BLIP Open-Source Causal Tracing Tool

BLIP open-source causal tracing tool: Palit et al. (2023) adapt Meng et al.'s causal intervention to BLIP VQA and release their tracer code. Prisma is separate.

In Palit, Pandey, Arora, and Liang, arXiv:2308.14179v1 (27 Aug 2023), the adaptation claim and the open-source claim are not the same object. The paper states that it adapts a unimodal causal tracing tool, specifically, the causal intervention method of Meng et al. (2022), which that paper describes as clean, corrupted, and patched runs on GPT-2 XL and GPT-NeoX with noise in subject-token embeddings. It states that the implementation follows Meng et al. most closely, then changes the corrupted input to BLIP image embeddings for visual question answering. Separately, it states that it releases its own BLIP causal tracing tool and fully open-sources its code. A visualizer and adaptations to other vision-language models are described as forthcoming (“will soon release”), not as already released. The extracted v1 text says the code is available at “this URL” but does not print a destination.

Joseph et al., arXiv:2504.19475v3 (3 Jun 2025) is a different paper. In Prisma’s interpretability library it lists circuit-analysis support for ablations via forward-pass interventions, HookedSAEViT sparse-feature circuit analysis by replacing layers with multiple sparse autoencoders (SAEs) or sparse coders, and HookedViT integration with ACDC for automatic circuit discovery. The visualization tools it lists are attention-head visualization and logit lens, also described as logit lens and attention visualization for intermediate activations. Activation caching is listed as infrastructure for circuit analysis, not as a separate discovery algorithm. Nothing in the extracted Prisma v3 text says these components implement, include, or replace the BLIP tracer, and the extracted BLIP v1 text does not mention Prisma.

The keyword “open-source interpretability tool BLIP” collapses three different referents: the BLIP vision-language model cited as Li et al. (2022), the 27 Aug 2023 causal-tracing paper that adapts a unimodal intervention to that model, and later open-source vision tooling. Palit et al. state both that they “adapt a unimodal causal tracing tool to BLIP” and that they “release our BLIP causal tracing tool as open source.” Those sentences name a borrowed procedure and a new artifact. Prisma v3 then lists its own circuit-analysis and visualization tools for vision and video transformers. Using one paper’s software claim as if it were the other’s changes both the intervention and the codebase.

This article separates those statements, then specifies the intervention Palit et al. say they adapted, the release language in v1, and the circuit-analysis and visualization components Prisma actually names. It does not treat either library as a substitute for the other. Dates below for the two papers are the version dates printed on the PDFs. Later revisions, hyperlink targets, and repository contents beyond those PDFs are not in the evidence used here.

