Pith. sign in

REVIEW 3 major objections 6 minor 2 cited by

Analyzing Finetuning Representation Shift for Multimodal LLMs Steering

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Fine-tuned MLLM concepts can be recovered by adding simple mean shift vectors to the original model's concepts, and the same additive shifts steer outputs for debiasing and safety—all without retraining.

desk verdict A useful MLLM steering framework whose headline recovery claim is undercut by in-sample evaluation, but the held-out steering results and code make it worth refereeing. read the letter →

arxiv 2501.03012 v2 pith:RX4XAAFV submitted 2025-01-06 cs.AI cs.CLcs.CV

Pegah Khayatan , Mustafa Shukor , Jayneel Parekh , Arnaud Dapogny , Matthieu Cord This is my paper · ORCID
classification cs.AIcs.CLcs.CV
keywords concept-basedexplainabilitymultimodalLLMrepresentationshiftconceptvectorsmodelsteeringgenderdebiasingsafetyalignmentlinearhypothesis
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

This paper tries to show that the changes a multimodal LLM undergoes during fine-tuning can be captured as simple additive shifts of its latent concepts, and that these same shifts can be used to steer the model's behavior without any additional training. The authors extract interpretable visual and textual concepts from the residual stream of an MLLM, compute for each concept the mean change in representations between the original and fine-tuned model, and add that shift vector to the original concept. On fine-tuning experiments for places, colors, and sentiments, the shifted concepts are textually closer to the fine-tuned ones than the original concepts are, indicating partial recovery. They then use the same additive mechanism to change answer types in VQA, alter caption style, convert gendered captions to neutral ones, and reduce the attack success rate on a multimodal safety benchmark. If correct, this gives a training-free way to inspect what fine-tuning changes and to control MLLM outputs cheaply.

What carries the argument

Concept shift vectors: for each concept $u^a_k$ extracted from the original model as a K-means centroid in the residual stream, the shift vector $\Delta^{a\to b}_k$ is the average of the per-sample differences $b_m - a_m$ over the samples that most activate that concept; the shifted concept is $u^s_k = u^a_k + \alpha \Delta^{a\to b}_k$. The same construction, computed either between sample sets or between concept pairs, yields steering vectors applied additively to hidden states. Concepts are grounded in text through the unembedding matrix (logit lens), and concept similarity is measured by text grounding overlap (T-Overlap), which is the evaluation metric for recovery.

What would settle it

Test concept recovery on a held-out set of images that were not used to compute the shift vectors: if the T-Overlap improvement of shifted over original concepts disappears on those images, then the shift vectors only memorize the training set. Alternatively, find a fine-tuning regime where per-sample shifts are consistent but the shifted concept fails to match the fine-tuned concept in T-Overlap, which would break the claimed link between consistency and recovery.

Watch

Extended reading notes

Core claim

The paper's central claim is that fine-tuned concepts in a multimodal LLM are approximately recoverable from the original model by translating each original concept along a concept shift vector, defined as the mean of per-sample representation changes between the two models (Equation 4). The same shift-vector geometry, applied to activations rather than to concept dictionaries, provides a steering mechanism: coarse steering vectors are differences between mean representations of target and source sample sets, and fine-grained steering vectors are differences between paired concepts. The authors report that steering significantly increases the frequency of target answers, changes caption styles, converts gendered captions to neutral ones across three MLLMs, and drops the attack success rate from 45/100 to 5/100 on a MM-SafetyBench subset. They frame this as evidence that MLLM behavior can be monitored and controlled at a concept level without training.

Load-bearing premise

The load-bearing premise is that a concept is represented as a linear direction in the model's latent space, so that adding a mean difference vector moves the concept to its fine-tuned position; if concepts are non-linear or the shift vectors do not transfer to unseen images, both concept recovery and steering lose their foundation.

Editorial extensions

If this is right

  • Fine-tuned concept recovery does not require access to the fine-tuned model after the shift vectors are computed: the vectors are learned once and then applied to the original model.
  • The same additive shift applied to activations, not just to concept centroids, can change the distribution of VQA answers toward a target type or a specific answer while leaving other answer types largely unchanged.
  • Steering can alter caption style (colors, places, sentiments) and can debias captions, with fine-grained concept-to-concept steering converting more gendered captions to neutral ones than coarse steering.
  • Safety behavior can be steered: applying a vector derived from safe text-only responses shifts unsafe multimodal responses toward refusal, reducing the attack success rate on a subset of MM-SafetyBench from 45% to 5%.
  • Recovery quality correlates with shift consistency: concepts whose individual sample shifts align with the mean shift vector are recovered better, a relationship the paper quantifies with a statistically significant Pearson correlation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The method's reliance on mean shifts suggests it will work best when fine-tuning changes are approximately linear and unimodal; fine-tunings that rotate or nonlinearly warp concept manifolds would be expected to show worse recovery, which the paper's own citation to non-linear features implies.
  • The steering vectors could in principle be transferred between models in the same family, since only hidden-state differences are used; this is an extension the paper does not test.
  • A natural testable extension is to use shift vectors computed on one set of images to steer on a held-out set, which would validate the implicit assumption that the shift direction generalizes beyond the samples used to compute it.
  • The safety steering result suggests a general recipe: contrasting a model's behavior on the same prompt with and without an image isolates the contribution of the image modality, which could be used more broadly for diagnosing and correcting modality-induced biases.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. This paper proposes a concept-based framework for analyzing and steering multimodal LLMs (MLLMs). Concepts are extracted as k-means centroids of residual-stream activations and grounded in text and image spaces. For fine-tuned models, the authors define per-concept 'shift vectors' as the mean change of the representations of the images assigned to each original concept (Section 3.2, Eq. 4) and claim these recover fine-tuned concepts from the original model without accessing the fine-tuned model at inference. The same additive-shift mechanism is used for steering: coarse steering vectors between sets of answers/styles, and fine-grained steering vectors between concept pairs, with applications to VQA answer manipulation, caption style control, gender debiasing, and safety (refusal) alignment. Experiments cover LLaVA-1.5, Qwen2-VL-Instruct, and Idefics2, with fine-tuning on Visual Genome place/color/sentiment subsets.

Significance. The paper addresses a worthwhile and underexplored question — how fine-tuning changes the internal concept space of MLLMs — and offers a simple, training-free tool with plausible applications. Its strengths are the breadth of models and tasks (three MLLMs; VG fine-tuning subsets; VQAv2 and COCO evaluations), the extensive ablations (number of concepts, layer, steering strength, number of samples, token position), the public code release, and an honest limitations section acknowledging the linear representation hypothesis. The steering experiments are held-out on validation sets and, if confirmed with proper baselines, would be practically useful. The decisive weakness is that the central interpretability claim — concept recovery — currently rests on an in-sample evaluation, and the steering/safety results lack error bars and random-direction controls. With a held-out recovery evaluation and stronger baselines, the paper would make a solid contribution.

major comments (3)
  1. [Section 3.2, Eq. (4); Section 4.1, Fig. 6] The concept-recovery evaluation is in-sample and, at the default α=1, is close to tautological. Since S(1)=S(2), both U^a and U^b are extracted from the same images, and Δ_k is the mean of (b_m − a_m) over A_k. For k-means centroids (up to the sign convention in the definition of A_k via |v_i^a(x_m)|), u_a^k equals the mean of the a_m in its cluster, so Eq. (4) reduces to u_s^k = mean_{m∈A_k} b_m: the 'shifted concept' is the conditional mean of the fine-tuned representations of the original cluster's own images, and it is compared against fine-tuned centroids estimated from those same images. The T-Overlap gain in Fig. 6 is therefore largely a consequence of the construction and does not demonstrate that a shift vector computed on one set recovers fine-tuned concepts on unseen images. Please add a held-out evaluation: compute Δ on a training split, extract U^a and U^b from a disjoint split (or evaluate on held-out images), and report T-Overlap there, alongside baselines such as a single global shift (mean of all b_m − a_m) and a random direction of matched norm. The claim in Section 3.2 that shifted concepts are computed 'without accessing the fine-tuned model' also needs this held-out evidence, since computing Δ already requires the fine-tuned model's representations on the paired set.
  2. [Section 4.2, Tables 1–3, Fig. 9; App. B.1, B.3] The steering experiments report no error bars and no random-direction baseline, and the reported numbers are partly selected on the validation set used for reporting. The steering layer is chosen by evaluation on the VQAv2 validation set (App. B.1), and the 'meaningful' fine-grained directions are selected by their measured impact on the same kind of set (App. B.3); the text does not clarify whether the 5k-sample subset underlying Table 2 is the same set used for these selections. A control in which a random direction of comparable norm (or a PCA direction) is added at the same layer would be needed to show that the answer-type shifts in Tables 1–3 are due to the semantic content of the steering vector rather than to any large perturbation. Bootstrap confidence intervals on the answer counts would also substantiate the 'targeted steering' claim, especially since the Yes→No row of Table 2 changes yes/no accuracy by 21.8 points while 'other' accuracy still drops by 2.1 points.
  3. [Table 5; App. D] The safety result risks circularity because the steering vector is derived from samples on MM-SafetyBench using the same ASR metric on which it is then evaluated. The derivation set is selected as the MM-SafetyBench samples where Qwen2-VL-Instruct is unsafe with an image and safe without an image (judged by ASR), and the main paper reports ASR 45/100 → 5/100 on 'a portion' of MM-SafetyBench without stating that the 100 evaluation samples are disjoint from the derivation subset. Please specify the exact split, evaluate on a fixed held-out subset (or the full benchmark), and state the number of derivation samples used. The helpfulness check in Table 9 rests on only 100 samples with 3 borderline cases, so a larger evaluation of benign queries is needed to support the claim that the safety steering preserves utility.
minor comments (6)
  1. [Section 3.3] The fine-grained steering vector is defined with a duplicated term, 'sf_ij = sf_ij = uj − ui'; the intended definition is sf_ij = uj − ui. Additionally, the definition of A_k in Section 3.2 uses the absolute activation |v_i^a(x_m)|, whereas standard k-means assignment uses signed projections; please specify the exact assignment rule so that A_k provably corresponds to the clusters whose centroids appear in Eq. (4).
  2. [Section 4.1; Eq. (3)] There is a typo, 'Throught', in the first sentence of Section 4.1. Also, Eq. (3) defines T-Overlap with |Twords(u)| in the denominator, making the metric asymmetric when the two concepts produce different grounding-word sets; please state the common value of Ngrounding used for both arguments.
  3. [Fig. 7; App. A.5] The reported p-values of '0.0000' should be expressed as, e.g., p < 10^-4, and the text should state the number of concept-level points pooled and whether the Pearson correlation is computed within or across the four tokens of interest.
  4. [Table 4; App. C] The gender-debiasing results report only counts of gendered-to-neutral conversions; without a caption-quality metric such as CIDEr (or human evaluation), it is unclear whether the conversions preserve caption fidelity, particularly given the small totals (794, 815, and 926 samples, respectively).
  5. [Section 5] The limitations paragraph appropriately acknowledges the linear representation hypothesis via Ref. [14], but it should also note that all quantitative evaluations (T-Overlap, answer-type counts, ASR string matching) are automatic proxies and that no human evaluation of concept interpretability or of steered caption quality is provided.
  6. [App. D] There is a typo, 'safty', in App. D, and the relationship between the 733-sample steering-derivation subset of Table 8 and the 100-sample evaluation of Table 5 should be stated explicitly in the main text so that the reader can verify the train/evaluation split.

