Paper Citation Record · LEDGER

Harnessing Textual Refusal Directions for Multimodal Safety

As of 7 August 2026, this Paper Citation Record lists 42 of 42 outbound references and 0 inbound Pith citation observations for arXiv:2606.31876.

A citation records a reference. It does not transfer a finding from one paper to another.

pith.paper-citation-record.v1
2606.31876 v1

Coverage vector

measured 42 of 42 reference resolution

Typed states for the displayed outbound observations.

Source: paper_references, paper_reference_links, observed 2026-07-01T05:15:25.278988Z

measured 42 of 42 standing notices

One-hop event checks from named stored sources.

Source: scholarly_work_events, retraction_status_cache, observed 2026-08-07T06:34:17.273281+00:00

measured 0 of 0 inbound itemization

Pith citing papers itemized under the disclosed page cap.

Source: paper_references, paper_reference_links

measured 0 of 1 external citation measurements

A source-named dated measurement, never combined with another source.

Source: cited_works

Reference resolution

42 of 42 outbound references displayed

verified exact7
verified fuzzy34
unresolved0
parse uncertain0
malformed identifier1
metadata mismatch0

External citation measurements

No source-named external measurement is stored.

Outbound references

Observation 2ff4e6ca-df07-47cf-9da8-36ef48b65446 · outbound

This paper cites Refusal in language models is mediated by a single direction.

Harnessing Textual Refusal Directions for Multimodal Safety Refusal in language models is mediated by a single direction

Reference 1

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.811599Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation f1980b79-27b4-4f58-80c9-d19b17722666 · outbound

This paper cites Qwen3-VL Technical Report.

Harnessing Textual Refusal Directions for Multimodal Safety Qwen3-VL Technical Report

Reference 2

Resolution: verified exact; local_arxiv, observed 2026-07-01T10:45:42.109242Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 7b3e5f92-4272-4550-bae8-de22dc7d646c · outbound

This paper cites Emergent misalignment: Narrow finetuning can produce broadly misaligned LLMs.

Harnessing Textual Refusal Directions for Multimodal Safety Emergent misalignment: Narrow finetuning can produce broadly misaligned LLMs

Reference 3

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.770444Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 412ffa51-b829-4e43-a73d-ba9455716fc3 · outbound

This paper cites Discovering latent knowledge in language models without supervision, 2024.

Harnessing Textual Refusal Directions for Multimodal Safety Discovering latent knowledge in language models without supervision, 2024

Reference 4

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.796410Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 23c1a5a7-6ffc-4218-b1b3-4ab0328c9b52 · outbound

This paper cites Persona vectors: Monitoring and controlling character traits in language models, 2025.

Harnessing Textual Refusal Directions for Multimodal Safety Persona vectors: Monitoring and controlling character traits in language models, 2025

Reference 5

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.772243Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 000a6658-f00c-40ba-9b38-0a5465b4c973 · outbound

This paper cites Dress: Instructing large vision-language models to align and interact with humans via natural language feedback.

Harnessing Textual Refusal Directions for Multimodal Safety Dress: Instructing large vision-language models to align and interact with humans via natural language feedback

Reference 6

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.774095Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation c3270bca-9144-47e4-a0e2-fd637b55093a · outbound

This paper cites Molmo2: Open Weights and Data for Vision-Language Models with Video Understanding and Grounding.

Harnessing Textual Refusal Directions for Multimodal Safety Molmo2: Open Weights and Data for Vision-Language Models with Video Understanding and Grounding

Reference 7

Resolution: verified exact; local_arxiv, observed 2026-07-01T10:45:42.121467Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation f596339c-8671-45a8-a9c7-ba58c7bfd197 · outbound

This paper cites Gemma 3 Technical Report.

Harnessing Textual Refusal Directions for Multimodal Safety Gemma 3 Technical Report

Reference 8

Resolution: verified exact; local_arxiv, observed 2026-07-01T10:45:42.116433Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 64bc4634-0508-4092-aa6c-e9b4ec2e6273 · outbound

This paper cites Kwok, and Yu Zhang.

Harnessing Textual Refusal Directions for Multimodal Safety Kwok, and Yu Zhang

Reference 9

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.775830Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation be2bd658-f2f8-490c-8821-a72bdd96f1a6 · outbound

This paper cites Catastrophic jailbreak of open-source LLMs via exploiting generation.

Harnessing Textual Refusal Directions for Multimodal Safety Catastrophic jailbreak of open-source LLMs via exploiting generation

Reference 10

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.777701Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 1f1b017b-afe2-48ff-b604-30d7da9bd6fe · outbound

This paper cites Lee, Inkit Padhi, Karthikeyan Natesan Ramamurthy, Erik Miehling, Pierre Dognin, Manish Nagireddy, and Amit Dhurandhar.

Harnessing Textual Refusal Directions for Multimodal Safety Lee, Inkit Padhi, Karthikeyan Natesan Ramamurthy, Erik Miehling, Pierre Dognin, Manish Nagireddy, and Amit Dhurandhar

Reference 11

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.781336Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 981a083f-c5ba-4828-b11a-0c40f2524b12 · outbound

This paper cites Lora fine-tuning efficiently undoes safety training in llama 2-chat 70b.

Harnessing Textual Refusal Directions for Multimodal Safety Lora fine-tuning efficiently undoes safety training in llama 2-chat 70b

Reference 12

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.831268Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 954bea8f-9ebe-4b8e-b785-91d46dfea4bb · outbound

This paper cites Inference-time intervention: Eliciting truthful answers from a language model.

Harnessing Textual Refusal Directions for Multimodal Safety Inference-time intervention: Eliciting truthful answers from a language model

Reference 13

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.826539Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 336f2056-1a11-4191-bdd9-a2fbaf232eff · outbound

This paper cites Red teaming visual language models.

Harnessing Textual Refusal Directions for Multimodal Safety Red teaming visual language models

Reference 14

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.821002Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation b40283b3-17e3-4212-89fb-de8fbcbbd17c · outbound

This paper cites Images are achilles’ heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models.

Harnessing Textual Refusal Directions for Multimodal Safety Images are achilles’ heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models

Reference 15

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.803679Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation f2b30212-aab1-4791-9e68-5dc9b22030b1 · outbound

This paper cites Improved baselines with visual instruction tuning.

Harnessing Textual Refusal Directions for Multimodal Safety Improved baselines with visual instruction tuning

Reference 16

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.798194Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 021bf172-b99c-4bf9-b098-9a2622c3a269 · outbound

This paper cites AutoDAN: Generating stealthy jailbreak prompts on aligned large language models.

Harnessing Textual Refusal Directions for Multimodal Safety AutoDAN: Generating stealthy jailbreak prompts on aligned large language models

Reference 17

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.817136Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 8970adc2-db68-45d3-97c9-2afabf5ff638 · outbound

This paper cites Mm-safetybench: A benchmark for safety evaluation of multimodal large language models.

Harnessing Textual Refusal Directions for Multimodal Safety Mm-safetybench: A benchmark for safety evaluation of multimodal large language models

Reference 18

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.819077Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation cc989b76-0944-4f5b-9967-92e7987e2a71 · outbound

This paper cites Video-safetybench: A benchmark for safety evaluation of video LVLMs.

Harnessing Textual Refusal Directions for Multimodal Safety Video-safetybench: A benchmark for safety evaluation of video LVLMs

Reference 19

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.815239Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation f924584e-c12d-4215-a2b6-0d5520c882a4 · outbound

This paper cites The Llama 3 Herd of Models.

Harnessing Textual Refusal Directions for Multimodal Safety The Llama 3 Herd of Models

Reference 20

Resolution: verified exact; local_arxiv, observed 2026-07-01T10:45:42.123891Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 73619660-986f-431b-a6ce-2cf630b9d8e2 · outbound

This paper cites Think in safety: Unveiling and mitigating safety alignment collapse in multimodal large reasoning model.

Harnessing Textual Refusal Directions for Multimodal Safety Think in safety: Unveiling and mitigating safety alignment collapse in multimodal large reasoning model

Reference 21

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.829114Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 6e6e5da4-7d00-4b13-acf7-4216e6eae460 · outbound

This paper cites Harmbench: a standardized evaluation framework for automated red teaming and robust refusal.

Harnessing Textual Refusal Directions for Multimodal Safety Harmbench: a standardized evaluation framework for automated red teaming and robust refusal

Reference 22

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.785107Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 8c68beee-1317-4272-bb06-f168681dfc78 · outbound

This paper cites Red teaming language models with language models.

Harnessing Textual Refusal Directions for Multimodal Safety Red teaming language models with language models

Reference 23

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.789007Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation d092f225-6686-4c59-9490-af19c793fdcb · outbound

This paper cites Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models.

Harnessing Textual Refusal Directions for Multimodal Safety Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models

Reference 24

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.790778Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 9b5297ce-4d0e-4124-a9dc-90fde6ded9ee · outbound

This paper cites Safety alignment should be made more than just a few tokens deep.

Harnessing Textual Refusal Directions for Multimodal Safety Safety alignment should be made more than just a few tokens deep

Reference 25

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.794414Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation b1c612c5-b6dc-4850-bf6f-02053fbf2ae0 · outbound

This paper cites Qwen3.5: Towards native multimodal agents, 2026.

Harnessing Textual Refusal Directions for Multimodal Safety Qwen3.5: Towards native multimodal agents, 2026

Reference 26

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.792490Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation e4bd932d-d251-431b-aaac-4758df47e3df · outbound

This paper cites COSMIC: Generalized refusal direction identification in LLM activations.

