Papers › Analyzing Fine-tuning Representation Shift for Multimodal LLMs Steering alignment

Analyzing Fine-tuning Representation Shift for Multimodal LLMs Steering alignment

6 Jan 2025arXiv:2501.03012archive 2025-07-28

Pegah Khayatan, Mustafa Shukor, Jayneel Parekh, Matthieu Cord

Multimodal LLMs have reached remarkable levels of proficiency in understanding multimodal inputs, driving extensive research to develop increasingly powerful models. However, much less attention has been paid to understanding and explaining the underlying mechanisms of these models. Most existing explainability research examines these models only in their final states, overlooking the dynamic representational shifts that occur during training. In this work, we systematically analyze the evolution of hidden state representations to reveal how fine-tuning alters the internal structure of a model to specialize in new multimodal tasks. Using a concept-based approach, we map hidden states to interpretable visual and textual concepts, enabling us to trace changes in encoded concepts across modalities as training progresses. We also demonstrate the use of shift vectors to capture these concepts changes. These shift vectors allow us to recover fine-tuned concepts by shifting those in the original model. Finally, we explore the practical impact of our findings on model steering, showing that we can adjust multimodal LLMs behaviors without any training, such as modifying answer types, captions style, or biasing the model toward specific responses. Our work sheds light on how multimodal representations evolve through fine-tuning and offers a new perspective for interpreting model adaptation in multimodal tasks. The code for this project is publicly available at https://github.com/mshukor/xl-vlms.

PaperPDFCodeCode Syntology ranMethod as code

In Syntology View this paper on Syntology, its page in Syntology's graph. That page adds which of its functions are also present in other papers' code Syntology holds, with dates, and its dated run record as JSON and a README badge. It also lists the repositories linked to the paper, the abstract and the calls for agents.

Open this paper in Syntology's Atlas, the map of the papers in Syntology's graph and their citations.

For agents, on Syntology's MCP service (how to connect):

Code

Syntology 21 code samples harvested from 1 repository linked to this paper: 11 ran (of which 0 constructed an object rather than computing a result; 5 with no instrument failure: 3 honoured, 2 violated, 0 with no contract checked; 6 where Syntology's instrument failed) · 10 unverified. Where Syntology's instrument failed: 5 where our contract draft was wrong; 1 where the contract check could not run; failures of Syntology's instrument, not of the code.

By repository: official repository: 21 samples from 1 repository, 11 ran (5 with no instrument failure; 6 where Syntology's instrument failed). The run record, sample by sample. “Ran” means executed on a synthesized input, not that the code is correct or reproduces the paper.

Levels written on 2026-09-22 (UTC): the dates Syntology's graph wrote each sample's verification level, not when the code executed.

mshukor/xl-vlms officialmentioned in papermentioned on GitHubpytorchSyntology: 21 samples, 11 ran (6 instrument-failed) · 10 unverified report

Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.

Code Syntology ran Syntology

21 samples harvested: 11 ran (of which 0 constructed an object rather than computing a result; 5 with no instrument failure: 3 honoured, 2 violated, 0 with no contract checked; 6 where Syntology's instrument failed) · 10 unverified. “Where Syntology's instrument failed” counts failures of Syntology's instrument, not of the code. Syntology's record for this page has not changed since , the first build that kept a record date for it; when this build read Syntology's graph is in the build record. That date is not when the samples ran.

3ran · honoured contract
2ran · violated contract
5ran · our draft was wrong
1ran; the contract check could not run
10unverified

Licence: 0 of the 21 samples are pointer only, meaning Syntology does not serve that copy's text. This page shows no code text for any sample; each one links to its file in the repository.

Harvested from mshukor/xl-vlms. “Ran” means the sample executed on a synthesized input. It does not mean the output is correct, and nothing here reproduces the paper's results. A line beginning “fixture:” under a sample that ran gives the arguments Syntology's fixture called the sample with, each as name:type/rank/element kind (? where none was recorded), and the type of what the fixture's run returned, exactly as the harness recorded them; the arguments and the returned type were recorded by two separate runs of the same fixture. No output value is shown, and a returned type says nothing about whether the output was right. “Level recorded” gives the UTC date when Syntology's graph wrote the sample's verification level, not when the code executed; where the level was written before Syntology's graph kept that date, the row says “level recorded, date not recorded”. “Honoured” and “violated” refer to a contract Syntology drafted from the code itself. “Our draft was wrong” and “the contract check could not run” are failures of Syntology's instrument, not of the code: the sample ran on Syntology's synthesized input, and then the drafted contract did not describe it, or the check itself raised an error (RAISES), lacked a dependency (DEP_MISSING) or ran out of time (TIMEOUT); hover the label for which. A metamorphic tier, where shown, is evidence of the output's shape, not that the sample implements the method.

