{"id":"6d97f553-8663-4449-b4a8-b1a59601f2a3","arxiv_id":"2505.20029","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Instruction-tuned multimodal LLMs predict fMRI responses to natural images better than vision-only models and on par with CLIP, though most explained variance is shared across instructions.","lead":"This paper asks whether multimodal AI models that can answer questions about pictures produce internal signals that resemble how the human visual brain responds to the same pictures. Using fMRI data from people watching natural scenes, it finds these AI models predict brain activity as well as earlier vision-language models, though different question types mostly explain the same shared brain signal.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Instruction-specific brain alignment may largely reflect text-generation artifacts, because features are mean-pooled over variable-length generated output tokens rather than over a fixed, instruction-controlled visual representation.","rationale":"The reader's weakest assumption identifies exactly the load-bearing point: Section 4's mean-pooled hidden states over generated tokens are treated as faithful, instruction-specific image representations. My stress-test sharpens this into a concrete confound mechanism and proposes a decisive control. This concern does not invalidate the broader empirical result that MLLM representations predict fMRI better than vision-only models; that comparison is likely robust to generation length because it is averaged over all instructions. However, the more novel and prominent claim—that varying instructions steers the model to encode different visual concepts aligned with the brain—is only as strong as the assumption that the pooled token embeddings isolate instruction-driven visual content. Appendix N's acknowledgment that NSD is passive-viewing data further weakens any direct brain-side validation of 'task-specific' processing. Since the paper is already CONDITIONAL with moderate confidence and this is precisely the stated weakest assumption, the appropriate action is to keep the verdict unchanged while requiring the proposed control for full acceptance. I found no circularity, invented entities, or internal inconsistency; the empirical pipeline is standard and the code link is provided, though not pinned to a commit.","tokens_in":24529,"tokens_out":3775,"duration_ms":66066,"concrete_test":"Re-extract MLLM features for all 10 instructions using the hidden states immediately after encoding the image and instruction but before any output token is generated (e.g., the last hidden state of the instruction token, or the Q-Former output in InstructBLIP). If the per-voxel argmax maps (Fig. 3) and the instruction ranking of brain alignment stay largely unchanged, the effect is not a generation artifact. If they collapse or reorder substantially, the reported instruction-specificity is driven by generated text. A complementary cheap check: truncate every generation to the first K=5 tokens and recompute Fig. 3; large map changes would confirm a length confound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novelty is that one MLLM can be steered by natural-language instructions to capture different visual concepts, evidenced by instruction-specific alignment maps (Figs. 3–4) and shared-variance decompositions (Figs. 6–7). This claim rests on the feature extraction of Section 4: 'We then average these hidden state representations of the output generated tokens to obtain the final embedding.' Mean-pooling over all generated tokens conflates several confounds: (1) generation length differs widely across the 10 instructions (e.g., 'There are two animals' vs. a multi-sentence caption), changing the effective feature norm and direction; (2) the lexical content of generated text carries strong language-model priors (object names, scene words) that may align with brain responses independently of any visual processing; (3) autoregressive generation conditions each token on the previous text, so the pooled embedding reflects a text-generation trajectory, not a stable image-plus-instruction representation. Because the fMRI data were collected during passive viewing with no task instruction given to participants (Appendix N), there is no independent neural measure of instruction-specific processing; the only 'instruction-specific' signal comes from the model's own outputs. The authors' control in Appendix M (no image input) shows that visual input is necessary, but it does not rule out the text-generation confound, since both conditions still use generated tokens. Thus the paper's strongest claim—instruction steering produces concept-specific brain alignment—could be an artifact of variable-length, lexically heterogeneous text outputs rather than a genuine instruction-specific visual representation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates whether instruction-tuned multimodal large language models (MLLMs) can serve as brain-encoding models for visual stimuli. Using the Natural Scenes Dataset, the authors extract embeddings by mean-pooling hidden states of output tokens generated by InstructBLIP, mPLUG-Owl, and IDEFICS under ten natural-language instructions, then train voxel-wise ridge regression encoding models. They report that MLLMs show significantly better normalized brain alignment than a vision-only ViT-H baseline and perform comparably to CLIP-Text, that some instructions (e.g., image captioning, visual question answering) align better with specific visual ROIs, that MLLMs capture count- and recognition-related concepts, and that most explained variance is shared across instructions. The paper includes extensive appendices with layer-wise maps, category analyses, variance partitioning, and a limitations statement.","tokens_in":24848,"tokens_out":2498,"duration_ms":32814,"significance":"If the central claims hold, the paper would extend brain-encoding methodology by showing that a single instruction-tuned MLLM can be steered by natural-language prompts to produce representations with different degrees of alignment to visual cortex, potentially offering a flexible alternative to training separate task-specific models. The work uses a standard encoding pipeline on four NSD subjects with held-out test images, and it makes code publicly available, which are strengths. However, the load-bearing comparisons and the instruction-specific conclusions are currently undermined by the feature-extraction procedure, the unmatched CLIP baseline, and the absence of a stated significance test, so the significance of the claimed findings depends on whether these issues can be resolved.","major_comments":[{"comment":"The instruction-specific alignment results in Figures 3-7 rest on features obtained by mean-pooling hidden states over all generated output tokens for each instruction. This procedure conflates genuine image-plus-instruction representation with several generation artifacts: generation length varies widely across instructions, the lexical content of generated text carries language-model priors that may align with brain responses independently of visual processing, and autoregressive generation makes the pooled embedding depend on the whole text-generation trajectory rather than on a fixed, instruction-controlled visual representation. Because the fMRI data were collected during passive viewing with no task instruction given to participants (Appendix N), the only instruction-specific signal in the pipeline comes from the model's own outputs. The control in Appendix M (image only, instruction only) shows that visual input is necessary, but it does not exclude the text-generation confound because both compared conditions still use generated tokens. The authors should provide a fixed-length or generation-independent feature extraction (e.g., prompting only, or using the representation at the final prompt token before generation) and should control for generation length and lexical content, or explicitly show that the instruction-specific alignment patterns survive such controls.","section":"Section 4 (feature extraction)"},{"comment":"The claim that MLLMs perform 'comparably to' CLIP is based on an unmatched comparison: the CLIP-Text model receives ground-truth oracle captions, while the MLLM embeddings come from model-generated tokens. The appendix also compares LLaMA-2-7B using captions, again with oracle input. This asymmetry can either inflate or deflate the relative performance in a way that is not controlled. Additionally, the asterisk and caret markers in Figure 2 are said to indicate 'statistically significantly better' with p <= 0.05, but no statistical test, multiple-comparison correction, or error-bar definition is given in the main text or figure caption; the only test described in the appendices (a one-way ANOVA in Appendix F) compares early versus higher visual ROIs, not model pairs. The authors should state the exact test used for each marker, report effect sizes and variance across subjects, and either match the CLIP baseline by using captions generated from the same model or clearly frame the comparison as 'MLLM-predicted text versus oracle captions.'","section":"Section 6.1, Figure 2 and Appendix K"},{"comment":"The conclusion that MLLMs 'effectively capture count-related and recognition-related concepts' is based on grouping the ten instructions into five concept categories after the fact, with instruction-concept memberships chosen by the authors (e.g., VQ2 appears in both Count and Recognition, so the groups are not disjoint). The voxel-level winner-take-all maps in Figures 3-4 then select, for each voxel, the instruction with the highest alignment on the same data used to motivate the groupings. This post-hoc selection can create apparent concept specificity even if the underlying instruction representations are not specifically aligned with count or recognition processing. The authors should pre-specify the concept groupings, use disjoint instruction sets, or validate the grouping with a held-out set of instructions; without this, the 'effective capture' claim is not yet established.","section":"Section 6.2 (visual concept grouping)"},{"comment":"The limitations section states that NSD was collected while subjects passively viewed images, 'so the dataset may not fully capture how brain activity aligns with task-specific instructions.' This is a direct acknowledgment that the passive-viewing data cannot independently ground the instruction-specific alignment maps that form the paper's central novelty. The authors should either soften the main-text claims (e.g., 'instruction-conditioned model representations' rather than 'instruction-specific neural processing') or add an analysis that ties the instruction-specific model differences to a neural measure that is not solely derived from the model's own outputs. The current framing overstates what can be concluded from passive-viewing fMRI.","section":"Appendix N (Limitations)"}],"minor_comments":[{"comment":"The caption says the diagonal cells are empty, but