.5 arms used the exact original and AntiLoop NVFP4
checkpoints with the same vLLM build and serving configuration: TP1, FP8 KV
cache, FlashInfer attention, Marlin NVFP4 MoE, and MTP speculative decoding
with three draft tokens.
LoopHard is judged by GLM-5.2 using a convergence-aware rubric: systematic
reasoning and verification that reaches a conclusion are not loops, and a trace
that notices its own circling and exits is classified as recovered. The
calibration set contained 42 manually labeled traces. Across three judge runs,
accuracy was 88.1%, 92.9%, and 95.2%; all three runs identified all 17 labeled
loops, with 2–5 false positives among the 25 non-loop traces.
The 285 prompts, metadata, and GLM-5.2 evaluation code are published in the
LoopHard dataset on
Hugging Face.
Capability preservation
The capability checks below compare the same round-2 AntiLoop adapter against
its FP8 reference model under a matched runtime-LoRA setup. These runs used the
default presence penalty and should not be interpreted as evaluations of the
exact mixed-precision artifact at presence_penalty=1.5.
GPQA Diamond
Model
Accuracy
Qwen3.6-35B-A3B official model card
86.0%
FP8 reference, our matched harness
167 / 198 (84.34%)
AntiLoop FP8, our matched harness
166 / 198 (83.84%)
The official-model-card number is included for context and was not produced by
our harness.
Our GPQA run used thinking mode, paired per-question seeds,
temperature=0.7, top_p=0.95, top_k=20, MTP3 speculative decoding, a
65,536-token reasoning budget, and 4,096 tokens of answer headroom. The difference was not significant.
The GSM8K run used the exact 1,319-example openai/gsm8kmain test split,
thinking mode, paired seeds, temperature=0.7, top_p=0.95, top_k=20, MTP3,
an 8,192-token reasoning budget and 1,024 tokens of answer headroom. The difference was not significant.
Taken together, the matched GPQA and GSM8K results show no material or
statistically detectable capability loss at these sample sizes. They do not
establish equivalence across other tasks, modalities, or sampling settings.
Usage
Use a recent vLLM build with ModelOpt mixed-precision support:
