bojianzhou/mer2026-track3-lora

MER2026 Track 3 (MER-Prefer) — LoRA adapters

LoRA adapters for our 5th-place system on MER2026 Track 3, plus the full single-model zoo behind the paper's negative results.

• 📄 Paper: Parameter-Efficient Post-Training for Multimodal Emotional Preference Prediction. MRAC 2026 @ ACM Multimedia.
• 💻 Code & per-sample scores: (https://github.com/zhoubojian-stevenchow/mer2026-track3-reproduction)
• 🏆 Official result: 91.8240 (Stage 1) / 69.6388 (Stage 2), average 80.7314, 5th of 21

Every number in the paper is verifiable without these weights — the per-sample scores are in the GitHub repo and verify_paper.py recomputes 34 checks on CPU. The adapters are here for anyone who wants to regenerate those scores from video, or build on the models.

Adapters

subfolder	checkpoint	role	Holdout-99	Test
v3	350	base recipe, no subtitles	76.75	89.72
bt	240	Bradley–Terry pairwise objective	82.81	90.25
MT590	590	individual-emotion multi-task, ρ=0.5	89.90	89.46
r70	620	individual-emotion multi-task, ρ=0.7	85.83	89.19
all3	1170	three-task mixture	88.90	90.51
s220	220	seed variant, subtitle recipe	78.77	88.66
s480	480	+474 majority-voted pairs	86.86	89.46
vlora	270	+vision-tower LoRA	82.81	89.46
cons	230	swap-consistency regularizer	80.79	89.18
synth	920	+synthetic preference pairs	82.81	86.82
vl	390	cross-backbone probe (Qwen2.5-VL base)	81.74	86.82

vl adapts Qwen2.5-VL, not Qwen2.5-Omni. Every other adapter is over Qwen/Qwen2.5-Omni-7B (Thinker), LoRA r=16 α=32 on the language layers only — 40.4 M trainable parameters, 0.45 % of the backbone. The vision and audio towers are frozen, which the paper argues is why input-side diversification fails to decorrelate members.

Checkpoint indices are only meaningful at the training length in the paper (mixed dataset 13,573 examples, 3 epochs ≈ 2,547 optimizer steps at accumulation 16). all3 in particular must be checkpoint 1170: that is the checkpoint that produced the Stage-1 scores the slot-averaging result rests on. A different checkpoint is a different model and the gains do not transfer.

The submitted system

Both official submissions use the same recipe — four slots at equal weight, with two models averaged into one vote in the third slot:

SLOTS = [['v3'], ['bt'], ['MT590', 'all3'], ['r70']]

all3 as a separate fifth vote contributes exactly +0.0000, because it near-duplicates MT590 (soft-score correlation 0.969). Folded into MT590's vote it lowers that vote's variance without adding a decision direction, and gains on both boards: 91.5612 → 91.8240 (Stage 1), 69.6208 → 69.6388 (Stage 2). This is score-space averaging, not weight-space model souping.

Usage

from transformers import Qwen2_5OmniThinkerForConditionalGeneration
from peft import PeftModel

base = Qwen2_5OmniThinkerForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2.5-Omni-7B", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(base, "<user>/<repo>", subfolder="MT590")

Scoring is a single forward pass, not generation: append the shared "a" prefix so the next position is the decision token, take a binary softmax over the 1 / 2 token ids, then cancel position bias with a swapped second pass:

p_a1 = 0.5 * p_fwd + 0.5 * (1 - p_bwd)

Decision rule as submitted:

p_a1 > 0.5  -> a1
p_a1 < 0.5  -> a2
p_a1 = 0.5  -> a1 iff p_fwd >= 0.5

The tie branch is structural: when the forward and swapped passes agree exactly, p_a1 is exactly 0.5 by construction. See code/01_inference_engine.py in the GitHub repo for the exact implementation, including the MT590 exception (its submission left ties at a2).

Training data

Official MER2026 data only: EmoPrefer-Data (574 pairs), EmoPrefer-Data-V2 (2,096), the official subtitle file, and the Track-1 individual-emotion set (9,395 clips). Not redistributed here — obtain it from the MER2026 organizers.

Limitations

Single task, single backbone family, one competition. The decorrelation principle the paper argues for is an empirical regularity at this task and scale, with a proposed mechanism (frozen perception towers), not a theorem. See the Limitations section of the paper.

Citation

@inproceedings{zhou2026emopref,
  title     = {Parameter-Efficient Post-Training for Multimodal Emotional
               Preference Prediction: A Top-5 System, the Limits of Improving
               It, and a Holdout That Fails to Predict the Test},
  author    = {Zhou, Bojian},
  booktitle = {Proceedings of the Workshop on Multimodal, Robust, and Affective
               Computing (MRAC) at ACM Multimedia},
  year      = {2026}
}

Adapters inherit the licence of their base model (Qwen2.5-Omni / Qwen2.5-VL).