the displayed matrix appears to contain values near 1 on the diagonal; please clarify whether the diagonal is excluded or set to 1 by construction.","section":"Section 6.3, Figure 6 caption"},{"comment":"The brain-map figures use a color code for ten instructions and then reuse the same color bar for concept groups in Figure 4; the captions should state whether the color assignment per instruction is identical across all subjects and whether the maps are thresholded, because the current presentation makes it hard to compare across subjects.","section":"Section 6.2, Figure 3 and Appendix G"},{"comment":"The sentence 'The hidden states are dynamic and continuously evolve until the model generates the final token, which involves multiple forward passes in the process' is grammatically awkward; it should also specify whether the hidden states are taken from the language model backbone only or also from the vision encoder and Q-Former, since PCA is applied after pooling.","section":"Section 4"},{"comment":"The normalized-brain-alignment procedure divides Pearson correlation by cross-subject prediction accuracy and then averages across voxels, regions, and participants; the paper should report the distribution of this normalized measure across voxels or subjects rather than only the mean, especially because Figure 2 reports values up to 0.9 that strongly depend on the choice of the voxel-inclusion threshold (cross-subject accuracy >= 0.05).","section":"Section 5 (Evaluation metrics)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable empirical study with a standard encoding pipeline and useful appendices, but the main novelty—instruction-specific brain alignment in MLLMs—currently rests on a feature-extraction procedure that pools variable-length generated text and on comparisons that are not fully matched or statistically documented. The passive-viewing limitation is acknowledged in Appendix N, which weakens the central claim as stated. I do not see evidence of circularity in the sense of using the model's own outputs as ground truth: the alignment is measured against external fMRI data, and the post-hoc grouping and winner-take-all selection are selection concerns rather than circularity. The issues are fixable in principle (e.g., re-extracting features at a fixed token position, adding a paired significance test, and pre-registering or cross-validating the concept groupings), so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the bottom line: this is a legitimate and useful paper. It's the first to test instruction-tuned multimodal LLMs as brain encoding models with ten different natural-language prompts, and the central empirical finding—MLLMs beat ViT-H and land about on par with CLIP on NSD—looks credible given the standard four-subject, held-out-1000-image encoding pipeline, random-initialization controls, and public code. The variance-partitioning analysis is a nice addition; it quantifies how much of the explained variance is shared across instructions, and the finding that most is shared is a sobering, honest result.\n\nThe soft spots are real but not load-bearing for the main comparison. First, the significance stars in Figures 2 and 10 are not backed by a stated test or error bars; the appendix has some ANOVAs, but the pairwise star comparisons are under-specified. Second, CLIP uses oracle captions while MLLMs use their own generated tokens. That actually makes the MLLMs' parity with CLIP more impressive, but the comparison should be labelled that way. Third—and this is the one I'd push hardest on—the instruction-specific steering claims rest on mean-pooling hidden states over variable-length generated output. Generation length and lexical content differ a lot across prompts, so part of the apparent 'instruction-specific' brain alignment could come from text-statistics confounds rather than genuinely different visual representations. The appended stress-test note is right that the passive-viewing NSD data gives you no independent neural measure of instruction processing; the no-image control rules out the pure text-only case but not the variable-length generation confound. The authors acknowledge the passive-viewing limitation in Appendix N, which is fair.\n\nThe overall claim—that instruction-tuned MLLMs are viable brain encoders—holds up despite that. The instruction-differentiation claim is more fragile, but it's a suggestive result you can address with fixed-length representations, length-matched controls, or per-token position analyses in a revision.\n\nI'd take this for peer review. A good referee should engage with the method and push on the feature-extraction controls, but the core benchmark is worth publishing. I'd also bring it to a reading group; it's a good example of how a practical benchmark and a subtle confound can coexist. I'd cite it if I were writing about multimodal brain encoding.","headline":"A solid, useful brain-encoding benchmark for instruction-tuned MLLMs, with a real but addressable confound in the instruction-specific steering claim.","tokens_in":25387,"tokens_out":3502,"would_cite":true,"duration_ms":38933,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Instruction-tuned multimodal LLMs, prompted with natural-language instructions, predict human visual brain activity significantly better than vision-only models and on par with CLIP, with the prompting instruction selecting which visual…","keywords":["instruction tuning","multimodal large language models","brain encoding","fMRI","Natural Scenes Dataset","representational alignment","variance partitioning","vision-language models"],"falsifier":"A control that would settle the claim: take the same images and models and replace the ten task instructions with ten length-matched, semantically empty prompts (for example, matched nonsense strings or unrelated sentences). If the instruction-specific differences in normalized brain alignment persist under nonsense prompts, then the task-specific alignment is driven by generation statistics or prompt wording rather than by the semantics of the instruction. A complementary check would record fMRI while participants actually perform each of the ten tasks; if the region-wise instruction alignment found under passive viewing does not sharpen under active task performance, the claim that these instructions capture task-specific neural processing would be weakened.","tokens_in":24388,"feed_emoji":"🧠","tokens_out":8072,"duration_ms":71407,"temperature":0.7,"pith_summary":"This paper asks whether instruction-tuned multimodal large language models (MLLMs), which answer natural-language questions about images by generating text, can predict human visual brain activity. The authors prompt three MLLMs (InstructBLIP, mPLUG-Owl, IDEFICS) with ten instructions spanning six visual tasks, obtain an image representation from each prompt by averaging the hidden states of the model's generated output tokens, and fit voxel-wise ridge regression encoding models to fMRI responses from participants watching natural scenes. They find that instruction-tuned MLLM representations predict brain activity significantly better than a vision-only model (ViT-H) and on par with CLIP, even though CLIP is given ground-truth captions while the MLLMs only see their own generated text. A reader should care because a single model, steerable by language, could replace collections of task-specific models in brain encoding, and instruction identity could act as a probe for which visual properties the brain's regions compute. The paper also shows the limits of that steerability: captioning dominates, most explained variance is shared across instructions, and several instructions do not differentiate brain-alignment patterns.","feed_headline":"Prompted multimodal AI rivals CLIP at predicting brain activity","feed_subtitle":"With ten natural-language prompts, instruction-tuned MLLMs beat vision-only models and match CLIP on fMRI prediction.","key_machinery":"The load-bearing object is the instruction-conditioned image representation: for each image and each of ten natural-language instructions, the model generates text tokens zero-shot, and the hidden states of those generated tokens are averaged and PCA-reduced to 1024 dimensions. That embedding is fed to voxel-wise bootstrap ridge regression encoding models trained to predict fMRI responses, with performance scored as Pearson correlation normalized by an estimated cross-subject prediction ceiling. Variance partitioning, which fits separate and joint encoding models for pairs of instructions and uses set arithmetic to split unique from shared explained variance, is what lets the paper claim that captioning subsumes most of the other instructions' signal. The instruction set itself, ten prompts borrowed from the MultiInstruct task taxonomy covering captioning, question answering, relationships, reasoning, understanding, and scene recognition, is the manipulator that makes a single model act as many task-specific encoders.","core_discovery":"On the paper's own terms, the discovery is that prompting an instruction-tuned MLLM with different natural-language instructions changes the brain alignment of its representations in a structured way. Averaged over ten instructions, all three MLLMs significantly outperform randomly initialized versions and the unimodal ViT-H model, and perform comparably to CLIP-Text across the whole visual cortex and in five visual ROIs. The instruction that wins varies by region: image captioning aligns best in high-level visual areas (EBA, PPA, FFA), while image-understanding instructions such as 'describe the most dominant color' align best in early visual regions. Grouping instructions into visual concepts, the MLLMs capture count-related and recognition-related concepts with clear brain alignment, but color, positional understanding, and general scene understanding produce similar alignment patterns regardless of the specific concept. Variance partitioning shows that most of the explained variance is shared between image captioning and the other instructions, with captioning acting as an umbrella category; the authors interpret the large shared variance as identifying where instruction-tuned MLLMs could improve at differentiating tasks.","pith_inferences":["If instruction prompts genuinely select distinct neural computations, prompt variation becomes a cheap experimental probe: rotating instructions on a single MLLM could generate hypotheses about which cortical region computes which visual property, testable against targeted fMRI experiments.","The shared-variance portrait suggests a testable design for the next generation of instruction-tuned models: training to minimize representational overlap across instructions should reduce the shared explained variance in brain responses, and if the mapping from instructions to neural processes is real, it should increase unique variance in the corresponding ROIs.","The study cannot separate task semantics from surface phrasing with only ten hand-picked prompts; a follow-up using multiple paraphrases