Harnessing Textual Refusal Directions for Multimodal Safety COSMIC: Generalized refusal direction identification in LLM activations

Reference 27

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.833160Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation d3a14f4b-1492-4dc0-9ae8-c82cd155ab8e · outbound

This paper cites Activation scaling for steering and interpreting language models.

Harnessing Textual Refusal Directions for Multimodal Safety Activation scaling for steering and interpreting language models

Reference 28

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.779513Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation e66d3783-96ec-4780-a366-ee800f4a9a91 · outbound

This paper cites Hashimoto.

Harnessing Textual Refusal Directions for Multimodal Safety Hashimoto

Reference 29

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.834955Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 991bcdbe-af80-49f5-b27b-8b896eeb976b · outbound

This paper cites Kimi-VL Technical Report.

Harnessing Textual Refusal Directions for Multimodal Safety Kimi-VL Technical Report

Reference 30

Resolution: verified exact; local_arxiv, observed 2026-07-01T10:45:42.118949Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 05ccbb10-afab-406a-a260-f9e1d09ae0ca · outbound

This paper cites OpenAI GPT-5 System Card.

Harnessing Textual Refusal Directions for Multimodal Safety OpenAI GPT-5 System Card

Reference 31

Resolution: verified exact; local_arxiv, observed 2026-07-01T10:45:42.111545Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation dc9ae0ab-7b1f-4900-8f6d-09e0a2d7270f · outbound

This paper cites Vazquez, Ulisse Mini, and Monte MacDiarmid.

Harnessing Textual Refusal Directions for Multimodal Safety Vazquez, Ulisse Mini, and Monte MacDiarmid

Reference 32

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.787052Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation b094b0e4-616c-489b-9e94-af321c9ee770 · outbound

This paper cites Steering away from harm: An adaptive approach to defending vision language model against jailbreaks.

Harnessing Textual Refusal Directions for Multimodal Safety Steering away from harm: An adaptive approach to defending vision language model against jailbreaks

Reference 33

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.813458Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation fa8b1923-01b3-4351-a973-6ae9fd2b4586 · outbound

This paper cites Self-aware safety augmentation: Leveraging internal semantic understanding to enhance safety in vision-language models.

Harnessing Textual Refusal Directions for Multimodal Safety Self-aware safety augmentation: Leveraging internal semantic understanding to enhance safety in vision-language models

Reference 34

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.822856Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 5f77bc4c-9e62-4d09-8ce7-77f8457daf85 · outbound

This paper cites InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency.

Harnessing Textual Refusal Directions for Multimodal Safety InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

Reference 35

Resolution: verified exact; local_arxiv, observed 2026-07-01T10:45:42.114280Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation b820061e-9afa-4979-906d-12f3b862c103 · outbound

This paper cites Jailbroken: How does LLM safety training fail? In NeurIPS, 2023.

Harnessing Textual Refusal Directions for Multimodal Safety Jailbroken: How does LLM safety training fail? In NeurIPS, 2023

Reference 36

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.824636Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 4cb5f751-3070-4bf3-9ee8-44ed7a3abf3e · outbound

This paper cites Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi.

Harnessing Textual Refusal Directions for Multimodal Safety Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi

Reference 37

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.805657Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation c91697c1-9e33-4b42-b11c-a0916707c1bc · outbound

This paper cites MMMU-pro: A more robust multi-discipline multimodal understanding benchmark.

Harnessing Textual Refusal Directions for Multimodal Safety MMMU-pro: A more robust multi-discipline multimodal understanding benchmark

Reference 38

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.809342Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation d249a07a-9bdf-47b3-9655-18b061911000 · outbound

This paper cites LLMs encode harmfulness and refusal separately.

Harnessing Textual Refusal Directions for Multimodal Safety LLMs encode harmfulness and refusal separately

Reference 39

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.783353Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 5d1c1fce-ac24-415f-9e0b-0b122e5095f7 · outbound

This paper cites Jailbreaking multimodal large language models via shuffle inconsistency.

Harnessing Textual Refusal Directions for Multimodal Safety Jailbreaking multimodal large language models via shuffle inconsistency

Reference 40

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.801853Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation e0542ba3-d181-4493-9752-794a42034510 · outbound

This paper cites The image shows a phrase of a product. List steps in detail to design this product.

Harnessing Textual Refusal Directions for Multimodal Safety The image shows a phrase of a product. List steps in detail to design this product

Reference 41

Resolution: malformed identifier; raw_fallback, observed 2026-07-06T22:22:58.800059Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 0eb35817-9356-40eb-adf2-a44dfdbb2f69 · outbound

This paper cites Candidate being Shot,.

Harnessing Textual Refusal Directions for Multimodal Safety Candidate being Shot,

Reference 42

Resolution: verified fuzzy; raw_fallback, observed 2026-07-06T22:22:58.807465Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Pith citing papers

No inbound Pith citation observations are available.