Key takeaways

  • Palit et al. state that they adapt Meng et al.’s unimodal causal intervention, clean run, corrupted run, and patching of intermediate states, not that they are publishing Meng et al.’s code. The change they specify is corruption of BLIP image embeddings, with patching in the question encoder and answer decoder, rather than noise in subject tokens of GPT-2 XL or GPT-NeoX.
  • The open-source object stated as available in v1 is their BLIP causal tracing code. The visualizer and ports to other vision-language models are prospective. The extracted PDF does not include the code URL.
  • Prisma lists forward-pass ablations, HookedSAEViT layer replacement for sparse-feature circuits, ACDC integration through HookedViT, attention visualization, and logit lens. That list is not a claim that Prisma is the BLIP tracer.
  • https://www.anthropic.com/research/open-source-circuit-tracing prints PublishedMay 29, 2025 and describes an attribution-graph release. https://www.anthropic.com/research/donating-open-source-petri prints PublishedMay 7, 2026 and describes the Petri alignment-test toolbox. Those pages are not evidence for what the tracing paper adapted or released, and they are not evidence that Prisma implements the BLIP tracer.
What each paper states, and what that statement does not establish. This is a claim separation, not a performance comparison and not a substitution of one tool for the other.
Statement Where it is stated What the paper states What it does not establish
Adapted from unimodal causal tracing https://arxiv.org/pdf/2308.14179v1 prints arXiv:2308.14179v1 and the date 27 Aug 2023. The adaptation wording is in the abstract, the related-work account of Meng et al., and section 3. The abstract says they “adapt a unimodal causal tracing tool to BLIP.” Section 3 says they “adapt the causal intervention method from Meng et al.” Related work says the implementation “follows this work most closely,” after describing Meng et al.’s three runs on GPT-2 XL and GPT-NeoX: clean input, corrupted input produced by noise in subject tokens, and patching of layer outputs from the clean run into the corresponding layers of the corrupted run, to measure causal relevance for factual knowledge. It does not state that the released code is Meng et al.’s repository. It does not keep the corrupted object as subject-token embeddings. It does not name a public unimodal repository that was forked.
Released as open source Same paper, abstract and conclusion The abstract says “we release our BLIP causal tracing tool as open source” and “Our code is available at this URL.” The conclusion says “we fully open source our code and will soon release a visualizer as well as adaptations to other vision-language models.” The extracted v1 text does not include the URL target. A visualizer and other vision-language adaptations are not stated as already released. “The first causal tracing tool for a vision-language model” is the authors’ priority claim, not an independent check.
Prisma circuit-analysis and visualization tools Joseph et al., arXiv:2504.19475v3, 3 Jun 2025; abstract, section 1, and section 3.1.2. Codebase URL printed in the paper: github.com/Prisma-Multimodal/ViT-Prisma Circuit analysis: ablations via forward-pass interventions; HookedSAEViT sparse-feature circuit analysis by layer replacement with multiple SAEs or sparse coders; HookedViT integration with ACDC. Visualization: attention-head visualization and logit lens, also “logit lens and attention visualization for intermediate activations.” Activation caching is listed with those tools as support for circuit analysis. v3 does not state that these components are the BLIP causal tracer, that they include BLIP, or that they replace Palit et al.’s tool. Activation patching and attribution patching are discussed as field techniques in related work, not named as Prisma library modules in the tool bullets.

What the BLIP paper states that it adapts

Palit et al. use three wordings, and they should be kept distinct. The abstract says they “adapt a unimodal causal tracing tool to BLIP to enable the study of the neural mechanisms underlying image-conditioned text generation.” The introduction places that effort in causal mediation analysis, citing Pearl, and says causal mediation analysis “has so far been implemented only for the unimodal language domain,” with the citation pointing to Meng et al. Section 3 then says: “We adapt the causal intervention method from Meng et al. [27] to investigate visual question answering (COCO-QA) in a vision-language model (BLIP).”

The unimodal procedure it names

The procedure they say the implementation follows most closely is Meng et al., “Locating and editing factual associations in GPT” (their reference [27]; NeurIPS 2022, as cited). They also discuss Vig et al. as an earlier application of causal mediation analysis to GPT-2 for grammatical gender bias, and they cite probing work as insufficient to show that a model uses the information a probe finds. The “follows this work most closely” sentence is attached to Meng et al., not to Vig et al.

As Palit et al. describe Meng et al., the unimodal method corrupts token-embedding inputs by introducing noise into a sentence’s subject tokens, on GPT-2 XL and GPT-NeoX, to measure causal relevance of states for factual knowledge. The model is observed in three runs: a clean-input run, a corrupted-input run, and an intervention that patches layer outputs from a clean run of the same sentence into the corresponding layer outputs of the corrupted run. This article reports that description only as Palit et al. state it. Meng et al.’s paper is not itself in the evidence used here, so no further details of that method, noise distribution, layer set, or metrics, are added from outside this citation.

