Source-linked AI summary
Taking Shortcuts for Categorical VQA Using Super Neurons
Pierre Musacchio, Jaeyi Jeong, Dahun Kim, Jaesik Park
TL;DR
Prior VLM probing emphasizes aggregated, macro-level representations, leaving open whether individual scalar activations can answer visual questions accurately. This paper turns raw LLM activations into Super Neuron classifiers and reports that they outperform base models across diverse VQA benchmarks, with up to 5.10x speedup while maintaining model-level performance.
Problem
Prior VLM probing emphasizes aggregated, macro-level representations, leaving whether individual scalar activations can answer visual questions accurately as an open question.
Source: §1 · p. 1Method
The paper identifies Super Neurons by thresholding raw LLM activations and converting selected scalar values into categorical classification predictions.
Source: §1 · p. 2Results
SNs outperform base models across diverse VQA benchmarks and achieve up to 5.10x speedup while maintaining model-level performance.
Source: §1 · p. 2Takeaways & Limitations
SNs support extreme early exiting at the first generated token and first LLM layer, substantially reducing inference time.
Source: §5 · p. 14Whether SNs provide useful signals on complex open-ended prompts and reasoning, and whether robustness and early exiting hold there, remains undetermined.
Source: §5 · p. 14
Abstract
from arXiv · showhide
Sparse Attention Vectors (SAVs) have emerged as an excellent training-free alternative to supervised finetuning or low-rank adaptation to improve the performance of Vision Language Models (VLMs). At their heart, SAVs select a few accurate attention heads for a task of interest and use them as classifiers, rather than relying on the model's prediction. In a similar spirit, we find that directly probing the raw activations of the VLM, in the form of scalar values, is sufficient to yield accurate classifiers on diverse visually grounded downstream tasks. Shifting focus from attention vectors to scalar activations dramatically increases the search space for accurate parameters, allowing us to find more discriminative neurons immediately from the first generated token. We call such activations Super Neurons (SNs). In this probing setting, we discover that enough SNs appear in the shallower layers of the large language model to allow for extreme early exiting from the first layer of the model at the first generated token. Compared to the original network, SNs robustly improve the classification performance while achieving a speedup of up to 5.10x.
1 Introduction
The introduction proposes training-free probing of scalar LLM activations to identify Super Neurons (SNs), which act as categorical VQA classifiers and enable early exit. It frames this approach as a shift from macro-level representations to micro-level activations.
1 Introduction
- SNs outperform the base models on a diverse suite of unseen categorical VQA validation benchmarks.Thresholding raw activations yields accurate classifiers across a wide diversity of datasets.Source: §1 · p. 2
- The method shifts from macro-level representations to scalar activations, thresholding stored LLM activations from probing data into classification predictions.It is a training-free strategy inspired by, rather than clustering attention heads.Source: §1 · pp. 1–2
- The introduction says the study investigates SNs’ discriminative power, model location, quantity, and robustness, and introduces agreement rate to quantify divergence from model predictions.Source: §1 · p. 2
- Up to 5.10× speedup is achieved by exiting at the LLM’s first layer while generating the first token, with model-level performance maintained.The introduction presents extreme early exit as a byproduct of SNs in shallow layers.Source: §1 · p. 2
2 Related work
Prior VLM-efficiency work spans pruning, token reduction, encoder training, early exit, and sparse attention vectors, while this paper uses raw activations as categorical predictors.
Efficient VLMs
- Efficiency approaches include parameter pruning by distillation or weight-search policies, token reduction [14] [43], encoder training, and early exit [2].Sparse attention vectors [27] have also been used for training-free VLM classification; accurate prediction from selected weights had been shown for convnets and LSTMs [32], but not established for multimodal transformer tokens.Source: §2 · p. 3
- The method identifies expert neural activations for categorical VQA without changing model weights or relying on token similarity, then enables early exit as early as the first layer.The paper describes SNs as raw activations repurposed for categorical predictions.Source: §2 · pp. 3–4
- For LLaVA-v1.5-7b, SNs search 131,072 activation scalars versus 1,024 SAV attention vectors.SNs probe activations at the micro level rather than clustering attention heads, and are reported to be robust to prompt variations and distribution shifts.Source: §2 · p. 3, Figure 2 · p. 3
Explainable VLMs
- Interpretability work includes neuron concept tagging, studies of attention sinks and object-aligned attention maps, and sparse autoencoders identifying object-specific concepts.Linear probing suggests reasoning occurs at different, non-monolithic stages and.Source: §2 · p. 4
Efficient VLMs
- Raw activations can predict categorical answers, with expert neurons sometimes appearing by the first LLM layer at the first answer token; the paper introduces agreement rate to study SN–model disagreement.Robustness experiments suggest SNs generalize to new distributions and neighboring prompts rather than exploiting spurious input correlations.Source: §2 · p. 4
3 Method
The method searches a VLM’s layer activations for scalar Super Neurons (SNs) that can classify categorical VQA answers, then evaluates their agreement with the model. It defines how to gather, select, aggregate, and assess these neurons.
3.1 Preliminaries
- Inspired by, the method hypothesizes that individual neuron activations can encode answers well enough to match or outperform the full VLM.It defines a VLM as vision and text encoders feeding an LLM, and denotes each layer’s activation by H_l.Source: §3.1 · p. 5
3.2 Super Neurons
- SN discovery gathers probing examples, thresholds raw activations into binary predictions, selects neurons above a metric cutoff, and aggregates their validation predictions.The method can aggregate SN predictions by averaging or majority voting.Source: §3.2 · pp. 6–7
3.3 Agreement rate
- Agreement rate (AR) measures how often SN predictions match the VLM’s answers and can be reported for neurons above a specified SN accuracy threshold.For example, AR@0.8 covers SNs whose accuracy exceeds 0.8 on the set being evaluated.Source: §3.3 · p. 7
4 Experiments
Across seven categorical VQA datasets, probed Super Neurons (SNs) outperform or compete with their source models, and experiments show they can support early-exit inference. The evaluation also examines robustness, data requirements, and metric-specific optimization.
4.1 Datasets
- The experiments cover seven categorical VQA datasets spanning hallucination, occlusion, depth, visual understanding, geometry, general knowledge, and scientific reasoning.The datasets are Pope, InstaOrder (Occ.), InstaOrder (Depth), VizWiz, Clevr, A-OKVQA, and ScienceQa.Source: §4.1 · pp. 7–8
4.2 Evaluation protocol
- The evaluation uses balanced probing sets sampled from training data and reports accuracy, precision, recall, and F1 to account for class imbalance.Each set contains 3,000 samples except VizWiz; A-OKVQA and ScienceQa are converted into binary questions.Source: §4.2 · pp. 8, 10
4.3 Results
- SNs outperform or match the source models across the evaluated suite, including on occlusion understanding and the challenging A-OKVQA and ScienceQa tasks.The validation comparison covers LLaVA-v1.5-7b and Qwen3-VL-4b-Instruct; Qwen’s occlusion F1 with SNs also exceeds vanilla LLaVA’s.Source: Table 2 · p. 9, §4.3 · p. 12
- Accurate SNs appear even in shallow LLM layers, and surpassing the model’s accuracy requires them to disagree with its predictions more often.On Pope, agreement drops sharply once SNs exceed the model’s performance.Source: Figure 4 · p. 8, §4.3 · pp. 11–12
- SNs outperform or match n-shot and SAV baselines on accuracy and F1, and compare positively to SAVs on every reported metric.For the balanced VizWiz comparison, both methods use 40 probing samples; the authors note that SAVs’ high recall accompanies lower accuracy and answer bias.Source: Table 3 · p. 9, §4.3 · p. 12, Table 8 · p. 13
- SNs also emerge in LLaVA-v1.5-13b and Qwen3-VL-32b-Instruct, supporting their evaluation at larger model scales.The reported scaled-up evaluations use Pope and ScienceQa, respectively.Source: Table 4 · p. 10, §4.3 · p. 13
4.4 Ablation studies
- SNs remain competitive when probed on Pope and evaluated on Pope-Voc, and are robust to prompt changes; performance improves with more probing data, surpasses the base model above 100 samples, and degrades at 5K.The authors default to 3K samples and report that optimizing for a chosen metric strongly influences validation performance.Source: Table 7 · p. 12, §4.4 · pp. 13–14, Table 9 · p. 14
- 5.10× faster inference than the original model is achieved by stopping at the first layer while maintaining model-level performance.SNs perform best on the first generated token, enabling early exit before the autoregressive process.Source: §4.4 · p. 14
5 Conclusion
Super Neurons use scalar activations as robust categorical classifiers that compare favorably with base models and sparse attention vectors, while enabling early exiting and reducing VLM inference time. Their usefulness on complex open-ended prompts and reasoning remains to be established.