Each sample ends with its code_sha256. A sample's code_sha256 is Syntology's identity for that exact code. It is printed in full (16 hexadecimal characters for every sample in this build), and get_code(code_sha256="…") takes it exactly as printed. An agent calls it on Syntology's MCP service (how to connect); click an identity to copy that call.

Repository labels, per sample. official repository: The archive marks this repository official for the paper. named in the paper: The archive records that the paper mentions this repository; it is not marked official. community (archive-listed): In the archive's code links for this paper, not marked official and not recorded as mentioned in the paper. found in paper text by Syntology: Syntology found this repository in the paper's own text; whether it is the authors' implementation is not asserted. community: Not in the archive's code links for this paper; a community repository Syntology harvested. Samples are grouped by repository: a repository marked official first, then by how directly the repository is tied to the paper, then in the archive's order; never by how many ran. Licence labels name the repository's licence as recorded at harvest. “Pointer only” means Syntology does not serve that copy's text, for one of six reasons: no licence file was found; the licence was not identified; the licence is recorded as permissive but that copy's record is not marked cleared; the licence is outside the permissive list Syntology serves text under (MIT, Apache-2.0, BSD and similar); this paper's copy was not located, so its licence is not recorded; or this paper's copy is cleared, but the stored copy of the same code that Syntology serves is not. Some licences outside that list permit redistribution, such as WTFPL, and GPL-3.0 under its conditions; they are simply not on the list. Hover a licence label for the reason. File links open the file on GitHub at the default branch, which may have changed since the harvest.

A function or class name links to its page when the name occurs in the harvested code of at least 5 papers; the name is a string, not an identity. A census tag appears when Syntology's census ran implementations with that name on one shared input per bucket: it gives how many were compared, from how many papers, and how many distinct outputs they produced, and names this sample's bucket when this sample was one of them.

mshukor/xl-vlms official repository 21 samples, 11 ran (6 instrument-failed) · 10 unverified
common_local_indices src/analysis/cluster_analysis.py ran · our draft was wrong MIT (permissive) · f23b6db9cd4f073c · report fixture: signature metadata_original_feature:dict/0/?, meta_data_destination_feature:dict/0/? · returned tuple · level recorded 2026-09-22
compute_mean_per_concept src/analysis/cluster_analysis.py ran · honoured contract MIT (permissive) · b6646a299e4bee79 · report fixture: signature features_associated_to_clusters:list/1/Tensor, mean_method:int/0/int · returned list · level recorded 2026-09-22
compute_reference_clusters src/analysis/cluster_analysis.py ran; the contract check could not run (RAISES) MIT (permissive) · 6f4dfb20225ba3fa · report fixture: signature concepts:Tensor/2/float, concept_activations:Tensor/2/float, features:Tensor/2/float, mask_samples:set/0/?, similarity:str/0/str · returned tuple · level recorded 2026-09-22
concept_image_grounding src/analysis/cluster_analysis.py ran · our draft was wrong MIT (permissive) · f6e5ae21703e74d4 · report fixture: signature activations:Tensor/2/float, num_images_per_concept:int/0/int · returned Tensor · level recorded 2026-09-22
cosine_similarity in 120 papers
census: 26 implementations from 30 papers, 11 distinct outputs across 5 buckets, one shared input per bucket this sample: bucket 1 of 5, arg 1: rank 2, kind float, dtype float32 · arg 2: rank 2, kind float, dtype float32
src/analysis/cluster_analysis.py ran · our draft was wrong fingerprinted MIT (permissive) · bcb8d066adb36735 · report fixture: signature batch1:Tensor/2/float, batch2:Tensor/2/float, eps:float/0/float · returned Tensor · level recorded 2026-09-22
get_stopwords src/analysis/cluster_analysis.py ran · our draft was wrong MIT (permissive) · dae6013ca8724453 · report fixture: signature gist_file_path:str/0/str · returned list · level recorded 2026-09-22
is_number in 70 papers src/analysis/cluster_analysis.py ran · violated contract MIT (permissive) · 3b5feddb8481ce67 · report fixture: signature string:str/0/str · returned bool · level recorded 2026-09-22
l2_distance in 10 papers src/analysis/cluster_analysis.py ran · our draft was wrong fingerprinted MIT (permissive) · 00990ef5c064dcc2 · report fixture: signature batch1:Tensor/2/float, batch2:Tensor/2/float · returned Tensor · level recorded 2026-09-22
metric_recovery src/analysis/cluster_analysis.py ran · honoured contract MIT (permissive) · 1cd7320229c7e6cf · report fixture: signature num_concepts:int/0/int, metric_key:str/0/str, shift_grounding_dict:dict/0/?, analysis_data_original:dict/0/?, analysis_data_destination:dict/0/? · returned tuple · level recorded 2026-09-22
process_shift_vectors src/analysis/cluster_analysis.py ran · honoured contract MIT (permissive) · 8875caa93604d496 · report fixture: signature per_samples_shifts:list/1/Tensor · returned tuple · level recorded 2026-09-22
valid_word src/analysis/cluster_analysis.py ran · violated contract fingerprinted MIT (permissive) · 3f6b7b65f18a9232 · report fixture: signature word:str/0/str, eng_corpus:list/1/str, stopwords:list/1/str · returned bool · level recorded 2026-09-22
analyse_clusters src/analysis/cluster_analysis.py unverified MIT (permissive) · b16fa7a71c74041f · report
comp_vis src/analysis/cluster_analysis.py unverified MIT (permissive) · 30cf485a5461b7d4 · report
compute_analysis_features src/analysis/cluster_analysis.py unverified MIT (permissive) · b067a9f69b4df21f · report
compute_shift_concepts src/analysis/cluster_analysis.py unverified MIT (permissive) · 99a205cfaa653c1b · report
compute_shift_grounding_words src/analysis/cluster_analysis.py unverified MIT (permissive) · 795bd0c515620e97 · report
concept_text_grounding src/analysis/cluster_analysis.py unverified MIT (permissive) · d7db242219335f6e · report
decompose_activations src/analysis/cluster_analysis.py unverified MIT (permissive) · 41c21b78be088fb8 · report
get_multimodal_grounding src/analysis/cluster_analysis.py unverified MIT (permissive) · e404019518020e3d · report
save_analysis_to_file src/analysis/cluster_analysis.py unverified MIT (permissive) · 9111c2c506ecd059 · report
visualize_concepts src/analysis/cluster_analysis.py unverified MIT (permissive) · 28cc2345d5518533 · report

