---
title: 'SAVE: Sparse Autoencoder-Driven Visual Enhancement'
url: https://www.emergentmind.com/topics/save
type: topic
---

# SAVE: Sparse Autoencoder-Driven Visual Enhancement

SAVE, short for **Sparse Autoencoder-Driven Visual Information Enhancement**, is a training-free framework for mitigating object hallucination in Multimodal Large Language Models (MLLMs). It addresses hallucination as a consequence of **language priors** and **visual information loss** by identifying Sparse Autoencoder (SAE) latent directions that are most indicative of grounded visual processing and then steering the model along those directions during inference. In the reported evaluations, SAVE outperforms state-of-the-art training-free methods on standard benchmarks, including a **10\%p improvement in CHAIR\_S** together with consistent gains on **POPE** and **MMHal-Bench** [2512.07730].

## 1. Problem setting and central idea

Object hallucination in MLLMs refers to the generation of object mentions that are not supported by the image. SAVE is motivated by two causes stated explicitly: **visual information loss**, where intermediate representations may under-attend or suppress crucial image features, and **language priors**, where the language model’s statistical tendencies dominate when visual evidence is weak [2512.07730].

Within this framing, hallucination is treated as a systematic failure mode rather than a random error. SAVE is designed to diagnose which internal latent directions encode genuine visual understanding and which correlate with hallucination, and then to selectively boost the former. The framework is therefore not based on additional supervised fine-tuning of the base MLLM, but on inference-time intervention in its internal hidden states [2512.07730].

The method has three core stages. First, an SAE is trained on hidden representations from chosen transformer layers. Second, a binary object-presence question-answering probe is used to identify the SAE latent feature most associated with correct visual grounding, termed the **visual understanding feature**. Third, the model is steered along that feature’s decoder direction during inference, with the goal of reinforcing grounded visual understanding and suppressing hallucinated object generation [2512.07730].

## 2. Sparse autoencoder formulation

SAVE builds on a **two-layer sparse autoencoder** inserted into chosen transformer layers of an MLLM. Given a hidden representation

$$
x \in \mathbb{R}^{T\times d_n},
$$

where $T$ is the number of tokens and $d_n$ the model hidden size, the SAE computes

$$
a(x) = \mathrm{TopK}[ \mathrm{ReLU}( W_{\mathrm{enc}} \cdot(x - b_{\mathrm{pre}}) + b_{\mathrm{enc}} ) ]
$$

and

$$
\mathrm{SAE}(x) = W_{\mathrm{dec}} \cdot a(x) + b_{\mathrm{dec}}.
$$

Here $a(x)\in\mathbb{R}^{T\times d_M}$ is a sparse activation in which only the top $k$ entries per token survive, with $d_M\gg d_n$, and each row of $W_{\mathrm{dec}}$ is interpreted as a latent **feature direction** [2512.07730].

The training objective is

$$
L_{\mathrm{SAE}}(x) = \| x - \mathrm{SAE}(x)\|_2^2  +  \lambda \| a(x)\|_0.
$$

The reconstruction term enforces fidelity to the original hidden representation, while the sparsity term encourages only a few latent features to activate per token. The hyperparameter $\lambda$ balances reconstruction fidelity against **monosemantic sparsity** [2512.07730].

The reported SAE training setup uses **$\sim1.2$ M image-caption pairs** such as ShareGPT4V, optimized with **Adam**, with **density scheduling of $\lambda$ from $0 \to$ value over initial steps**. For any activation $x$, the vector $a(x)$ constitutes a sparse code whose nonzero indices identify which latent features are active. The stated hyperparameters include **Top-k in SAE encoder: $k \approx 256$**, and **SAE sparsity $\lambda$: typically increased linearly to 5 over early steps** [2512.07730].

## 3. Identification of visual understanding features

To isolate SAE features that reflect genuine visual grounding, SAVE uses a **balanced set of 10 K yes/no queries**, comprising **5 K positive** cases in which the object is present and **5 K negative** cases in which the object is absent. The model is asked a binary question of the form **“Is there a … in this image?”**, and the response is labeled as correct or hallucinated [2512.07730].

Let $X_{\mathrm{correct}}$ and $X_{\mathrm{hallu}}$ denote the sets of SAE activations $a(x)$ associated with correct and hallucinated answers. For each feature index $j$, SAVE defines activation frequencies

$$
f_j^{\mathrm{correct}} = \frac{1}{|X_{\mathrm{correct}}|} \cdot \sum_{x\in X_{\mathrm{correct}}} \mathbf{1}[a_j(x)>0]
$$

and

$$
f_j^{\mathrm{hallu}}   = \frac{1}{|X_{\mathrm{hallu}}|}   \cdot \sum_{x\in X_{\mathrm{hallu}}}   \mathbf{1}[a_j(x)>0].
$$

The **separation score** is then

$$
s_j = f_j^{\mathrm{correct}} - f_j^{\mathrm{hallu}}.
$$

The feature with the largest positive $s_j$ is defined as the **visual understanding feature**, whereas reversing the difference identifies a top **hallucination feature** [2512.07730].