5 Conclusion
- Super Neurons provide robust categorical classifiers that compare favorably with base models and sparse attention vectors, with accurate shallow-layer scalars enabling first-token, first-layer early exiting and lower inference time.The conclusion presents these properties as improving runtime efficiency of VLMs.Source: §5 · p. 14
- The authors plan to apply the method to vision language action models, where accurate discrete action decisions could be made faster.Source: §5 · p. 14
Limitations
- Whether Super Neurons provide useful signals for complex open-ended prompts and reasoning, and retain robustness and early exiting in those contexts, remains an open research question.Source: §5 · p. 14
A Datasets and prompts
The section defines a varied VQA evaluation suite spanning hallucination, spatial reasoning, constrained visual inputs, synthetic questions, commonsense, and science. It also describes dataset filtering and prompt choices for probing and validation.
Pope
- Pope contains 2.9K COCO-based VQA questions about object presence or absence, designed to evaluate VLM visual hallucinations.Source: A Datasets and prompts · p. 19
InstaOrder
- InstaOrder tests instance-wise occlusion understanding and depth-order reasoning; prior work found LLaVA-v1.5-7b struggles with geometry-related prompts.Source: A Datasets and prompts · p. 19
VizWiz
- VizWiz, created by people with visual impairments, tests reasoning under constrained visual inputs; the authors retain only yes/no questions for probing and validation [10].Source: A Datasets and prompts · p. 19
Clevr
- Clevr tests spatial understanding, including counting and positional understanding, using only classification questions for probing and validation.Source: A Datasets and prompts · p. 19
A-OKVQA
- A-OKVQA is a four-choice multiple-choice dataset requiring commonsense reasoning grounded in the scene.Source: A Datasets and prompts · p. 19
ScienceQA
- ScienceQA covers 26 topics, 127 categories, and 379 skills; its multiple-choice data is converted to binary visual question answering for inference.Source: A Datasets and prompts · p. 20
Prompts
- Clevr and VizWiz use no prompt template after yes/no filtering, while other datasets use templates in Table 1; InstaOrder includes object locations only to disambiguate repeated categories.Table 1 defines variables for object category, bounding-box location, original question, and multiple-choice answers.Source: A Datasets and prompts · p. 20, Table 1 · p. 20
B Baselines
The section reports the experimental configurations for LLaVA-v1.5-7b and Qwen3-VL-4b-Instruct and compares Super Neurons with finetuned alternatives across dataset scales, finding a budget-dependent tradeoff.
