{"id":"04ec247b-9021-4703-bd0a-e7a06e010337","arxiv_id":"2507.10442","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"By probing visual, projection, and response representations, the authors find that most VLM visual knowledge loss for recognition and counting occurs in the language decoder, while spatial understanding is lost in the visual encoder.","lead":"This paper probes the internal layers of popular vision-language models to find where basic visual abilities such as object recognition, counting, and spatial relations are lost. It reports that for recognition and counting, the visual encoder and alignment layers already contain the right information, but the language decoder fails to output it, while spatial understanding fails earlier in the visual encoder.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Response-space probe uses mean-pooling over all generated tokens; a token-level or non-linear probe could recover the 'lost' information, weakening the decoder-bottleneck attribution.","rationale":"The paper makes a genuinely useful contribution by locating failures at the module level, and I have no reason to doubt the raw response accuracies. The strongest and most novel claim, however, is the attribution: that the visual/VL spaces contain the answer while the response space loses it. That attribution is exactly where the average-pooled linear probe is least secure. The paper's control experiment is necessary but insufficient because it only checks that the probe does not overfit the labels; it does not check that the probe's readout is capable of extracting the task information. Since the entire 'Response Wide Shut' argument is a comparison of readouts across spaces, an asymmetric readout choice can manufacture the observed gap. The proposed test is cheap, direct, and would either substantiate or correct the attribution. Accordingly, the conditional verdict is appropriate; no verdict change is needed, but the authors should add this analysis before treating the bottleneck claim as established.","tokens_in":21246,"tokens_out":5654,"duration_ms":64528,"concrete_test":"Re-run the response-space evaluation on LLaVA-1.5 and LLaVA-NEXT for Stanford Dogs and CUB with three alternative readouts on the same decoder outputs: (1) the hidden state at the final generated token; (2) the hidden state at the first answer position under forced decoding of the class label; (3) a 2-layer MLP (non-linear probe) on the average-pooled features. Also extract the logits for the class-label tokens at generation step 1. If any readout reaches accuracy comparable to the visual probe (>80%), the 'information loss' conclusion is a probing artifact and the paper must refine its bottleneck attribution; if all remain near the reported response-probe accuracies, the central claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4 average-pools the T×F language-decoder output before linear probing. For autoregressive LLMs, class information is typically concentrated in the answer token; mean-pooling over all generated tokens dilutes the signal and can make a decoder that preserves the information look like it has lost it. The shuffled-label control only rules out memorization, not readout mismatch. In contrast, the visual/VL-projection probes use average-pooled features with a readout standard for those encoders, so the cross-space comparison is not apples-to-apples. If the response-space signal is token-localized or non-linear, low probe accuracy is an artifact, and the claim that 'knowledge does not translate to the response space' becomes the weaker claim that the decoder cannot verbalize the answer despite having the information.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a three-space analysis of vision-language models: linear probes on the visual encoder output, the VL projection output, and the language decoder output, compared against the free-form textual response. Using PaintSkills, Pascal VOC, Stanford Dogs, and CUB, the authors find that for coarse and fine-grained recognition and counting, probe accuracy in the visual and VL-projection spaces is high (often >90%) while response-space probe accuracy and text accuracy drop sharply, especially for fine-grained recognition. For spatial understanding the trend reverses: visual and projection probes are weak (~50%) while the response space is relatively stronger for some models. The paper also studies robustness to corruptions and background/visual-prompting transformations, and hypothesizes that the response-space drop stems from ineffective joint fine-tuning of the projection and decoder. The three-space methodology and the broad model/task coverage are valuable, but the central claim relies on a response-space probe whose readout (average pooling over generated tokens) is likely to discard the token-localized information in autoregressive decoder outputs.","tokens_in":21335,"tokens_out":8776,"duration_ms":90923,"significance":"If the central claim holds, it would challenge the common attribution of VLM visual failures to the visual encoder (Tong et al., 2024) and redirect efforts to the projection/decoder alignment stage for most tasks. The paper contributes a reusable three-space evaluation protocol, open-source model coverage, a shuffled-label control, and detailed per-corruption/transformation results. However, the significance is contingent on the validity of the