Circularity Check

1 steps flagged · score 6.0 of 10

Concept recovery is evaluated on the same images used to compute the shift vectors, so the main interpretability result is partly an in-sample fit rather than an independent prediction.

  1. fitted input called prediction [Section 3.2, Eq. (4); evaluation in Section 4.1, Fig. 6]
    "we fix the dataset S(1) = S(2), and obtain two sets of embeddings from f a, fb respectively... To compute the concept shift vector ∆a→b_k (ua_k) associated with ua_k, we aggregate shifts of its associated samples specified by Ak: ∆a→b_k (ua_k) = 1/|Ak| Σ_{m∈Ak} (bm − am) ... us_k = ua_k + α ∆a→b_k (ua_k). ... We evaluate how well a shifted concept us_k (Equ. (4)) is similar to its match ub_m(k)."

    Because the concepts are k-means centroids, each original concept ua_k equals the mean of the representations a_m over its own assigned set Ak. With the default α=1, Eq. (4) becomes us_k = mean_{m∈Ak} b_m, i.e. the 'shifted concept' is precisely the mean of the fine-tuned representations of the same images that defined the original cluster. The recovery metric in Fig. 6 then compares this within-sample conditional mean to fine-tuned k-means centroids estimated from the same set of fine-tuned representations B. There is no held-out split: the shift vector is fitted to and evaluated on the same images, so the reported improvement in T-Overlap is at least partly forced by the construction and does not demonstrate generalization to unseen images.

full rationale

The clearest circularity is in the fine-tuning concept-recovery experiment. The shift vector for a concept is the average original-to-finetuned representation difference over the images assigned to that concept, and recovery is measured on those same images; at α=1 the shifted concept is just the conditional mean of the fine-tuned representations, so Fig. 6 partly measures an in-sample fit rather than an independent recovery of fine-tuned concepts. The steering experiments are more independent: coarse and fine-grained vectors are computed from a train subset and evaluated on a validation set, and the paper reports ablations over layer, strength, and sample count. The use of the authors' earlier CoX-LMM framework [47] is a self-citation, but it is a published, code-released method and is not used to forbid alternatives, so it does not by itself make the derivation circular. The linear-representation assumption is an acknowledged limitation, not a circular step. Overall, the central recovery claim is partially circular due to the in-sample evaluation, while the steering applications retain independent empirical content.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The framework rests on the linear representation hypothesis and on the concept extraction and grounding machinery borrowed from the authors' CoX-LMM. The manual keyword lists for fine-tuning subsets and the layer and sample choices are additional degrees of freedom that shape all reported results. No genuinely new entities (particles, forces, dimensions) are introduced.

free parameters (5)
  • alpha (steering and shift magnitude) = 1 (default)
    Controls the size of the additive shift in Eqs. (4), (6), (10); ablated in App. B.5.3 and Fig. 14, but no principled selection criterion is given.
  • K (number of concepts) = 20
    Number of k-means clusters in main experiments; ablated in App. A.4 (Fig. 15), showing low sensitivity.
  • Layer l for extraction and steering = last layer for VQAv2, 20th for COCO, 19 for safety
    Chosen per application based on validation performance (App. B.5.2, App. D), a form of model selection on the evaluation distribution.
  • VG keyword subsets = Lists in App. A.2
    The fine-tuning datasets (places, colors, sentiments) are curated by hand-picked keyword lists, which determine what concepts emerge.
  • Number of samples for steering vector = A few hundred
    App. B.5.1 shows 50 samples already work, but the exact count is not fixed in the main protocol.
assumptions (5)
  • domain assumption Linear representation hypothesis: concepts correspond to linear directions and additive arithmetic works in activation space.
    Relied on for both concept shift (Eq. 4) and steering (Eq. 6); supported only by PCA visualizations (Fig. 8, 29) and prior LLM work [42, 48], not by a derivation or a targeted MLLM test.
  • domain assumption K-means clustering of residual stream features yields interpretable concepts.
    Inherited from CoX-LMM [47] by the same authors; the paper does not validate that clusters map to distinct semantic concepts beyond text grounding.
  • domain assumption Text grounding overlap (T-Overlap, Eq. 3) is a valid measure of concept similarity.
    The central quantitative metric for concept recovery; it only compares top vocabulary words after unembedding, which may miss geometric similarity.
  • domain assumption Unembedding matrix W_U applied to concept vectors yields meaningful text labels (logit lens).
    Used in Eq. (2) for all text groundings; a standard but heuristic interpretability tool.
  • domain assumption Bijective matching between original and fine-tuned concepts via cosine similarity is semantically meaningful.
    App. A.1; matching by cosine similarity in the same space may pair unrelated concepts, and the optimal transport constraint forces a bijection even when concepts do not correspond.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Analyzing Finetuning Representation Shift for Multimodal LLMs Steering." pith.science (2026). https://pith.science/paper/RX4XAAFV

@misc{pith2026250103012,
  author       = {Pith},
  title        = {Pith review of: Analyzing Finetuning Representation Shift for Multimodal LLMs Steering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RX4XAAFV}},
  note         = {Machine review of arXiv:2501.03012}
}
read the original abstract