Empirically, the top visual feature achieves **AUROC$\approx0.93$** for separating correct from hallucinated activations. This suggests that a single SAE latent direction can carry a strong signal about whether the model’s answer is grounded in the image rather than driven by hallucination-related internal dynamics [2512.07730].

## 4. Steering mechanism and inference procedure

Once the most indicative visual feature $j^\*$ has been selected, SAVE performs **latent-space activation steering** by modifying the hidden representation along that feature’s decoder direction:

$$
x_{\mathrm{steered}} = x + \alpha W_{\mathrm{dec}}[j^\*,:].
$$

Here $\alpha$ is the **steering strength**. The intended effect is to boost the internal activation along the identified visual-understanding direction and thereby reinforce the model’s visual grounding [2512.07730].

The intervention is applied **during inference** and is explicitly described as simple. Steering is applied at **a single transformer layer**, with **layer 24 in LLaVA-1.6** given as an example. Only **query (input) tokens** are modified, while **generated tokens** are left unchanged. The steering strength is tuned per layer, with **typical values 3–15** and common choices **$\{3,5,10,15\}$** [2512.07730].

Layer selection is part of the method’s practical behavior. SAEs are trained at layers **8, 12, 16, 20, 24**, and the results indicate that steering in **early (8, 12)** or **late (24)** layers is most effective, while **mid-layers (16, 20)** are less so. The optimal steering strength increases with network depth, reported as **$\alpha=3$ at layer 8** and **$\alpha=15$ at layer 24** [2512.07730].

## 5. Benchmarks and quantitative results

SAVE is evaluated on three benchmark families. **CHAIR\_S** and **CHAIR\_I** measure the proportion of captions or object mentions that hallucinate, where lower is better. **POPE** evaluates binary presence questions on MSCOCO and reports **F1** and **accuracy** across **random**, **popular**, and **adversarial** splits. **MMHal-Bench** contains **96 image-question pairs spanning eight categories**, scored by **GPT-4**, and reports **average score $\uparrow$** and **hallucination rate $\downarrow$** [2512.07730].

On **LLaVA-1.6 (7B)**, the reported gains are as follows [2512.07730]:

| Benchmark | Baseline | SAVE |
|---|---:|---:|
| CHAIR\_S | 31.2% | 21.4% |
| CHAIR\_I | 7.9% | 5.4% |
| MMHal-Bench score | 2.88 | 3.12 |
| MMHal-Bench HalRate | 0.41 | 0.36 |

For POPE, the paper reports that **average F1/Acc improve by $\sim0.5$–1 pts** on LLaVA-1.6 (7B). The comparisons are made against **state-of-the-art, training-free methods** including **VCD, DeCO, Devils, VTI, VISTA** [2512.07730].

The method is also reported to generalize across architectures. Similar consistent gains appear on **LLaVA-NeXT (8B)** and **Qwen2-VL (7B)**. The paper describes this as evidence of both **robustness** and **generalizability** across multiple models and layers [2512.07730].

## 6. Mechanistic analysis, ablations, and stated limitations

SAVE is accompanied by analyses intended to explain why the steering intervention reduces hallucination. At the token level, under the same prefix such as **“In the distance, a…”**, the vanilla model’s penultimate layer can sharply increase the probability of a hallucinated token such as **“boat”**; SAVE prevents that spike and instead favors a grounded alternative such as **“mountain”** [2512.07730].

Attention analysis shows that SAVE **decreases text-token attention and increases image-token attention** when measuring average cross-attention from the final generated token back to query versus image tokens. The reported interpretation is that steering along visual understanding features reduces reliance on language priors and increases grounding in the visual input [2512.07730].

The ablations further constrain the claim space. Steering toward the **top-1 visual feature** outperforms steering toward **hallucination features**, **randomly chosen features**, and combinations of the **top-3** or **top-5** features. The paper also reports **balanced yes–no transition counts ($\approx37$ “yes $\to$ no” vs. $41$ “no $\to$ yes”)**, which is presented as evidence that SAVE does not merely bias the model toward a fixed answer. Qualitative examples on **CHAIR** and **adversarial POPE/MMHal-Bench** are described as showing visually accurate outputs relative to hallucinated baselines [2512.07730].

The released implementation includes `/save/sae_train.py` for SAE training, `/save/probe.py` for binary object-presence probing and activation collection, `/save/steer.py` for inference-time steering wrappers for **LLaVA** and **Qwen2-VL**, `/configs/` for YAML hyperparameter files, and `/experiments/` for reproducing **CHAIR**, **POPE**, and **MMHal-Bench** evaluations [2512.07730].

The paper also states potential limitations and future directions. The listed limitations are **reliance on a single-feature steering**, which may miss **multi-faceted visual concepts**, and the need to **train separate SAEs per layer**. Proposed future directions include extending feature identification to richer probes such as **counting** and **relations**, exploring **multi-direction steering strategies**, and integrating SAVE into **end-to-end fine-tuning** for additional gains [2512.07730].

Source: https://www.emergentmind.com/topics/save