Pith. sign in

REVIEW 3 cited by

Textual Steering Vectors Can Improve Visual Understanding in Multimodal Large Language Models

T0 review · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Text-derived steering vectors, especially mean shift, improve spatial relation and counting accuracy in multimodal LLMs by up to 7.3% on CV-Bench and by larger margins on several out-of-distribution datasets.

desk verdict Text-only steering vectors can shift MLLM visual behavior — the color intervention is convincing — but the benchmark table needs error bars and a null-intervention control before 'consistent enhancement' is secure. read the letter →

arxiv 2505.14071 v1 pith:P3LPV2VN submitted 2025-05-20 cs.LG cs.CLcs.CV

Woody Haosheng Gan , Deqing Fu , Julian Asilis , Ollie Liu , Dani Yogatama , Vatsal Sharan , Robin Jia , Willie Neiswanger This is my paper · ORCID
classification cs.LGcs.CLcs.CV
keywords steeringaccuracylanguagelargemllmsmodelsmultimodalvectors
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Multimodal large language models (MLLMs) such as PaliGemma2 and Idefics3 are built by taking a text-only language model, such as Gemma2 or Llama-3.1, and adding an image encoder on top. The authors ask whether the text-only part still contains useful directions, called steering vectors, that can be added to the model's hidden states during inference to change how it processes images. They build these vectors from ordinary English sentences: for example, the word 'on' in 'The cat is on the table' marks a spatial relationship, and 'five' marks counting. Three vector construction methods are tested: sparse autoencoder features, the mean difference between anchor words and other words, and a linear probe. They also compare with a prompting baseline.

On the CV-Bench benchmark, the best configurations show gains of 7.3 percentage points on spatial relationship accuracy and 3.3 points on counting for PaliGemma2-3B, with mean shift the strongest method. On out-of-distribution sets such as What'sUp, BLINK, CLEVR, and Super-CLEVR, the average improvement over methods is 7.6 points for mean shift, and some CLEVR cells gain more than 20 points. A toy experiment shows that adding a color direction from the text backbone can push the model's reported color for a yellow-orange image toward red.

The results are not uniform: many cells in the main table show no gain or a small decline, and the strongest numbers are chosen after grid searching the layer and scale. The test sets are small and error bars are not reported for the main table. The method also depends on the assumption that text-space directions survive vision-language fine-tuning. Still, the evidence supports a real but uneven transfer effect, and the approach is cheap because no weights are updated.

Extended reading notes

Core claim

Abstract: "We find that text-derived steering consistently enhances multimodal accuracy across diverse MLLM architectures and visual tasks. In particular, mean shift boosts spatial relationship accuracy on CV-Bench by up to +7.3% and counting accuracy by up to +3.3%, outperforming prompting and exhibiting strong generalization to out-of-distribution datasets." The load-bearing version is that concept directions extracted from the text-only backbone remain effective in the multimodal model after vision-language fine-tuning, so a single vector per concept per layer improves visual reasoning without any parameter update.

Load-bearing premise

The method assumes that the semantic direction for a visual concept in the backbone LLM's activation space is preserved and still causally aligned in the MLLM after vision-language post-training, and that adding alpha times that vector to image or text token hidden states changes visual processing in the intended way without disrupting other computations. This is stated in Section 2 as "preserved semantics of the text backbone" and is the basis for extracting vectors from Gemma2/Llama-3.1 and applying them to PaliGemma2/Idefics3. If post-training rotates or suppresses these directions, the central claim fails. The toy color intervention and the CV-Bench gains are evidence, not proof, of this premise.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

No new physical or architectural entities are introduced. The main hidden cost is model selection: layer, scale, token type, and several design constants are chosen to maximize CV-Bench accuracy or small OOD validation subsets. The text concept directions themselves are derived from hand-curated sentences and pretrained SAEs, so they are not fitted to the benchmark labels.

free parameters (6)
  • Steering layer index ℓ* = e.g., L5 for PaliGemma2-3B MeanShift spatial; varies per model/method/task
    Selected by grid search maximizing CV-Bench grid-split accuracy (Section 5.1).
  • Steering strength α* = e.g., 1.0 for PaliGemma2-3B MeanShift spatial; search ranges {0.1..1.0} unnormalized, {10..60} normalized for SAE and…
    Same grid search; each model/method/task has its own optimum.
  • Token intervention type (image, text, both) = chosen per model/task; for OOD chosen on 25 to 50 point validation subsets
    Section 6.1 uses a small validation subset to pick token type before final OOD evaluation.
  • SAE top-n feature count n = not specified
    Algorithm 1 ranks 'top n' features but n is never given; this affects vector aggregation.
  • Sentence-anchor pair count K = 20
    Appendix A.1 sets K=20 with hand-curated anchor words such as 'on'.
  • PCA dimension d for linear probe = K/2 = 10
    Section 4.4 fixes d=K/2 before training the probe.
assumptions (4)
  • domain assumption Semantic directions in the text-only backbone remain meaningful after vision-language post-training
    The whole method relies on this transfer; stated in Section 2 as 'preserved semantics of the text backbone' and used in Section 4. The toy color experiment supports but does not prove it.
  • domain assumption Adding a constant vector to residual stream activations h' = h + alpha*v preserves model coherence without renormalization or distributional correction
    All interventions in Section 5 use this linear addition with no other adjustment; this is a strong assumption about activation space.
  • domain assumption Pretrained SAE features and the o3-mini verification filter correctly isolate the intended visual concept
    Algorithm 1 uses pretrained SAEs and an LLM judge to select features; no human verification or exact acceptance threshold is described.
  • domain assumption The CV-Bench grid-search split is a valid proxy for choosing hyperparameters that transfer to out-of-distribution sets
    Section 6.1 uses CV-Bench optimal layer and alpha for OOD; if this transfer fails, OOD gains are not attributable to the method.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Textual Steering Vectors Can Improve Visual Understanding in Multimodal Large Language Models." pith.science (2026). https://pith.science/paper/P3LPV2VN

@misc{pith2026250514071,
  author       = {Pith},
  title        = {Pith review of: Textual Steering Vectors Can Improve Visual Understanding in Multimodal Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P3LPV2VN}},
  note         = {Machine review of arXiv:2505.14071}
}
read the original abstract

Steering methods have emerged as effective and targeted tools for guiding large language models' (LLMs) behavior without modifying their parameters. Multimodal large language models (MLLMs), however, do not currently enjoy the same suite of techniques, due in part to their recency and architectural diversity. Inspired by this gap, we investigate whether MLLMs can be steered using vectors derived from their text-only LLM backbone, via sparse autoencoders (SAEs), mean shift, and linear probing. We find that text-derived steering consistently enhances multimodal accuracy across diverse MLLM architectures and visual tasks. In particular, mean shift boosts spatial relationship accuracy on CV-Bench by up to +7.3% and counting accuracy by up to +3.3%, outperforming prompting and exhibiting strong generalization to out-of-distribution datasets. These results highlight textual steering vectors as a powerful, efficient mechanism for enhancing grounding in MLLMs with minimal additional data collection and computational overhead.

Figures

Figures reproduced from arXiv: 2505.14071 by the authors.

Figure 1
Figure 1. Overview of our steering methodology. For an MLLM with a text-only LLM backbone and a given [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Effect of steering strength on color token probabilities. To demonstrate that textual representations can effectively inter￾vene in visual understanding, we conduct a simple color percep￾tion experiment using GemmaScope [Lieberum et al., 2024a] for Gemma-2-9B for feature extraction and PaliGemma2-10B-mix￾448 [Beyer et al., 2024] as our target model. We present the model with a yellow-orange image (whose RGB hex code… view at source ↗
Figure 3
Figure 3. Left: Number of SAE features associated with each taxonomy (counting, spatial relationship, entity, and attribute) across the layers of Llama-3.1-8B, Gemma2-2B, and Gemma2-9B. Notably, SAE features for such visual concepts are sparse, numbering fewer than 10 across 16k total SAE features (Gemma2-2B/9B) or 32k features (Llama￾3.1-8B). Right: Examples of features corresponding to visual concepts, identified by the lay… view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Left: Depiction of mean shift method for the counting feature for Gemma-2-2B. The mean shift direction points from the mean hidden state of irrelevant tokens to the mean hidden state of relevant tokens (i.e., counting￾related tokens). Activations are projected to two d…
Figure 5
Figure 5. Figure 5: Efficient Grid Search with PaliGemma2-3B on the Spatial Relationship Task. Counting Spatial Relationship Entity Attribute Count Relation Distance Depth +0.7% (L7@0.8) +1.3% (L14@1) +0.0% (L9@0.8) +0.0% (L16@0.6) +3.3% (L9@1) +7.3% (L5@1) +0.0% (L9@0.6) +2.7% (L6@1) +1.…
Figure 6
Figure 6. Figure 6: Performance improvements on CV-Bench tasks when steering PaliGemma2-3B with MeanShift vectors. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Prompt template for querying GPT-o3-mini to verify whether a given feature is related to a visual taxonomy. For each taxonomy, the template employs a brief definition of the taxonomy, two example features that align with each taxonomy (for few-shot learning), and the t…
Figure 8
Figure 8. Figure 8: System and user prompt template for generating MLLM prompts. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Additional color perception intervention examples. In each case, we apply the normalized textual steering [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Example prompt for the CV-Bench Relation dataset. [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Example prompt for the CV-Bench Count dataset. [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Example prompt for the What’sUp-A dataset. [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]
Figure 13
Figure 13. Figure 13: Example prompt for the What’sUp-B dataset. [PITH_FULL_IMAGE:figures/full_fig_p021_13.png]
Figure 14
Figure 14. Figure 14: Example prompt for the BLINK Object Localization dataset. [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]
Figure 15
Figure 15. Figure 15: Example prompt for the CLEVR dataset. SUPER-CLEVR Image: [Model Prefix] answer en [Task Prefix] Answer the question by only responding the number. [Taxonomy Prefix] Prioritize counting objects and quantifying elements over other analysis. [Question] How many different…
Figure 16
Figure 16. Figure 16: Example prompt for the Super-CLEVR dataset. [PITH_FULL_IMAGE:figures/full_fig_p022_16.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Do Unified Multimodal Models Think in One Space? A Lens Through Cross-Branch Steering