per task would show whether the instruction-specific alignment tracks the task or the wording.","Because the fMRI data come from passive viewing, the alignment measured here is instruction-to-observation, not instruction-to-behavior; recording brain activity while subjects actually perform the instructed tasks could reveal stronger or different task-specific alignments."],"forward_implications":["Instruction-tuned MLLMs can be used off the shelf as brain encoders: one model, prompted with natural language, predicts visual fMRI as well as a model trained on ground-truth image-caption pairs, and no task-specific fine-tuning is needed.","The instruction itself acts as a selection knob for brain regions: captioning-style prompts best predict high-level visual areas (EBA, PPA, FFA), while color-understanding prompts best predict early visual cortex.","The large shared variance between image captioning and other instructions quantifies how much instruction-tuned MLLM task-following still collapses into generic captioning, giving a concrete metric for future instruction-tuning improvements.","Because instruction-tuned models beat a non-instruction-tuned multimodal model (BLIP-2) at the same task, the brain-alignment gain is attributed to instruction tuning itself rather than to the underlying vision-language backbone."],"supporting_citations":[{"why":"Supplies the Natural Scenes Dataset: fMRI recordings of participants viewing natural images, plus the ROI definitions used for region-level analysis.","marker":"(Allen et al., 2022)"},{"why":"Contributes the cross-subject prediction accuracy estimation used to normalize brain alignment scores.","marker":"(Schrimpf et al., 2021)"},{"why":"Provides the bootstrap ridge regression voxel-wise encoding framework used to predict fMRI responses.","marker":"(Deniz et al., 2019)"},{"why":"Introduces InstructBLIP, one of the three instruction-tuned MLLMs whose representations are evaluated.","marker":"(Dai et al., 2023)"},{"why":"Introduces mPLUG-Owl, the second instruction-tuned MLLM evaluated.","marker":"(Ye et al., 2023)"},{"why":"Introduces IDEFICS, the third instruction-tuned MLLM evaluated.","marker":"(Laurençon et al., 2023)"},{"why":"Provides CLIP, the non-instruction-tuned multimodal baseline that MLLMs are shown to match.","marker":"(Radford et al., 2021)"},{"why":"Supplies the MultiInstruct task taxonomy from which the ten natural-language instructions are borrowed.","marker":"(Xu et al., 2023)"},{"why":"Supplies the variance partitioning approach used to separate unique and shared explained variance across instruction pairs.","marker":"(de Heer et al., 2017)"},{"why":"Establishes the prior result that multimodal models beat vision-only models in high-level visual cortex, which this paper extends to instruction-tuned models.","marker":"(Wang et al., 2022)"}],"fun_headline_variants":["Prompted MLLMs beat ViT, match CLIP on brain alignment","Instructions alter MLLM-brain alignment across visual cortex","MLLM prompts align to brain by region and task","Task prompts change how MLLMs predict brain activity","Prompt-specific MLLM embeddings mirror brain encoding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The analysis assumes that averaging the hidden states of the text tokens an MLLM generates under a given instruction yields a faithful, instruction-specific representation of the image, so that differences in brain alignment across instructions reflect genuine task-specific neural processing rather than artifacts of generation length, language-model priors, or the exact wording of the ten chosen prompts.","fun_headline_variants_meta":{"raw":{"variants":["Prompted MLLMs beat ViT, match CLIP on brain alignment","Instructions alter MLLM-brain alignment across visual cortex","MLLM prompts align to brain by region and task","Task prompts change how MLLMs predict brain activity","Prompt-specific MLLM embeddings mirror brain encoding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000425,"raw_usage":{"total_tokens":2240,"prompt_tokens":1069,"completion_tokens":1171,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":685,"completion_tokens_details":{"reasoning_tokens":1090}},"tokens_in":685,"tokens_out":1171,"duration_ms":13004,"temperature":1.0,"reasoning_tokens":1090,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:00:22.867966+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A control that would settle the claim: take the same images and models and replace the ten task instructions with ten length-matched, semantically empty prompts (for example, matched nonsense strings or unrelated sentences). If the instruction-specific differences in normalized brain alignment persist under nonsense prompts, then the task-specific alignment is driven by generation statistics or prompt wording rather than by the semantics of the instruction. A complementary check would record fMRI while participants actually perform each of the ten tasks; if the region-wise instruction alignment found under passive viewing does not sharpen under active task performance, the claim that these instructions capture task-specific neural processing would be weakened.","supporting_citations":[{"cited_title":"Natural language supervision with a large and diverse dataset builds better models of human high-level visual cor- tex","cited_arxiv_id":null,"evidence_quote":"Establishes the prior result that multimodal models beat vision-only models in high-level visual cortex, which this paper extends to instruction-tuned models."}],"review_version":1}