Research map / Vision-language and multimodal / Multimodal LLMs

MERLIN: Building Low-SNR Robust Multimodal LLMs for Electromagnetic Signals

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds EM-134K (134k EM signal-text instruction pairs) and EM-Bench (4,200 QA pairs over 14 subtasks), and proposes MERLIN, a two-stage training scheme whose second stage uses feature- and logit-level knowledge distillation from a frozen high-SNR teacher to a low-SNR student for robustness.

Problem

Electromagnetic signal understanding with MLLMs lacks large paired signal-text data and standardized benchmarks, and encoder-LLM models degrade sharply in low-SNR conditions where features collapse into homogeneous embeddings.

Approach

Treat low-SNR robustness as a feature-level problem: distill from a frozen teacher fed the high-SNR version of the same signal into a student fed the low-SNR version, using a denoising subspace projection and logit KL distillation alongside the task loss.

Key results

• MERLIN reaches 78.27 average on EM-Bench perception and reasoning versus 72.94 for Qwen3-VL-4B-Instruct and 32.35 for the best proprietary baseline (Claude-4-Sonnet).
• Ablation: low-SNR accuracy rises from 59.7 (Stage-1) to 62.9 (Stage-2 SFT) and 65.1 with full MERLIN distillation.
• Overall ablation accuracy improves from 71.8 (Stage-1) to 78.6 (full MERLIN).

Surprising findings

• Adding more low-SNR training data during pre-training gives only marginal low-SNR gains, while linearly interpolating noisy embeddings toward clean ones recovers generative performance.
• Text-only frontier LLMs given textualized IQ sequences score near zero on reasoning (strategy generation) tasks.

Limitations

• Low-SNR training uses Gaussian noise injection, which may not reflect real-world interference.
• Inconsistencies in the paper: dataset sizes (4,200 vs 4,800 vs 48,000 QA pairs; 134k vs 100k) and sub-task counts (14 vs 16) differ across sections.
• Baselines receive textualized signals only, so comparisons do not isolate the signal encoder from the distillation method.

Experimental setup

Datasets: EM-134K EM-Bench

Benchmarks: EM-Bench

Metrics: accuracy Rouge-L BLEU

Baselines: GPT-5 Claude-4-Sonnet DeepSeek-v3.2-exp Qwen3-Next-80b-A3B Gemini-2.5-Pro Qwen3-VL-4B-Instruct Stage-1 pretrained model

Builds on: LLaVA EMind knowledge distillation

Closest papers by meaning

• EEE-Bench: A Comprehensive Multimodal Electrical And Electronics Engineering Benchmark (CVPR 2025 · similarity 0.85)
• R-Sparse: Rank-Aware Activation Sparsity for Efficient LLM Inference (ICLR 2025 · similarity 0.84)
• S$^{2}$FT: Efficient, Scalable and Generalizable LLM Fine-tuning by Structured Sparsity (NeurIPS 2024 · similarity 0.83)
• HaploVL: A Single-Transformer Baseline for Multi-Modal Understanding (ICML 2025 · similarity 0.83)
• LLMs can see and hear without any training (ICML 2025 · similarity 0.82)
• MIA-Bench: Towards Better Instruction Following Evaluation of Multimodal LLMs (ICLR 2025 · similarity 0.82)