    cs.CV 2026-07 conditional novelty 7.0 of 10

    Steering vectors from the understanding branch can control image generation, but vectors from the generation branch cannot control understanding, showing UMMs are architecturally unified but representationally asymmetric.

  2. VISOR++: Universal Visual Inputs based Steering for Large Vision Language Models

    cs.CV 2025-09 conditional novelty 6.0 of 10

    A single adversarially optimized image can reproduce activation-steering behavior in multiple VLMs and partially transfer to unseen models.

  3. Resa: Transparent Reasoning Models via SAEs

    cs.CL 2025-06 conditional novelty 6.0 of 10

    SAE-Tuning, a sparse-autoencoder-guided SFT procedure, elicits RL-comparable reasoning in 1.5B models from CoT-free QA data at about $1 and 20 minutes of training.

Reference graph

Works this paper leans on

12 extracted references · 10 canonical work pages · cited by 3 Pith papers

  1. [1]

    Feature’s explanation: {feature_explanation}

    work page

  2. [2]

    {taxonomy}

    URL https://openreview.net/forum?id=I4e82CIDxv. Team OLMo, Pete Walsh, Luca Soldaini, Dirk Groeneveld, Kyle Lo, Shane Arora, Akshita Bhagia, Yuling Gu, Shengyi Huang, Matt Jordan, et al. 2 olmo 2 furious. arXiv preprint arXiv:2501.00656, 2024. OpenAI. o3-mini. https://openai.com/index/openai-o3-mini/ , 2025. Accessed: 2025-05-13. Nina Panickssery, Nick Ga...

    arXiv 2024

  3. [3]

    {activation_example_3}

    work page

  4. [4]

    Top activation examples (tokens wrapped in <top>...</top> have the highest activation values and are the most important to focus on):

    work page

  5. [5]

    {activation_example_1}

    work page

  6. [6]

    {activation_example_2}

    work page

  7. [8]

    {activation_example_4}

    work page

  8. [9]

    {activation_example_5} Examples of features that DO align with the {taxonomy} taxonomy (notice how the key words are highlighted with <top>...</top> tags): Example 1: - Explanation: {explanation_1} - Activations: {activations_1} Example 2: - Explanation: {explanation_2} - Activations: {activations_2} When making your decision, you should follow these rules:

    work page

Show all 12 references
  1. [10]

    First pay attention to the feature’s explanation

  2. [11]

    If you cannot decide, you should then pay special attention to the tokens highlighted with <top>...</top> tags, as these are the most highly activated tokens and strongest indicators of what the feature detects

  3. [12]

    INSTRUCTION:

    Also consider the diversity of the activation examples provided. If one feature only activates one particular word, it may not be as aligned as a feature that activates on a variety of words. Based on the feature’s explanation and the highlighted tokens in the activation examp...

  4. [2025]

    Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna

    URL https://openreview.net/forum?id=Zy2XgaGpDw. Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. In European Conference on Computer Vis...

    2025 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.