What they change for BLIP

The adaptation is not a claim that subject-token corruption is reused unchanged. Section 3.1 states that BLIP takes a pre-processed image and question tokens and returns a single-word answer. The authors corrupt the image embeddings before they are fed into the question encoder, which they say yields an incorrect output, then try to restore the correct answer by patching individual intermediate states of a clean run into the corrupted run. States that produce the greatest answer improvement are treated as causally relevant. That last sentence is their criterion for calling a state causally relevant, not a separate statistical test reported in the extracted text.

Figure 2’s caption describes the architecture they intervene on: an image-patch embedding is fed into the question encoder with question tokens; cross-attention produces image-conditioned question embeddings; those embeddings go to the answer decoder. They attribute this BLIP-for-VQA architecture to Li et al. (2022). The tracing paper does not name a checkpoint, weight hash, or parameter count for that pretrained VQA model. Those identifiers are unknown from the extracted v1 text and are not inferred here from any other model.

The corruption implementation is more specific than “add noise,” and the paper’s own phrases do not all use the same notation. It says the second image embedding of the batch is corrupted by adding noise to all 577 patch embeddings, producing clean/corrupted pairs (E, E*). It also says that for each image they sample one noise instance ε ∼ N(1, ν), with ν an adjustable standard deviation, and multiply ε onto each patch embedding. The corrupted encoding is passed into the question encoder with the question tokens. Patching overwrites the output of each state, layer L, token T, of the corrupted question-encoder run with the corresponding state from the clean run. The same intervention is replicated for the answer decoder. Figure 1’s caption is a schematic of that contrast: a clean image encoding cross-attends with the question and yields “brown”; the corrupted encoding yields an incorrect answer; an intermediate state is patched from clean to corrupted to observe the effect on answer probabilities. That caption is an illustration, not the 200-sample result.

What v1 states that it releases

The release claim is about their implementation, not about the unimodal tool they say they adapted. The abstract’s release sentence is: “we release our BLIP causal tracing tool as open source to enable further experimentation in vision-language mechanistic interpretability by the community. Our code is available at this URL.” The conclusion repeats the scope: “Towards this end, we fully open source our code and will soon release a visualizer as well as adaptations to other vision-language models.” They also write that they “introduced the first causal tracing tool for a vision-language model.” That priority sentence is an author claim. This article does not independently verify it, and it does not treat it as a description of Prisma.

Release boundary in v1. On the 27 Aug 2023 text, the stated open-source object is the BLIP causal tracing code. The visualizer and adaptations to other vision-language models are future tense. In the extracted PDF, “this URL” has no printed destination, so v1 as extracted cannot be cited for a repository address. Whether a later version or the original hyperlink supplies one is unknown here.

How the adapted intervention is scored

The metric is part of what “adapting causal tracing” means in this paper, because they define causal relevance by restoration of answer probability rather than by a probe correlation. Given question embedding Q, image pair (E, E*), and correct answer A, they compare the probability of A in the corrupted run with its probability in the restored run, and normalize by the gap between the clean and corrupted probabilities.

Equation (1) is that normalized recovery at layer L and token T. In the extracted PDF the two differences are concatenated without a visible division sign. The prose defines a ratio: restored-minus-corrupted, divided by clean-minus-corrupted. With that reading, and with patchL,T(E, E*) the run in which the clean state at (L, T) replaces the corrupted state:

ΓL,T = [p(A | patchL,T(E, E*), Q) − p(A | E*, Q)] / [p(A | E, Q) − p(A | E*, Q)]

No probability inputs for a particular image are printed in the extracted text, so this article does not compute a numerical example. The paper says it expects ΓL,T in [0, 1], where 0 is no improvement over complete corruption and 1 is full recovery of the clean answer probability. That range is their expectation. They later report negative values for a few settings of ν, so the expectation is not a guarantee that every estimate falls in [0, 1]. Darker cells in their heatmaps are defined as higher ΓL,T. Cell values themselves are not printed in the extracted text.