response-space probe; the current average-pooling readout is not a faithful information measure for autoregressive decoders, so the paper's headline conclusion is not yet established.","major_comments":[{"comment":"The response-space probe averages the T×F language-decoder output across all generated tokens before fitting a linear classifier. For autoregressive LLMs, task-relevant information is typically concentrated in the answer token (or a small set of tokens), and mean pooling over the full generated sequence—including prompt continuations, filler tokens, and separators—dilutes this signal. The shuffled-label control in Appendix A.3 rules out label memorization by the probe, but it does not address readout mismatch: a probe that cannot access the informative token will report low accuracy even when the decoder output contains the information. The visual and VL-projection probes, in contrast, use average-pooled features with a readout that is standard and near-lossless for those encoders. The comparison is therefore not apples-to-apples, and the low response-space probe numbers in Tables 2 and 4 cannot support the claim that 'knowledge does not translate to the response space' until a token-level probe (e.g., probing the final token or the answer span) or a nonlinear probe is added and the drop persists.","section":"§3.4, Response Space; Tables 2 and 4"},{"comment":"The spatial-understanding conclusion that the visual encoder is the bottleneck rests on the same average-pooling readout. For a task defined by relative position (left/right/above/below), mean pooling over image patches discards the spatial layout that the task requires. The near-chance visual-probe accuracy in Table 5 (47-50%) may therefore understate the information available in the visual encoder. Before asserting that 'the visual encoder (i.e., CLIP) is responsible for the loss of information' (§4.2), the authors should probe the visual space with a readout that preserves position (e.g., probing patch tokens or using positional features), or explicitly caveat that the conclusion is conditional on the pooling-based protocol.","section":"§4.2, Table 5"},{"comment":"The paper interchangeably compares supervised probe accuracy (intermediate spaces) with zero-shot text generation (response space). These are different evaluation protocols: the probe is fit with labels on the task, while the text response is generated with a VQA prompt and matched via string/fuzzy matching. Low text accuracy can result from instruction-following failures or decoding choices rather than absence of information; the authors themselves note in Appendix A.2 that adding 'answer in one word' reduced performance for BLIP2/InstructBLIP. To support the 'information is lost' framing, the paper should either use a supervised readout for the text channel (e.g., a probe on the generated answer tokens) or restrict the conclusion to a 'verbalization failure' rather than 'knowledge loss.'","section":"§4.1, Tables 1-5"}],"minor_comments":[{"comment":"The phrase 'We conjuncture that these spaces' should read 'We conjecture that these spaces.'","section":"Section 3"},{"comment":"The heading 'Course-grained' is a typo for 'Coarse-grained.'","section":"Table 1"},{"comment":"The numbered summary list jumps from item 2 to item 4; item 3 is missing.","section":"Section 5.3"},{"comment":"The sentence 'For BLIP2 and InstructBLIP model we use the same prompt i.e., What is the central object...' lacks punctuation, and the phrase 'continuous outputs' is ambiguous; please clarify.","section":"Appendix A.2"},{"comment":"The phrase 'sinces it is similar' should be 'since it is similar.'","section":"Appendix A.6"},{"comment":"The statement 'We use COCO (Malik et al., 2024)' mis-credits the COCO dataset; the reference should distinguish the COCO dataset from the ObjectCompose-derived masks or cite the appropriate COCO source.","section":"Section 5.2"},{"comment":"The legend uses 'Text space' while the paper elsewhere uses 'VL Proj.'; please align the terminology.","section":"Figure 4"},{"comment":"The shuffled-label control is reported for LLaVA-NEXT only; consider reporting it for at least one additional model family (e.g., InstructBLIP) since the probing conclusions are drawn across seven models.","section":"Appendix A.3"}],"recommendation":"major_revision","confidential_remarks":"The central measurement issue (average pooling of decoder outputs) is fixable in revision and the paper is otherwise a solid empirical study. If a token-level probe confirms the drop, the paper will make a strong contribution. I do not recommend rejection; the scope of the paper is appropriate for the venue, but the headline conclusion should be re-evaluated after the readout is corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the useful part: this paper does something simple that nobody had done cleanly—it evaluates the same visual tasks at three points in a VLM (visual encoder, VL projection, decoder output) across seven open models and finds that for fine-grained recognition and counting, probe accuracy stays above 90% through the visual and projection spaces and then collapses in the response space, while for spatial understanding the pattern reverses. The spatial reversal is a genuinely new observation and it lands because they control for language priors with a blind LLM baseline and include a shuffled-label probe control. The robustness and background-transformation sections are a bonus, not the core, but they are consistent with the main story.\n\nThe soft spot is the response-space probe. They average-pool over all T generated tokens before linear probing. For an autoregressive decoder, class information typically concentrates in the answer token; mean-pooling dilutes it and can make a decoder that preserves the information look like it has lost it. The shuffled-label control only rules out probe memorization, not readout mismatch. So the cross-space comparison is not apples-to-apples. That said, the paper also reports zero-shot text accuracy, which is low for fine-grained tasks, so the broad phenomenon—the decoder output is poor at fine-grained recognition—is not an artifact. What the probe flaw weakens is the precise attribution: the claim that 'knowledge does not translate to the response space' should be softened to 'the decoder cannot reliably verbalize the answer under this readout.' The authors themselves hedge by calling the joint-fine-tuning explanation a conjecture, which is fair, but the abstract and intro state the stronger version.\n\nMinor issues: the fine-grained datasets are subsampled to 15 classes with no error bars, and the causal story about 0.17% fine-grained samples in LLaVA-665k is suggestive but not tested by a counterfactual.\n\nOverall: this is a serious empirical study with an important central observation and a methodological wrinkle that a careful referee could push the authors to fix. It belongs in peer review, not at the desk. If you work on VLM internals, read it; if you plan to cite the attribution, wait for the revision.","headline":"A systematic three-space probe of VLMs that finds a plausible decoder bottleneck for fine-grained recognition, but the response-space readout is too crude to pin the blame.","tokens_in":21880,"tokens_out":2899,"would_cite":false,"duration_ms":29747,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that basic visual knowledge is present in early VLM representations but mostly fails to reach the text output, and that the joint fine-tuning of the projection and decoder is the weakest link.","keywords":["vision-language models","linear probing","intermediate representations","fine-grained recognition","object counting","spatial understanding","VLM robustness","instruction tuning"],"falsifier":"Train a multi-layer probe or a per-token probe on the language decoder's hidden states for the Stanford Dogs or CUB fine-grained task. If such a probe exceeds roughly 90 percent accuracy while the model's text answer stays near 30 percent, then the paper's 'knowledge loss' is an artifact of its linear average-pooled readout; if every readout stays low, the information is truly blocked before the decoder.","tokens_in":21040,"feed_emoji":"👁️","tokens_out":6872,"duration_ms":70731,"temperature":0.7,"pith_summary":"This paper tries to locate, inside a vision-language model, where basic visual abilities break down. The authors compare three stages of the model—the visual encoder's latent space, the vision-language projection space, and the language decoder's response—on object recognition, fine-grained recognition, counting, and spatial reasoning. They find that for recognition and counting, linear probes read the correct answer out of the early visual and projection spaces with accuracy above 90 percent, while the model's actual text response often falls to 10–35 percent. For spatial relations the pattern reverses: the early spaces are weak and the response is comparatively stronger. The paper concludes that the bottleneck for most basic visual tasks is not the visual encoder but the transfer of already-encoded knowledge into the decoder's output, and that this points to the joint fine-tuning stage as the target for improvement.","feed_headline":"Visual knowledge vanishes at the language decoder","feed_subtitle":"Probing seven VLMs shows early layers hold the answer while text accuracy drops to 10–35% on fine-grained recognition.","key_machinery":"The central object is the three-space diagnostic: the output of the visual encoder, the output of the vision-language projection, and the output of the language decoder. A linear logistic-regression probe is trained on average-pooled features from each space, with a shuffled-label control task verifying that probe accuracy reflects encoded information rather than probe memorization; the response space is evaluated both by such a probe and by the model's own text output. This decomposition is what lets the paper attribute failures to the translation between components rather than to any single component.","core_discovery":"The paper's central discovery is that current open vision-language models already encode the information needed for fine-grained recognition and object counting in their visual and vision-language projection spaces—linear probes exceed 90 percent accuracy there—yet this information does not reach the final text response, where accuracy on fine-grained recognition can fall to 10–35 percent. This contradicts the idea that the visual encoder is the main source of these failures. Spatial understanding is the exception: there the visual and projection representations themselves are weak, so improving the visual encoder is the right direction. The authors attribute the recognition and counting gap to ineffective joint fine-tuning of the projection layer and language decoder, supported by the observation that instruction-tuned models transfer better than the non-instruction-tuned BLIP-2, and that scaling the decoder only partially closes the gap.","pith_inferences":["If this diagnosis is right, then adding a modest number of fine-grained examples to the alignment or instruction-tuning stage is a cheap testable fix: response-space accuracy should jump without changing the visual encoder.","The same three-space probing design could sort encoder-side from decoder-side causes for other reported VLM failures, such as negation, compositional reasoning, and hallucination.","Because the probes are supervised while the text evaluation is zero-shot, part of the measured gap may be a measurement asymmetry rather than lost information; token-level or nonlinear readouts would separate the two.","The spatial reversal suggests a two-part design rule: improve encoders for geometry and improve alignment training for semantics."],"forward_implications":["Improving the joint fine-tuning of the projection layer and language decoder, with more fine-grained examples in the alignment data, should transfer the high-accuracy early-space knowledge into VLM text responses.","Scaling the language decoder alone is not a remedy: larger decoders improve response-space accuracy only partially, leaving a large gap on fine-grained recognition.","For spatial understanding, the visual encoder is the information bottleneck, so better spatial encoders or spatial training data are needed rather than decoder-side fixes.","Visual prompting effects such as clean backgrounds and reverse blur are absorbed mostly by early spaces; the response space benefits less, indicating that alignment loss is a separate target for improvement.","The response space's apparent robustness to image corruptions is partly a side effect of information not flowing into the response, so robustness numbers should be interpreted with that in mind."],"supporting_citations":[{"why":"Supplies the opposing claim that VLM failures stem from the visual encoder, which this paper tests and reverses for recognition and counting.","marker":"Tong et al., 2024"},{"why":"Provides the LLaVA architecture and the 665k instruction-tuning dataset whose sparse fine-grained content the paper uses to explain the response-space gap.","marker":"Liu et al., 2023"},{"why":"Provides BLIP-2 and its Q-former projection, the non-instruction-tuned baseline showing high visual and projection accuracy but weak text responses.","marker":"Li et al., 2023a"},{"why":"Provides InstructBLIP, the instruction-tuned counterpart whose improved response-space accuracy supports the joint-fine-tuning explanation.","marker":"Dai et al., 2023"},{"why":"Documents VLMs' weakness on fine-grained visual concept recognition, the phenomenon this paper localizes to the response space.","marker":"Kim and Ji, 2024"},{"why":"Earlier comparison of LLM-decoder classification output with CLIP that this work extends to three intermediate spaces across more tasks.","marker":"Zhang et al., 2024"},{"why":"Supplies the PaintSkills diagnostic dataset and the counting and spatial task design used in the experiments.","marker":"Cho et al., 2023"},{"why":"Supplies the shuffled-label control-task methodology that validates the linear probes as measures of encoded information.","marker":"Hewitt and Liang, 2019"}],"fun_headline_variants":["Visual knowledge is trapped before the language decoder","VLMs see but don't speak: fine-grained info lost at decoder","The bottleneck is the decoder, not the encoder","Probes find knowledge that VLM text fails to deliver"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The diagnosis assumes that a linear probe on average-pooled features measures the information genuinely available to the next component, so a probe that cannot read the answer may be a limitation of the readout rather than an absence of information.","fun_headline_variants_meta":{"raw":{"variants":["Visual knowledge is trapped before the language decoder","VLMs see but don't speak: fine-grained info lost at decoder","The bottleneck is the decoder, not the encoder","Probes find knowledge that VLM text fails to deliver"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000711,"raw_usage":{"total_tokens":3157,"prompt_tokens":858,"completion_tokens":2299,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":2234}},"tokens_in":474,"tokens_out":2299,"duration_ms":19864,"temperature":1.0,"reasoning_tokens":2234,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:30:47.965312+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a multi-layer probe or a per-token probe on the language decoder's hidden states for the Stanford Dogs or CUB fine-grained task. If such a probe exceeds roughly 90 percent accuracy while the model's text answer stays near 30 percent, then the paper's 'knowledge loss' is an artifact of its linear average-pooled readout; if every readout stays low, the information is truly blocked before the decoder.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the LLaVA architecture and the 665k instruction-tuning dataset whose sparse fine-grained content the paper uses to explain the response-space gap."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the PaintSkills diagnostic dataset and the counting and spatial task design used in the experiments."}],"review_version":1}