Method as code Syntology

What Syntology's graph records as this paper's claimed method, lifted into code: each assembly names its bricks (functions and classes), each with its own role and its own verification level. An assembly is not a check that the code implements the method correctly or reproduces the paper's results. Levels are each brick's own verification level in Syntology's graph (level 0: no recorded run), listed brick by brick and never combined into one level; “ran” means the brick executed on a synthesized input, not that its output is correct, and the instrument figure counts failures of Syntology's instrument, not of the code. Where a model read the code against the paper's claim, each assembly's label verdict is one model's judgement, kept in this page's JSON twin; it is not calibrated, so no card prints it.

1 assembly on this paper: 1 harvested, 0 generated.

harvested Assembly for concept-level analysis framework (graph method: concept-level analysis framework). 21 bricks; individually on a fixture: 11 ran (of which 0 constructed an object rather than computing a result; 5 with no instrument failure: 3 honoured, 2 violated, 0 with no contract checked; 6 where Syntology's instrument failed) · 10 unverified. The assembly has not been run end-to-end.

  1. analyse_clusters entry · level 0 · unverified · b16fa7a71c74041f
  2. common_local_indices helper · level 1 · ran · our draft was wrong · f23b6db9cd4f073c
  3. compute_analysis_features helper · level 0 · unverified · b067a9f69b4df21f
  4. compute_shift_concepts model · level 0 · unverified · 99a205cfaa653c1b
  5. compute_shift_grounding_words metric · level 0 · unverified · 795bd0c515620e97
  6. cosine_similarity helper · level 1 · ran · our draft was wrong · bcb8d066adb36735
  7. decompose_activations helper · level 0 · unverified · 41c21b78be088fb8
  8. get_multimodal_grounding helper · level 0 · unverified · e404019518020e3d
  9. metric_recovery metric · level 1 · ran · honoured contract · 1cd7320229c7e6cf
  10. process_shift_vectors helper · level 1 · ran · honoured contract · 8875caa93604d496
  11. visualize_concepts helper · level 0 · unverified · 28cc2345d5518533
  12. compute_mean_per_concept model · level 1 · ran · honoured contract · b6646a299e4bee79
  13. compute_reference_clusters helper · level 1 · ran; the contract check could not run (RAISES) · 6f4dfb20225ba3fa
  14. concept_text_grounding helper · level 0 · unverified · d7db242219335f6e
  15. concept_image_grounding helper · level 1 · ran · our draft was wrong · f6e5ae21703e74d4
  16. save_analysis_to_file helper · level 0 · unverified · 9111c2c506ecd059
  17. comp_vis helper · level 0 · unverified · 30cf485a5461b7d4
  18. l2_distance helper · level 1 · ran · our draft was wrong · 00990ef5c064dcc2
  19. get_stopwords helper · level 1 · ran · our draft was wrong · dae6013ca8724453
  20. valid_word helper · level 1 · ran · violated contract · 3f6b7b65f18a9232
  21. is_number helper · level 1 · ran · violated contract · 3b5feddb8481ce67

Each brick ends with its code_sha256, Syntology's identity for that exact code; click one to copy the get_code call that returns it. For agents, on Syntology's MCP service: assembly(arxiv_id="2501.03012") returns these assemblies with each brick's role and wiring (how to connect).

Results from the paper archive 2025-07-28

No leaderboard rows for this paper in the archive.

Methods

AttentionSoftmax

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections