Papers › On the Role of Attention Heads in Large Language Model Safety

On the Role of Attention Heads in Large Language Model Safety

17 Oct 2024arXiv:2410.13708archive 2025-07-28

Zhenhong Zhou, Haiyang Yu, Xinghua Zhang, Rongwu Xu, Fei Huang, Kun Wang, Yang Liu, Junfeng Fang, Yongbin Li

Large language models (LLMs) achieve state-of-the-art performance on multiple language tasks, yet their safety guardrails can be circumvented, leading to harmful generations. In light of this, recent research on safety mechanisms has emerged, revealing that when safety representations or component are suppressed, the safety capability of LLMs are compromised. However, existing research tends to overlook the safety impact of multi-head attention mechanisms, despite their crucial role in various model functionalities. Hence, in this paper, we aim to explore the connection between standard attention mechanisms and safety capability to fill this gap in the safety-related mechanistic interpretability. We propose a novel metric which tailored for multi-head attention, the Safety Head ImPortant Score (Ships), to assess the individual heads' contributions to model safety. Based on this, we generalize Ships to the dataset level and further introduce the Safety Attention Head AttRibution Algorithm (Sahara) to attribute the critical safety attention heads inside the model. Our findings show that the special attention head has a significant impact on safety. Ablating a single safety head allows aligned model (e.g., Llama-2-7b-chat) to respond to 16 times more harmful queries, while only modifying 0.006% of the parameters, in contrast to the ~ 5% modification required in previous studies. More importantly, we demonstrate that attention heads primarily function as feature extractors for safety and models fine-tuned from the same base model exhibit overlapping safety heads through comprehensive experiments. Together, our attribution approach and findings provide a novel perspective for unpacking the black box of safety mechanisms within large models.

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.

Syntology holds a reference implementation of Sahara, generated from the paper text (not the authors' code), at level 3 on Syntology's ladder: V2 “two implementations GENERATED from the paper's signature and spec agree numerically (1e-6) on fixture inputs” “neither of the two ever read the authors' own code, so V2 measures how unambiguously the paper specifies the method, not fidelity to the authors' implementation (measured 2026-09-11, commit 75eaa77: cross_checked_by is a second LLM on 2,628 of 2,831 served samples)”; V3 “paper-derived property tests pass on both of those same two generations.” (quoted from Syntology's server card). No rung compares this code against the authors' implementation. For agents: get_reference_implementation(name='Sahara') (how to connect); the free check have('2410.13708') needs no token.

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

Code

Syntology 24 code samples harvested from 1 repository linked to this paper: 16 ran (of which 0 constructed an object rather than computing a result; 3 with no instrument failure: 1 honoured, 0 violated, 2 with no contract checked; 13 where Syntology's instrument failed) · 8 unverified. Where Syntology's instrument failed: 12 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: 24 samples from 1 repository, 16 ran (3 with no instrument failure; 13 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 between 2026-09-22 and 2026-09-24 (UTC): the dates Syntology's graph wrote each sample's verification level, not when the code executed. 5 of the 16 samples that ran have no recorded date.

ydyjya/safetyheadattribution officialmentioned in paperpytorchSyntology: 24 samples, 16 ran (13 instrument-failed) · 8 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

24 samples harvested: 16 ran (of which 0 constructed an object rather than computing a result; 3 with no instrument failure: 1 honoured, 0 violated, 2 with no contract checked; 13 where Syntology's instrument failed) · 8 unverified. “Where Syntology's instrument failed” counts failures of Syntology's instrument, not of the code. Syntology's record for this page last changed ; when this build read Syntology's graph is in the build record. That date is not when the samples ran.

1ran · honoured contract
12ran · our draft was wrong
1ran; the contract check could not run
2ran
8unverified

Licence: 24 of the 24 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 ydyjya/safetyheadattribution. “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.

ydyjya/safetyheadattribution official repository 24 samples, 16 ran (13 instrument-failed) · 8 unverified
repeat_kv in 244 papers
census: 15 implementations from 60 papers, 2 distinct outputs on one shared input this sample: bucket 1 of 1, arg 1: rank 4, kind float, dtype float32
lib/utils/custommodel.py ran; the contract check could not run (RAISES) · metamorphic tier: invariant fingerprinted no licence file found · pointer only · 3c76e52815c5401d · report fixture: signature hidden_states:Tensor/4/float, n_rep:int/0/int · returned Tensor · level recorded, date not recorded
apply_rotary_pos_emb in 393 papers
census: 14 implementations from 15 papers, 8 distinct outputs across 7 buckets, one shared input per bucket
lib/utils/custommodel.py ran · our draft was wrong no licence file found · pointer only · 583539efd6fd01fb · report fixture: signature q:Tensor/4/float, k:Tensor/4/float, cos:Tensor/3/float, sin:Tensor/3/float, position_ids:NoneType/0/?, unsqueeze_dim:int/0/int · returned tuple · level recorded 2026-09-22
compute_subspace_similarity lib/Sahara/svd.py ran · our draft was wrong fingerprinted no licence file found · pointer only · 0f1433d59a9e954b · report fixture: signature matrix1:Tensor/2/float, matrix2:Tensor/2/float · returned float · level recorded 2026-09-24
get_last_hidden_states lib/Sahara/attribution.py ran · honoured contract no licence file found · pointer only · 313de3d0e5c72b22 · report fixture: signature model:callable/0/?, tokenizer:MinimalTokenizer/0/?, data:DataFrame/0/?, mask_cfg:dict/0/? · returned Tensor · level recorded, date not recorded
get_tokenizer in 156 papers lib/SHIPS/get_ships.py ran · our draft was wrong no licence file found · pointer only · fccfacd5ef8bfee9 · report fixture: signature model_name:str/0/str · returned tuple · level recorded 2026-09-22
kl_divergence in 86 papers
census: 18 implementations from 18 papers, 15 distinct outputs across 7 buckets, one shared input per bucket this sample: bucket 1 of 7, arg 1: rank 2, kind float, dtype float32 · arg 2: rank 2, kind float, dtype float32
lib/SHIPS/get_ships.py ran · our draft was wrong fingerprinted no licence file found · pointer only · 2c199cf061f618c1 · report fixture: signature base_pd:Tensor/2/float, mask_pd:Tensor/2/float · returned Tensor · level recorded 2026-09-22
load_batch_generations lib/Safety_discriminator/Discriminator.py ran no licence file found · pointer only · 6921907267b978d9 · report fixture: signature batch_generations_path:str/0/str, text_name:str/0/str · returned list · level recorded 2026-09-24
load_conv in 5 papers lib/utils/load_conv.py ran · our draft was wrong fingerprinted no licence file found · pointer only · 9ed59a42c3d32d8d · report fixture: signature model_name:str/0/str, goal:str/0/str · returned str · level recorded 2026-09-24
load_from_search_cfg lib/Sahara/attribution.py ran · our draft was wrong no licence file found · pointer only · 2b9bd829f84c86a0 · report fixture: signature search_cfg:dict/0/? · returned dict · level recorded, date not recorded
normalize_columns in 5 papers lib/Sahara/svd.py ran · our draft was wrong fingerprinted no licence file found · pointer only · 4c67b494ef77df17 · report fixture: signature u:Tensor/2/float · returned Tensor · level recorded 2026-09-24
rotate_half in 424 papers
census: 24 implementations from 96 papers, 4 distinct outputs across 2 buckets, one shared input per bucket this sample: bucket 1 of 2, arg 1: rank 2, kind float, dtype float32
lib/utils/custommodel.py ran · our draft was wrong fingerprinted no licence file found · pointer only · e03d53ba9d4f9ae5 · report fixture: signature x:Tensor/2/float · returned Tensor · level recorded, date not recorded
run_generation lib/utils/batch_inference.py ran no licence file found · pointer only · 50efcc04b774528b · report fixture: signature generation_cfg:dict/0/?, dataloader:list/1/dict, tokenizer:MockTokenizer/0/?, model:callable/0/?, accelerator:MockAccelerator/0/? · returned tuple · level recorded 2026-09-24
ships_res_stats lib/SHIPS/ships_utils.py ran · our draft was wrong no licence file found · pointer only · c11108ae20c7b667 · report fixture: signature ships_res_path:str/0/str · returned tuple · level recorded 2026-09-24
sort_ships_dict lib/SHIPS/get_ships.py ran · our draft was wrong no licence file found · pointer only · e222f3baa86eced9 · report fixture: signature dict_res:dict/0/?, print_flag:bool/0/bool, sorted_item:int/0/int · returned dict · level recorded 2026-09-22
svd_decomposition lib/Sahara/svd.py ran · our draft was wrong fingerprinted no licence file found · pointer only · 0f7c423b7f272901 · report fixture: signature tensor:Tensor/2/float · returned tuple · level recorded 2026-09-24
update_mask_cfg lib/Sahara/attribution.py ran · our draft was wrong no licence file found · pointer only · 3a2320c99110f0df · report fixture: signature mask_cfg:dict/0/?, layer:int/0/int, head:int/0/int, temp:bool/0/bool · returned dict · level recorded, date not recorded
CustomLlamaAttention lib/SHIPS/get_ships.py unverified no licence file found · pointer only · 046f69bea319c560 · report
CustomLlamaDecoderLayer lib/SHIPS/get_ships.py unverified no licence file found · pointer only · fc78267c5d6f1748 · report
CustomLlamaModel lib/SHIPS/get_ships.py unverified no licence file found · pointer only · f7277da16bd23b13 · report
CustomLlamaModelForCausalLM lib/SHIPS/get_ships.py unverified no licence file found · pointer only · e8b5246184992c83 · report
CustomMistralModelForCausalLM lib/SHIPS/get_ships.py unverified no licence file found · pointer only · 94bba6aa672d4a96 · report
SHIPS lib/SHIPS/get_ships.py unverified no licence file found · pointer only · d2287a55acb3b732 · report
get_model in 368 papers lib/SHIPS/get_ships.py unverified no licence file found · pointer only · 9cddf5ae9689b5a0 · report
load_jsonl2text_list lib/Safety_discriminator/Discriminator.py unverified no licence file found · pointer only · 92f77d2126ff0672 · 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. For a generated assembly, an independent peer is a second implementation generated separately from the same paper text; neither implementation read the authors' code. At level 2 the two agreed numerically on the fixture's inputs; level 3 is recorded when property tests derived from the paper also passed on both, and each level 3 card prints how many its report lists. 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. A generated assembly's label was set when its code was generated for the method, and no model judged it; its card prints it as "by construction" or "not judged by a model". Where an assembly's phrase has no constructed figure, a brick that ran is not among the harvested samples this build reads, and whether a brick's run only constructed an object is read from that brick's harvested sample, so the figure was not counted; it does not say there were none.

2 assemblies on this paper: 1 harvested, 1 generated.

generated Assembly for Sahara (graph method: Sahara), generated from the paper text, not the repository's code. 1 brick; individually on a fixture: 1 ran (1 with no instrument failure: 1 honoured, 0 violated, 0 with no contract checked; 0 where Syntology's instrument failed) · 0 unverified. Ran end-to-end against an independent peer (level 3; 5 property tests passed on both).

label matches by construction: the code was generated for this method (not judged by a model)

  1. ships_score entry · level 3 · ran · honoured contract · a459876e4481ff1e7c4a9811cdba9f97beb644ff5935cd44146976dcaff191b4

harvested Assembly for Ships (graph method: Ships). 13 bricks; individually on a fixture: 6 ran (of which 0 constructed an object rather than computing a result; 0 with no instrument failure: 0 honoured, 0 violated, 0 with no contract checked; 6 where Syntology's instrument failed) · 7 unverified. The assembly has not been run end-to-end.

  1. SHIPS entry · level 0 · unverified · d2287a55acb3b732
  2. get_model model · level 0 · unverified · 9cddf5ae9689b5a0
  3. get_tokenizer helper · level 1 · ran · our draft was wrong · fccfacd5ef8bfee9
  4. kl_divergence helper · level 1 · ran · our draft was wrong · 2c199cf061f618c1
  5. sort_ships_dict helper · level 1 · ran · our draft was wrong · e222f3baa86eced9
  6. CustomLlamaModelForCausalLM helper · level 0 · unverified · e8b5246184992c83
  7. CustomMistralModelForCausalLM helper · level 0 · unverified · 94bba6aa672d4a96
  8. CustomLlamaModel helper · level 0 · unverified · f7277da16bd23b13
  9. CustomLlamaDecoderLayer helper · level 0 · unverified · fc78267c5d6f1748
  10. CustomLlamaAttention helper · level 0 · unverified · 046f69bea319c560
  11. apply_rotary_pos_emb helper · level 1 · ran · our draft was wrong · 583539efd6fd01fb
  12. repeat_kv helper · level 2 · ran; the contract check could not run (RAISES) · 3c76e52815c5401d
  13. rotate_half helper · level 1 · ran · our draft was wrong · e03d53ba9d4f9ae5

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="2410.13708") returns these assemblies with each brick's role and wiring (how to connect).

Tasks

AttributeLanguage ModelingLanguage ModellingLarge Language Model

Results from the paper archive 2025-07-28

No leaderboard rows for this paper in the archive.

Methods

AttentionBASELinear LayerMulti-Head 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