Equation (2) averages over layers and tokens as a function of the noise factor. The extracted expression is an average, Γ(ν) = 1/(|L| · |T|) Σl Σt Γl,t, with the parenthetical E* = νE. That parenthetical is not the same expression as ε ∼ N(1, ν) multiplied onto each patch. Both notations are printed in v1. This article does not choose one and rewrite the other.

Demonstration conditions

The task is COCO-QA, which they describe as open-ended answer generation from an image and a question, with one-word answers in four categories: object identification, object counting, colour identification, and location identification. They state that the dataset has 123,287 images and 78,736 train and 38,948 test questions, sourced from MSCOCO, citing Ren, Kiros, and Zemel (2015) and Lin et al. (2014). They then say they divided the training subset into three splits, colour, location identification, and counting, and assessed BLIP’s zero-shot performance on each. Object identification is named as a category but is not one of those three splits, and no accuracy for it is given. It is not inferred here.

80.23% color identification is the zero-shot accuracy Palit et al. report for BLIP’s pretrained VQA model on their color split of the COCO-QA training subset (Table 1, arXiv:2308.14179v1, 27 Aug 2023). The same table reports 26.30% for location identification and 3.27% for object counting. These are not stated as official COCO-QA test-set scores, and the split sizes are not printed.

They use the color split because they want mechanisms for a behavior the model already performs well. Their further analysis, which is author interpretation of errors, not the Table 1 percentages, is that counting accuracy is low because BLIP tends to emit digits rather than natural-language number words, and that location answers differ in structure from the dataset. Those format mismatches are reasons they give for not tracing those splits. They are not themselves causal-tracing results.

For the noise sweep they vary ν over [0.1, 30], averaged over 200 samples with 10 runs each. They do not measure Γ at ν = 0, because patching a clean run into a clean run would give Γ(ν) = 1 by construction. They describe a decaying curve from 0.1 to 30, little variation at extremely large ν, and negative values for a few ν. They choose ν = 5 to avoid noise so small that patching is trivial and noise so large that restoration is impossible. Figure 5’s caption states that it plots this average probability difference for the question encoder and the answer decoder. Axis ticks in the extracted figure text are not used here as data.

What the interventions showed, and what the authors infer

The reported intervention result is layer-local, not a full circuit. Heatmaps for COCOQA-ID458864 (“What is the color of the animal?”) and COCOQA-ID220218 (“What is the color of the character?”) are averages across 10 runs. A third heatmap averages 200 COCO-QA samples. Encoder and decoder layers are indexed from 0 to 11, and question tokens are the vertical axis. Figure 4’s caption says most of the causal relevance is concentrated in the final layers of the encoder and decoder blocks. The results text is more specific: in the question encoder, only the final layer (11), for all tokens, has a larger effect than earlier layers or tokens; in the answer decoder, layers 9 to 11 play the most apparent role. The abstract’s phrase “later layer representations for all tokens” is their summary of that pattern.

The next step in the paper is interpretation, and the authors mark it as such. They write that BLIP “does not benefit from restored access to the correct image embeddings until the final few layers.” They offer two readings: the vision modality may be irrelevant to computations until the final layer, with vision and language processed independently in intermediate layers; or final layers may override preceding layers that are still weakly causally relevant. Neither reading is identified by the intervention alone. The intervention shows where restoring a clean state changes answer probability under their noise model. It does not, on the paper’s own wording, decide between those two mechanisms.

The conclusion lists method limits rather than a completed circuit. The role of ν is unclear; they ask why increasing ν does not monotonically reduce performance. Restoration at a single site, rather than at multiple sites, does not show cross-module coordination. Identifying larger task mechanisms, comparable to unimodal circuit papers they cite, is described as future work, including checks on compositional benchmarks such as Winoground. Those are stated open problems, not results.

Circuit-analysis and visualization tools Prisma lists

Prisma v3 (Joseph et al., 3 Jun 2025; CVPR 2025 Workshop on Mechanistic Interpretability in Vision) describes an open-source framework for vision and video mechanistic interpretability. The printed codebase link is https://github.com/Prisma-Multimodal/ViT-Prisma. The abstract’s toolkit list is broader than circuit analysis: access to 75+ vision and video transformers; SAE, transcoder, and crosscoder training; 80+ pretrained SAE weights; activation caching, circuit analysis tools, and visualization tools; and educational resources. The circuit-analysis and visualization items are the ones the reader question asks for. The other items are named here only so they are not mistaken for that list. Repository contents beyond the v3 text were not checked.

Circuit analysis

Section 1 names “methods for circuit analysis” inside a suite that also includes attention-head visualization and logit lens. Section 3.1.2 is the inventory of implementations:

  • Ablations via forward-pass interventions. The Circuit Analysis bullet says the codebase supports these ablations. It does not, in that bullet, specify contrast-pair construction, metric, or a BLIP VQA setting.
  • HookedSAEViT. This module “enables sparse feature circuit analysis through layer replacement with multiple SAEs or sparse coders.” The stated mechanism is replacement of layers by sparse coders, not patching of a clean image embedding into a corrupted BLIP run.
  • ACDC through HookedViT. “The HookedViT module integrates with ACDC for automatic circuit discovery.” ACDC is cited via Conmy et al. Prisma claims integration, not authorship of ACDC. A separate citation, Rajaram et al., is discussed in related work as automatic discovery of visual circuits. v3 does not say that paper’s method is shipped as an additional named module beyond the ACDC integration.

Activation caching belongs next to this list but is not a fourth discovery algorithm. The model-registry section says HookedViT uses TransformerLens-inspired activation caching “for easy circuit analysis and SAE training.” The abstract lists “activation caching, circuit analysis tools, and visualization tools” in one phrase. Prisma also says it keeps TransformerLens naming conventions and hook interfaces so researchers can move between language and vision tooling. That is a compatibility claim. It does not state that TransformerLens itself is part of Prisma.

The hooks are described for HookedViT configurations drawn from Hugging Face, OpenCLIP, and timm, including ViT, CLIP-ViT, and DINO, plus video encoders ViViT and V-JEPA, and an adapted Kandinsky ViT encoder with a pretrained SAE for steering diffusion. The extracted v3 text does not name BLIP in that registry. Absence from the paper’s list is not a test of the GitHub tree.

Related work discusses activation patching and attribution patching as field techniques, and cites language tooling such as CircuitsVis, SAELens, and NNsight. Those discussions are not the library bullets. They are not added here to Prisma’s circuit-analysis list.

Visualization

Two visualization techniques are named, twice, with slightly different wording. The introduction lists “attention head visualization, and logit lens techniques,” citing nostalgebraist (2020) for logit lens. Section 3.1.2 says visualization “provides logit lens and attention visualization for intermediate activations.” This article treats those as the same two techniques, not as four tools.

The supplement adds a constraint on the attention-head visualizer, not a new tool name. A toy ViT with patch size 32 is included so attention heads can be inspected by that visualizer; patch size 16 heads are “too large to easily render in JavaScript.” That is a stated rendering limit for those toy checkpoints. It is not a specification of every attention plot in the library.

Prisma also states that vision and language differ in decoding: vision outputs “typically call for specialized decoding like cosine similarity or diffusion decoding instead of analyzing the next-token logit prediction.” v3 still lists logit lens as a visualization tool. The extracted pages name the technique and state that decoding difference, but they do not specify which readout, token logits, cosine similarity, or another projection, the vision logit-lens component uses. That implementation detail is unknown from this text.

Listed nearby, and not those tools

SAE evaluation tools measure sparsity (L0), reconstruction and substitution losses, and maximally activating images. Maximally activating images are under evaluation, not under the Visualization bullet, so they are not counted here as one of the visualization tools Prisma lists in that bullet. Training support for SAEs, transcoders, and crosscoders, the 80+ pretrained weights, tutorial notebooks, and toy ViTs are also library components. They are outside the circuit-analysis and visualization inventory above. SAE widths, L0 tables, and loss-change findings in Prisma are about those sparse coders. They are not measurements of the BLIP tracer and are not reused as if they were.

What neither paper establishes

The two papers do not identify a shared artifact. Palit et al. specify state patching in BLIP’s question encoder and answer decoder after image-embedding corruption, on a COCO-QA color split, with Γ as normalized answer-probability recovery. Prisma specifies forward-pass ablations, SAE or sparse-coder layer replacement, ACDC integration, attention visualization, and logit lens on a HookedViT interface aimed at vision and video transformers. Using one as a stand-in for the other is not supported by either extracted paper.

There is an unresolved scope tension, and it should stay unresolved. Palit et al. claim, as authors, the first causal-tracing tool for a vision-language model and say that tool’s code is open source. Prisma’s related work says that, aside from a circuit framework by Joseph and Nanda which Prisma says it extends with sparse-coder training, “no other open-source toolkits exist specifically for vision mechanistic interpretability.” The extracted Prisma text does not mention Palit et al., so it does not show that the BLIP tracer was considered and excluded. A BLIP VQA tracer is also not the same kind of object as a ViT registry with SAE training. This article does not adjudicate priority and does not convert either sentence into a substitution claim.

Other limits that affect reuse of the stated results:

  • Versions. The BLIP claims are from arXiv:2308.14179v1, 27 Aug 2023. The Prisma claims are from arXiv:2504.19475v3, 3 Jun 2025. Later PDFs, code, and docs are outside this evidence.
  • Checkpoint. The tracing experiments are attributed to BLIP’s pretrained VQA model. No checkpoint name or hash is printed. Layer indices 0–11 and 577 patch embeddings are experimental statements in v1, not a license to copy dimensions from a different BLIP or ViT checkpoint.
  • Metric range and noise notation. Expected Γ in [0, 1] coexists with reported negative values. ε ∼ N(1, ν) and E* = νE are both printed and are not reconciled here. Equation (1)’s division sign is supplied from the paper’s prose because the extracted equation string omits it.
  • Causal versus interpretive. High Γ at late layers is the intervention result they report. Independent modality processing, and override by final layers, are the authors’ alternative readings.
  • No replication in this evidence. Neither paper’s experiments were re-run here.

Other retrieved open-source texts are not these tools

The adaptation, release, and Prisma tool-list questions are answered by the two papers above. Two other retrieved pages name different objects. Dates in this subsection are the publication dates printed on those pages. Each page is cited only for a statement that page makes, not as a substitute for the BLIP causal tracer or for the circuit-analysis and visualization tools Prisma lists.

On 29 May 2025, Anthropic stated that it was open-sourcing “a new method to trace the thoughts of a large language model.” The post states that the method generates attribution graphs, which “(partially) reveal the steps a model took internally to decide on a particular output,” that the library “supports the generation of attribution graphs on popular open-weights models,” and that a frontend hosted by Neuronpedia lets users explore the graphs. It also states that the tools had already been used to study multi-step reasoning and multilingual representations in Gemma-2-2b and Llama-3.2-1b. That page describes a language-model attribution-graph release. It does not state an adaptation of Meng et al. to BLIP, a BLIP code release, or Prisma’s HookedViT inventory.

On 7 May 2026, Anthropic stated that in October 2025 it launched Petri, “an open-source toolbox of alignment tests that can be applied to any large language model,” to test for tendencies including “deception, sycophancy, and cooperation with harmful requests.” The post states: “We have handed over its development to Meridian Labs, an AI evaluation nonprofit.” As described on that page, Petri is an alignment-evaluation toolbox, not the BLIP causal tracing implementation and not a Prisma circuit-analysis or visualization module.