B Baselines
- The experiments use the reported LLaVA-v1.5-7b and Qwen3-VL-4b-Instruct configurations without alteration unless otherwise specified.The configurations are reported in Tables 2a and 2b.Source: B Baselines · p. 21
- With a small annotated finetuning budget, Super Neurons outperform fully finetuned and LoRA-tuned LLaVA while remaining training-free; larger budgets favor full finetuning for a single task.Single-task finetuning can significantly degrade performance on other tasks, whereas the training-free approach preserves pretrained weights and general capabilities.Source: B Baselines · p. 21
C Profiling
On an NVIDIA A100 with LLaVA-v1.5-7b, Super Neurons (SNs) reduce answer-generation wall time by bypassing autoregressive decoding and Hugging Face post-processing, and remain faster when generation is capped at one token.
C Profiling
- 1.8× faster than the base model when maximum generation is capped at one token.The profiling benchmark used LLaVA-v1.5-7b from Hugging Face on an NVIDIA A100 while varying the maximum generated token count.Source: C Profiling · p. 21
- Most of SNs’ wall-time gain comes from bypassing autoregressive decoding and Hugging Face’s post-processing routine.SNs skip the network’s autoregressive process.Source: C Profiling · p. 21
- The profiling benchmark measured LLaVA-v1.5-7b on an NVIDIA A100 while varying the model’s maximum generated token count.Measurements were finer-grained than the main benchmark and are reported in Table 4.Source: C Profiling · p. 21
D Robustness
SNs remain robust to prompt variation and outperform the vanilla model on random comparative strings, but fail when image-prompt relevance is broken, suggesting their decisions rely on relevant input rather than spurious biases.
Prompt sensitivity
- On InstaOrder (Occ.), SNs show similar F1 scores when the prompt’s relational word is replaced with alternatives.The tested prompt template uses a relational word between the two compared instances.Source: D Robustness · p. 22
- On random comparative strings of 3–10 characters, SNs outperform the vanilla model, suggesting they encode general in-domain task knowledge.The authors attribute this to extracting SNs directly from a metric that evaluates the task of interest.Source: D Robustness · p. 22
Adversarial prompting
- SNs’ F1 score collapses when Pope validation images or prompts are shuffled, indicating that their decisions require image-prompt relevance rather than spurious biases.The adversarial evaluation uses LLaVA-v1.5-7b on Pope; the dataset is described as saturated and focused on class recognition.Source: D Robustness · pp. 22–23, Table 6 · p. 23
E Location of SNs
SNs are more numerous on datasets that challenge LLaVA-v1.5-7b, often appearing in shallower layers, and some SNs overlap across datasets, including a neuron that outperforms the model on four benchmarks.
Dataset-wise
- For Pope, few SNs outperform LLaVA-v1.5-7b because it already answers the dataset correctly, though the authors report SNs can improve performance while reducing runtime.The Pope explanation is presented as a hypothesis; the cited passage gives no numerical results for the performance or runtime changes.Source: E Location of SNs · p. 23
- On harder datasets, many more SNs emerge, often in shallower layers; the visualization includes only SNs that outperform LLaVA-v1.5-7b.The figure shows that few such SNs emerge on datasets the model already answers well.Source: Figure 1 · p. 23, E Location of SNs · p. 24
Cross-dataset
- Some neurons exceed the model’s performance across multiple datasets, suggesting shared useful information and possible links between different types of questions.The overlap heatmap considers neurons that outperform the model on all datasets indicated.Source: E Location of SNs · p. 24, Figure 2 · p. 25
- InstaOrder (Occ.) and Clevr share the most SNs, which the authors attribute to their shared focus on object-wise geometric understanding and LLaVA-v1.5-7b’s difficulty on both.The heatmap visualizes neurons that outperform the model on all indicated datasets.Source: E Location of SNs · p. 24, Figure 2 · p. 25
- A single neuron outperforms LLaVA-v1.5-7b on all four benchmarks, illustrating that sparse Super Neuron activations can generalize across diverse data.The authors also report cases of SNs surpassing the model on three datasets simultaneously.Source: E Location of SNs · p. 24, Figure 2 · p. 25