Multimodal LLMs (MLLMs) have reached remarkable levels of proficiency in understanding multimodal inputs. However, understanding and interpreting the behavior of such complex models is a challenging task, not to mention the dynamic shifts that may occur during fine-tuning, or due to covariate shift between datasets. In this work, we apply concept-level analysis towards MLLM understanding. More specifically, we propose to map hidden states to interpretable visual and textual concepts. This enables us to more efficiently compare certain semantic dynamics, such as the shift from an original and fine-tuned model, revealing concept alteration and potential biases that may occur during fine-tuning. We also demonstrate the use of shift vectors to capture these concepts changes. These shift vectors allow us to recover fine-tuned concepts by applying simple, computationally inexpensive additive concept shifts in the original model. Finally, our findings also have direct applications for MLLM steering, which can be used for model debiasing as well as enforcing safety in MLLM output. All in all, we propose a novel, training-free, ready-to-use framework for MLLM behavior interpretability and control. Our implementation is publicly available.

Figures

Figures reproduced from arXiv: 2501.03012 by the authors.

Figure 1
Figure 1. Framework overview. We apply concept-level analysis for MLLM behavior monitor and control, for (left) understanding and manipulating (through shift vectors) concept changes due to fine-tuning, as well as (right) MLLM steering for debiasing or safety control. Specifically, given a set of M images {x1, ..., xM} we ex￾tract a set of residual stream representations from some layer l of the MLLM f. These representations … view at source ↗
Figure 2
Figure 2. Concepts extracted from original and fine-tuned mod￾els. concepts from the original f a (top), and model fine-tuned to focus more on places f b (bottom), for TOI = person. The concepts from f b exhibit a stronger association with places. where α controls the steering strength and it is set to 1 (we study α in App. B.5.3). Thus, in this setup, all examples are coarsely steered in the direction of the steering vector … view at source ↗
Figure 3
Figure 3. Concepts text grounding change after fine-tuning. Text grounding for concepts (TOI = bus) from f a and their match from f b , (fine-tuned to focus more on places). Emerging concepts may include grounding words not explicitly included in the fine-tuning vocabulary for place (e.g., ”District”, ”Crossing”), while others evolve more smoothly (e.g., ”Street”). match. Specifically, we define a matching function m: i \righ… view at source ↗
Figures from the paper (26 more)
Figure 5
Figure 5. Figure 5: Concepts evolve differently due to fine-tuning. Left: concepts extracted from the original model f a . Right: matched concept extracted from f b , fine-tuned to focus on places. Each concept is grounded in image and text. We observe different levels of adaptation acros…
Figure 4
Figure 4. Figure 4: Text grounding overlap (T-Overlap) between original and fine-tuned model concepts. Different concepts change to different extents depending on the fine-tuning. Concept evolution. To quantify how much a concept \bm {u}^a_{i} \in \bm {U}^a is changed after fine-tuning, w…
Figure 6
Figure 6. Figure 6: Recovering fine-tuned concepts. Across different fine￾tunings (places (top), colors (middle) and sentiments (bottom)), we compute the average text grounding overlap of original and shifted concepts with the matched fine-tuned ones. Shifting the original concepts result…
Figure 8
Figure 8. Figure 8: Linear separability of concepts features in MLLMs. We visualize the features related to the concepts ”yes” and ”no” after PCA projections across MLLMs layers. In summary, we demonstrated the feasibility of recovering target fine-tuned concepts by applying simple per-co…
Figure 7
Figure 7. Figure 7: Correlation between shift consistency and concept recovery (Color finetuning). The more consistent and aligned the individual representation shifts associated with a concept, the better the recovery of the fine-tuned concept. Which concepts are recovered better? We hyp…
Figure 9
Figure 9. Figure 9: Discovering meaningful steering directions. Each line corresponds to a fine-grained steering direction to steer the model answer to: ”No” (yes/no), ”4” (number) and ”Red” (other). Some steering directions are targeted (e.g., ”No”) as there is slight change in both the …
Figure 10
Figure 10. Figure 10: Steering MLLMs captions style. Captions steered to focus more on colors (left), places (middle) and sentiments (right). Steering image caption styles. We previously applied steer￾ing on relatively brief answers from the VQAv2 dataset. Here, we extend this approach to …
Figure 11
Figure 11. Figure 11: Concepts change during training. Illustration of the similarity between the original concepts the concepts during fine-tuning. Top: individual concepts change. Bottom: average concepts change [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: VG subsets. Keywords used to extract VG subsets. Each subset is selected based on the presence of the corresponding words in the captions. From top to bottom, words related to: places, colors, and sentiments. 0.98 0.63 0.90 0.60 0.99 0.68 0.98 0.72 0.98 0.68 t-SNE Vis…
Figure 14
Figure 14. Figure 14: Shift magnitude (α) and recovering fine-tuned model concepts. Illustration of the average of T-Overlap between shifted and matched fine-tuned concepts when varying the shift magnitude. paper, we focus on VQAv2 dataset [24], a visual question￾answering corpus with imag…
Figure 13
Figure 13. Figure 13: t-SNE visualization of 5 original concepts (red), shifted concepts (orange), and their corresponding fine-tuned concepts (blue). Dotted lines connect original and fine-tuned pairs, while dashed lines connect shifted and fine-tuned pairs. Numerical values indicate cosi…
Figure 15
Figure 15. Figure 15: Number of concepts and recovery. Varying the number of concepts K has minimal impact on the recovery, as measured by the overlap metrics, indicating the robustness of the recovery process to the choice of K. 0 20 40 60 80 100 VG-place 0 20 40 60 80 100 VG-place 0 20 4…
Figure 17
Figure 17. Figure 17: Correlation between shift consistency and concept recovery (Place, Color and Sentiment finetuning). The more consistent and aligned are the individual shift vectors associated with a concept, the better is recovery of the fine-tuned concept that can be achieved using …
Figure 18
Figure 18. Figure 18: Discovering meaningful steering directions. Each line corresponds to a finegrained steering direction to steer the model answer to (from top to bottom): ”No” (yes/no), ”Yes” (yes/no), ”2” (number) and ”4” (number). First line corresponds to the original model without …
Figure 20
Figure 20. Figure 20: Discovering meaningful steering directions with image captioning. We report the relative increase in number of words counts. Each figure corresponds to different fine-grained steering direction. 0 500 1000 1500 Samples 60 65 70 Accuracy other overall yes/no number 0 5…
Figure 21
Figure 21. Figure 21: Ablation study: number of samples to compute steering vector. From top to bottom: steering answers from ”Yes” (yes/no), ”1” (number) to ”No”, ”3” respectively. We report different metrics as follows (from left to right): VQA accuracy per answer type, number of answers…
Figure 26
Figure 26. Figure 26: Ablation study: steering strength (α) and changing caption styles. From left to right: steering captions style to include more: colors, places and sentiments. We report the CIDEr score. Despite having more captions from the target style, significantly increasing α lea…
Figure 23
Figure 23. Figure 23: Ablation study: steering strength (α) and changing answer types. From left to right: steering answers type towards: yes/no, number and other. We report the number of answers in each answer type. Increasing α pushes the model to generate more answers from the target ty…
Figure 25
Figure 25. Figure 25: Ablation study: steering strength (α) and changing caption styles. From left to right: steering captions style to include more: colors, places and sentiments. We report the number words belonging to each type. Increasing α pushes the model to generate words related to…
Figure 28
Figure 28. Figure 28: Ablation study: steering strength (α). From top to bottom: steering answers from ”Yes” (yes/no), ”1” (number) to ”No”, ”3”. We report different metrics as follows (from left to right): VQA accuracy per answer type, number of answers belonging to each type, number of o…
Figure 29
Figure 29. Figure 29: Linear separability of concepts features in MLLMs. We visualize the features related to the concepts ”yes”/”no”, ”1”/”3” and ”white”/”black” after PCA projections across MLLMs layers. COCO GENDERED WORDS ”man”, ”woman”, ”boy”, ”girl”, ”gentleman”, ”lady”, ”male”, ”fem…
Figure 30
Figure 30. Figure 30: Words employed for neutral words-matching in the [PITH_FULL_IMAGE:figures/full_fig_p020_30.png]
Figure 32
Figure 32. Figure 32: Each image is presented with three captions: (1) the [PITH_FULL_IMAGE:figures/full_fig_p021_32.png]
Figure 33
Figure 33. Figure 33: Steering MLLMs answers. Each line corresponds to different steering vector that change a specific original answer to a target one. From top to bottom: ”white” to ”black”, ”1” to ”3” and ”yes” to ”no” [PITH_FULL_IMAGE:figures/full_fig_p024_33.png]
Figure 34
Figure 34. Figure 34: Steering MLLMs answers type. Each line corresponds to different steering vector that change answers type to a target one. Steering vectors correspond to changing the answers type to yes/no (top) and numbers (bottom) [PITH_FULL_IMAGE:figures/full_fig_p025_34.png]
Figure 35
Figure 35. Figure 35: Steering MLLMs captions type. Each line corresponds to different steering vector that change captions style to a target one. Steering vectors correspond to changing the captions style so that they contain more: colors (top), places (middle) and sentiments (bottom) [P…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Multimodal Model Diffing for Feature Discovery and Control

    cs.CV 2026-08 conditional novelty 7.0 of 10

    By diffing base-language and multimodal sparse autoencoder features, MMDiff isolates causally relevant features that can be ablated or steered to control spatial, OCR, and safety behaviors in multimodal LLMs.

  2. GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs

    cs.CL 2025-07 conditional novelty 6.0 of 10

    GrAInS uses Integrated Gradients to identify the most influential tokens, then builds layer-wise steering vectors that improve truthfulness, reduce hallucination, and preserve general capabilities in LLMs and VLMs.

Reference graph

Works this paper leans on

75 extracted references · 42 canonical work pages · cited by 2 Pith papers

  1. [1]

    Flamingo: a visual language model for few-shot learning

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems , 35:23716–23736,

  2. [2]

    Refusal in language models is mediated by a single direction

    Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Rimsky, Wes Gurnee, and Neel Nanda. Refusal in language models is mediated by a single direction. arXiv preprint arXiv:2406.11717, 2024. 2

    arXiv 2024

  3. [3]

    Qwen-vl: A frontier large vision-language model with versatile abilities

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1, 2

    arXiv 2023

  4. [4]

    What makes multimodal in-context learning work? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1539–1550, 2024

    Folco Bertini Baldassini, Mustafa Shukor, Matthieu Cord, Laure Soulier, and Benjamin Piwowarski. What makes multimodal in-context learning work? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1539–1550, 2024. 1, 2

    work page 2024

  5. [5]

    Ostrovsky, Lev McKinney, Stella Biderman, and Jacob Steinhardt

    Nora Belrose, Zach Furman, Logan Smith, Danny Halawi, Igor V . Ostrovsky, Lev McKinney, Stella Biderman, and Jacob Steinhardt. Eliciting latent predictions from transformers with the tuned lens. ArXiv, abs/2303.08112, 2023. 3

    arXiv 2023

  6. [6]

    Leace: Perfect linear concept erasure in closed form

    Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. Leace: Perfect linear concept erasure in closed form. Advances in Neural Information Processing Systems, 36, 2024. 2

    work page 2024

  7. [7]

    Lan- guage models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Lan- guage models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877–1901,

    work page 1901

  8. [8]

    Rex: Reasoning-aware and grounded explanation

    Shi Chen and Qi Zhao. Rex: Reasoning-aware and grounded explanation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15586– 15595, 2022. 2

    work page 2022

Show all 75 references
  1. [9]

    Understanding and improving in-context learning on vision-language models

    Shuo Chen, Zhen Han, Bailan He, Mark Buckley, Philip Torr, V olker Tresp, and Jindong Gu. Understanding and improving in-context learning on vision-language models. In ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models, 2024. 2

    2024

  2. [10]

    Palm: Scaling language modeling with pathways

    Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1

    2022 arXiv

  3. [11]

    Eta: Evaluating then aligning safety of vision language models at inference time

    Yi Ding, Bolian Li, and Ruqi Zhang. Eta: Evaluating then aligning safety of vision language models at inference time. ArXiv, abs/2410.06625, 2024. 8, 10, 11

    2024 arXiv

  4. [12]

    Palm-e: An embodied multimodal language model

    Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1

    2023 arXiv

  5. [13]

    A mathematical framework for transformer circuits

    Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yun- tao Bai, Anna Chen, Tom Conerly, et al. A mathematical framework for transformer circuits. Transformer Circuits Thread, 1, 2021. 1

    2021

  6. [14]

    Not all language model features are linear

    Joshua Engels, Eric J Michaud, Isaac Liao, Wes Gurnee, and Max Tegmark. Not all language model features are linear. arXiv preprint arXiv:2405.14860, 2024. 8

    2024 arXiv

  7. [15]

    Multimodal task vectors enable many- shot multimodal in-context learning

    Brandon Huang et al. Multimodal task vectors enable many- shot multimodal in-context learning. In NeurIPS, 2024. 2

    2024

  8. [16]

    LIVE: Learnable in-context vector for visual question answering

    Yingzhe et al. LIVE: Learnable in-context vector for visual question answering. In NeurIPS, 2024. 2

    2024

  9. [17]

    A holistic approach to unifying automatic concept extraction and concept importance estimation

    Thomas Fel, Victor Boutin, Louis B ´ethune, R ´emi Cad`ene, Mazda Moayeri, L ´eo And ´eol, Mathieu Chalvidal, and Thomas Serre. A holistic approach to unifying automatic concept extraction and concept importance estimation. Ad- vances in Neural Information Processing Systems, ...

    2023

  10. [18]

    Craft: Concept recursive activation factorization for explainability

    Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, R´emi Cad`ene, and Thomas Serre. Craft: Concept recursive activation factorization for explainability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognitio...

    2023

  11. [19]

    Mul- timodal autoregressive pre-training of large vision encoders

    Enrico Fini, Mustafa Shukor, Xiujun Li, Philipp Dufter, Michal Klein, David Haldimann, Sai Aitharaju, Victor Guil- herme Turrisi da Costa, Louis B´ethune, Zhe Gan, et al. Mul- timodal autoregressive pre-training of large vision encoders. arXiv preprint arXiv:2411.14402, 2024. 2

    2024 arXiv

  12. [20]

    From wrong to right: A recursive approach towards vision- language explanation

    Jiaxin Ge, Sanjay Subramanian, Trevor Darrell, and Boyi Li. From wrong to right: A recursive approach towards vision- language explanation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1173–1185, 2023. 2

    2023

  13. [21]

    Towards automatic concept-based explanations

    Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. In Ad- vances in Neural Information Processing Systems (NeurIPS), pages 9277–9286, 2019. 2

    2019

  14. [22]

    Fig- step: Jailbreaking large vision-language models via typo- graphic visual prompts

    Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tian- shuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Fig- step: Jailbreaking large vision-language models via typo- graphic visual prompts. ArXiv, abs/2311.05608, 2023. 11

    2023 arXiv

  15. [23]

    Kwok, and Yu Zhang

    Yunhao Gou, Kai Chen, Zhili Liu, Lanqing Hong, Hang Xu, Zhenguo Li, Dit-Yan Yeung, James T. Kwok, and Yu Zhang. Eyes closed, safety on: Protecting multimodal llms via image- to-text transformation. In European Conference on Computer Vision, 2024. 11

    2024

  16. [24]

    Making the v in vqa matter: Elevating the role of image understanding in visual question answering

    Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Ba- tra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904–6913, ...

    2017

  17. [25]

    Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen

    J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685,

    arXiv

  18. [26]

    Miner: Mining the underlying pattern of modality-specific neurons in multimodal large language models

    Kaichen Huang, Jiahao Huo, Yibo Yan, Kun Wang, Yutao Yue, and Xuming Hu. Miner: Mining the underlying pattern of modality-specific neurons in multimodal large language models. arXiv preprint arXiv:2410.04819, 2024. 2

    2024 arXiv

  19. [27]

    Sparse autoencoders find highly interpretable features in language models

    Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representa- tions, 2024. 2

    2024

  20. [28]

    Mistral 7b

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825,

    arXiv

  21. [29]

    Mimic in-context learning for multimodal tasks

    Yuchu Jiang, Jiale Fu, Chenduo Hao, Xinting Hu, Yingzhe Peng, Xin Geng, and Xu Yang. Mimic in-context learning for multimodal tasks. arXiv preprint arXiv:2504.08851, 2025. 2

    2025 arXiv

  22. [30]

    Interpretability beyond feature attribution: Quantitative testing with concept activa- tion vectors (tcav)

    Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activa- tion vectors (tcav). In International conference on machine learning, pages 2668–2677. PMLR, 2018. 2

    2018

  23. [31]

    Shamma, Michael S

    Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalan- tidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. ...

    2017

  24. [32]

    Decoderlens: Layerwise interpretation of encoder-decoder transformers

    Anna Langedijk, Hosein Mohebbi, Gabriele Sarti, Willem Zuidema, and Jaap Jumelet. Decoderlens: Layerwise interpretation of encoder-decoder transformers. ArXiv, abs/2310.03686, 2023. 3

    2023 arXiv

  25. [33]

    Obelics: An open web-scale filtered dataset of interleaved image-text documents

    Hugo Laurenc ¸on, Lucile Saulnier, L´eo Tronchon, Stas Bek- man, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Sid- dharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. Advances in Neural Informa...

    2024

  26. [34]

    What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024

    Hugo Laurenc ¸on, L´eo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 1, 2, 3

    2024 arXiv

  27. [35]

    Inference-time intervention: Eliciting truthful answers from a language model

    Kenneth Li, Oam Patel, Fernanda Vi´egas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36, 2024. 2

    2024

  28. [36]

    Salad-bench: A hierarchical and comprehensive safety benchmark for large language models

    Lijun Li, Bowen Dong, Ruohui Wang, Xuhao Hu, Wangmeng Zuo, Dahua Lin, Yu Qiao, and Jing Shao. Salad-bench: A hierarchical and comprehensive safety benchmark for large language models. In Annual Meeting of the Association for Computational Linguistics, 2024. 10

    2024

  29. [37]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceeding...

    2014

  30. [38]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2023. 4

    2023

  31. [39]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296–26306, 2024. 1, 2, 3, 7

    2024

  32. [40]

    Mm-safetybench: A benchmark for safety evaluation of multimodal large language models

    Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. InEuropean Conference on Computer Vision, 2023. 8, 10

    2023

  33. [41]

    Mapl: Parameter-efficient adaptation of unimodal pre-trained mod- els for vision-language few-shot prompting

    Oscar Ma ˜nas, Pau Rodriguez Lopez, Saba Ahmadi, Aida Nematzadeh, Yash Goyal, and Aishwarya Agrawal. Mapl: Parameter-efficient adaptation of unimodal pre-trained mod- els for vision-language few-shot prompting. In Proceedings of the 17th Conference of the European Chapter of t...

  34. [42]

    Actually, othello-gpt has a linear emergent world model, 2023

    Neel Nanda. Actually, othello-gpt has a linear emergent world model, 2023. 6, 8

    2023

  35. [43]

    Interpreting gpt: The logit lens

    Nostalgebraist. Interpreting gpt: The logit lens. https : / / www . lesswrong . com / posts / AcKRB8wDpdaN6v6ru / interpreting - gpt - the - logit-lens, 2020. Accessed: [date of access]. 3

    2020

  36. [44]

    Gpt-4 technical report

    OpenAI. Gpt-4 technical report. arXiv, 2023. 1

    2023

  37. [45]

    Find- ing and editing multi-modal neurons in pre-trained trans- former

    Haowen Pan, Yixin Cao, Xiaozhi Wang, and Xun Yang. Find- ing and editing multi-modal neurons in pre-trained trans- former. arXiv preprint arXiv:2311.07470, 2023. 2

    2023 arXiv

  38. [46]

    Steering llama 2 via contrastive activation addition

    Nina Panickssery, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Matt Turner. Steering llama 2 via contrastive activation addition. arXiv preprint arXiv:2312.06681, 2023. 2

    2023 arXiv

  39. [47]

    A concept-based explainability framework for large multimodal models

    Jayneel Parekh, Pegah Khayatan, Mustafa Shukor, Alasdair Newson, and Matthieu Cord. A concept-based explainability framework for large multimodal models. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2

    2024

  40. [48]

    The linear representation hypothesis and the geometry of large language models

    Kiho Park, Yo Joong Choe, and Victor Veitch. The linear representation hypothesis and the geometry of large language models. In Forty-first International Conference on Machine Learning, 2024. 6, 8

    2024

  41. [49]

    What factors affect multi-modal in- context learning? an in-depth exploration

    Libo Qin, Qiguang Chen, Hao Fei, Zhi Chen, Min Li, and Wanxiang Che. What factors affect multi-modal in- context learning? an in-depth exploration. arXiv preprint arXiv:2410.20482, 2024. 2

    2024 arXiv

  42. [50]

    Ui-tars: Pioneering automated gui inter- action with native agents

    Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shi- jue Huang, et al. Ui-tars: Pioneering automated gui inter- action with native agents. arXiv preprint arXiv:2501.12326,

    arXiv

  43. [51]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  44. [52]

    Improving dictionary learning with gated sparse autoencoders

    Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Tom Lieberum, Vikrant Varma, J´anos Kram´ar, Rohin Shah, and Neel Nanda. Improving dictionary learning with gated sparse autoencoders. In Advances in Neural Information Processing Systems, 2024. 2

    2024

  45. [53]

    Linear adversarial concept erasure

    Shauli Ravfogel, Michael Twiton, Yoav Goldberg, and Ryan D Cotterell. Linear adversarial concept erasure. In International Conference on Machine Learning, pages 18400– 18421. PMLR, 2022. 2

    2022

  46. [54]

    Mansi Sakarvadia, Arham Khan, Aswathy Ajith, Daniel Grzenda, Nathaniel Hudson, Andr´e Bauer, Kyle Chard, and Ian T. Foster. Attention lens: A tool for mechanistically inter- preting the attention head information retrieval mechanism. CoRR, abs/2310.16270, 2023. 3

    arXiv

  47. [55]

    Multimodal neurons in pretrained text-only transformers

    Sarah Schwettmann, Neil Chowdhury, Samuel Klein, David Bau, and Antonio Torralba. Multimodal neurons in pretrained text-only transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2862– 2867, 2023. 1, 2

    2023

  48. [56]

    Implicit multimodal alignment: On the generalization of frozen llms to multimodal inputs

    Mustafa Shukor and Matthieu Cord. Implicit multimodal alignment: On the generalization of frozen llms to multimodal inputs. Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2

    2024

  49. [57]

    Skipping computations in multimodal llms

    Mustafa Shukor and Matthieu Cord. Skipping computations in multimodal llms. arXiv preprint arXiv:2410.09454, 2024. 1

    2024 arXiv

  50. [58]

    ep- alm: Efficient perceptual augmentation of language models

    Mustafa Shukor, Corentin Dancette, and Matthieu Cord. ep- alm: Efficient perceptual augmentation of language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22056–22069, 2023. 1

    2023

  51. [59]

    Beyond task performance: evaluating and re- ducing the flaws of large multimodal models with in-context- learning

    Mustafa Shukor, Alexandre Rame, Corentin Dancette, and Matthieu Cord. Beyond task performance: evaluating and re- ducing the flaws of large multimodal models with in-context- learning. In The Twelfth International Conference on Learn- ing Representations, 2024. 1, 2

    2024

  52. [60]

    Smolvla: A vision-language-action model for afford- able and efficient robotics

    Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, Michel Ar- actingi, Caroline Pascal, Martino Russi, Andres Marafioti, et al. Smolvla: A vision-language-action model for afford- able and efficient robotics. arXiv preprint arXiv...

    arXiv

  53. [61]

    Scaling laws for native multimodal models

    Mustafa Shukor, Enrico Fini, Victor Guilherme Turrisi da Costa, Matthieu Cord, Joshua Susskind, and Alaaeldin El- Nouby. Scaling laws for native multimodal models. arXiv preprint arXiv:2504.07951, 2025. 1

    2025 arXiv

  54. [62]

    Extracting latent steering vectors from pretrained language models

    Nishant Subramani, Nivedita Suresh, and Matthew E Peters. Extracting latent steering vectors from pretrained language models. In Findings of the Association for Computational Linguistics: ACL 2022, pages 566–581, 2022. 2

    2022

  55. [63]

    Gemma 2: Improving open language models at a practical size

    Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupati- raju, L´eonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ram´e, et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118,

    arXiv

  56. [64]

    Linear representations of sentiment in large language models

    Curt Tigges, Oskar John Hollinsworth, Atticus Geiger, and Neel Nanda. Linear representations of sentiment in large language models. arXiv preprint arXiv:2310.15154, 2023. 2

    2023 arXiv

  57. [65]

    Llama 2: Open foundation and fine-tuned chat models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Am- jad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2

    2023 arXiv

  58. [66]

    Activation addition: Steering language models without opti- mization

    Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. Activation addition: Steering language models without opti- mization. arXiv preprint arXiv:2308.10248, 2023. 2

    2023 arXiv

  59. [67]

    Improved baselines for data-efficient percep- tual augmentation of llms

    Th´eophane Vallaeys, Mustafa Shukor, Matthieu Cord, and Jakob Verbeek. Improved baselines for data-efficient percep- tual augmentation of llms. arXiv preprint arXiv:2403.13499,

    arXiv

  60. [68]

    Adashield: Safeguarding multimodal large lan- guage models from structure-based attack via adaptive shield prompting

    Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large lan- guage models from structure-based attack via adaptive shield prompting. In European Conference on Computer Vision ,

  61. [69]

    Reft: Representation finetuning for language models

    Zhengxuan Wu, Aryaman Arora, Zheng Wang, Atticus Geiger, Dan Jurafsky, Christopher D Manning, and Christo- pher Potts. Reft: Representation finetuning for language models. arXiv preprint arXiv:2404.03592, 2024. 2

    2024 arXiv

  62. [70]

    Few- shot multimodal explanation for visual question answering

    Dizhan Xue, Shengsheng Qian, and Changsheng Xu. Few- shot multimodal explanation for visual question answering. In ACM Multimedia 2024, 2024. 2

    2024

  63. [71]

    On concept- based explanations in deep neural networks

    Chih-Kuan Yeh, Been Kim, Sercan O Arik, Chun-Liang Li, Pradeep Ravikumar, and Tomas Pfister. On concept- based explanations in deep neural networks. arXiv preprint arXiv:1910.07969, 2019. 2

    1910 arXiv

  64. [72]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975–11986, 2023. 2

    2023

  65. [73]

    From redundancy to relevance: Enhancing ex- plainability in multimodal large language models

    Xiaofeng Zhang, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, Hao Tang, and Jieping Ye. From redundancy to relevance: Enhancing ex- plainability in multimodal large language models. arXiv preprint arXiv:2406.06579, 2024. 1 Analyzing Fine-tuning ...

    2024 arXiv

  66. [74]

    Different than VG, COCO contains captions describing the image general, often focusing on the central object

    for hidden states extraction, throughout the quantitative experiments. Different than VG, COCO contains captions describing the image general, often focusing on the central object. A.3. Concepts change during training In this section, we study how fine-tuning deviates the fine...

  67. [75]

    Words employed for gendered words-matching in the COCO dataset

    to extract gendered and neutral samples based on spe- COCO GENDERED WORDS ”person”, ”individual”, ”child”, ”kid”, ”children”, ”youth”, ”adult”, ”human” Figure 31. Words employed for gendered words-matching in the COCO dataset. cific word lists. We define the set of gendered wo...

    2012

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.