Your Action Plan

  1. Separate the adapted procedure from the released code. Cite Palit et al. for a transfer of Meng et al.’s three-run causal intervention, as that paper describes it, onto BLIP image embeddings. Do not cite v1 as a release of Meng et al.’s repository, and do not describe the corrupted input as subject tokens once the method has moved to the 577 patch embeddings.
  2. Bound the open-source claim by the v1 tense. The released object stated as available is the BLIP causal tracing code. Do not cite this version for a visualizer or for ports to other vision-language models. Do not invent a code URL: the extracted sentence stops at “this URL.” If a hyperlink or a later version supplies an address, record that version separately from v1.
  3. If you reproduce the demonstration, match the stated conditions and label the inference. The paper specifies a pretrained BLIP-for-VQA model, whose checkpoint is not named; the COCO-QA color split of the training subset, not the location or counting splits; multiplicative noise ε ∼ N(1, ν) on every patch embedding; ν = 5 after a sweep on [0.1, 30] over 200 samples and 10 runs; and one-site patches at (L, T) in the question encoder and the answer decoder. Compute Γ as the normalized probability gap in equation (1). Report late-layer concentration as the intervention pattern they describe. Do not report either of their “may mean” readings as what the patch established. Table 1’s 80.23%, 26.30%, and 3.27% are zero-shot accuracies on those training-subset splits, not tracing scores.
  4. Use Prisma’s named components without a substitution assumption. For circuit analysis, v3 supports forward-pass ablations, HookedSAEViT layer replacement with SAEs or sparse coders, and ACDC via HookedViT. For visualization, it lists attention visualization, including the attention-head visualizer, and logit lens. The paper’s codebase link is ViT-Prisma. Do not assume those hooks load BLIP, and do not treat activation patching or attribution patching as shipped modules unless a version newer than this v3 text says so. The vision logit-lens readout is unspecified in the extracted pages.
  5. Pin versions before comparing either claim to current code. https://arxiv.org/pdf/2308.14179v1 prints arXiv:2308.14179v1 and the date 27 Aug 2023. https://arxiv.org/pdf/2504.19475v3 prints arXiv:2504.19475v3 and the date 3 Jun 2025. A newer repository can add a visualizer, a BLIP hook, or a different noise model. That would be a new source, not a restatement of these two papers.
  6. Do not treat other open-source posts as these tools. https://www.anthropic.com/research/open-source-circuit-tracing prints PublishedMay 29, 2025 and describes an attribution-graph library. https://www.anthropic.com/research/donating-open-source-petri prints PublishedMay 7, 2026 and describes the Petri toolbox. Those pages do not alter the v1 adaptation or release wording, and they do not extend Prisma’s circuit-analysis or visualization list.

How We Sourced This

Primary pages used for the claims above: https://arxiv.org/pdf/2308.14179v1, which prints arXiv:2308.14179v1 and the date 27 Aug 2023; https://arxiv.org/pdf/2504.19475v3, which prints arXiv:2504.19475v3 and the date 3 Jun 2025; https://www.anthropic.com/research/open-source-circuit-tracing, which prints PublishedMay 29, 2025; and https://www.anthropic.com/research/donating-open-source-petri, which prints PublishedMay 7, 2026.

  1. What the BLIP paper states that it adapts
  2. What v1 states that it releases
  3. How the adapted intervention is scored
  4. Circuit-analysis and visualization tools Prisma lists
  5. What neither paper establishes
  6. Your Action Plan
Reading map: section order only, not a causal sequence or experimental result.
Editorial Team

Editorial Team

Editorial Team

Publisher description: The J-Space Editorial Team curates and synthesizes cutting-edge research in mechanistic interpretability, AI alignment, and cognitive science. With a commitment to precision, empirical rigor, and clear exposition, the team distills complex technical developments into accessible yet authoritative analyses for researchers, engineers, and policymakers working at the frontier of artificial intelligence.