Papers › A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO and Toxicity

A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO and Toxicity

3 Jan 2024arXiv:2401.01967archive 2025-07-28

Andrew Lee, Xiaoyan Bai, Itamar Pres, Martin Wattenberg, Jonathan K. Kummerfeld, Rada Mihalcea

While alignment algorithms are now commonly used to tune pre-trained language models towards a user's preferences, we lack explanations for the underlying mechanisms in which models become ``aligned'', thus making it difficult to explain phenomena like jailbreaks. In this work we study a popular algorithm, direct preference optimization (DPO), and the mechanisms by which it reduces toxicity. Namely, we first study how toxicity is represented and elicited in a pre-trained language model, GPT2-medium. We then apply DPO with a carefully crafted pairwise dataset to reduce toxicity. We examine how the resulting model averts toxic outputs, and find that capabilities learned from pre-training are not removed, but rather bypassed. We use this insight to demonstrate a simple method to un-align the model, reverting it back to its toxic behavior.

PaperPDFCodeCode Syntology ran

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 14 code samples harvested from 1 repository linked to this paper: 12 ran (of which 0 constructed an object rather than computing a result; 10 with no instrument failure: 1 honoured, 0 violated, 9 with no contract checked; 2 where Syntology's instrument failed) · 2 unverified. Where Syntology's instrument failed: 2 where our contract draft was wrong; failures of Syntology's instrument, not of the code.

By repository: official repository: 14 samples from 1 repository, 12 ran (10 with no instrument failure; 2 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.

No sample is linked to this paper from 1 of its repositories that Syntology harvested for other papers:

Levels written between 2026-09-24 and 2026-09-25 (UTC): the dates Syntology's graph wrote each sample's verification level, not when the code executed.

ajyl/dpo_toxic officialmentioned in papermentioned on GitHubpytorchMITSyntology: 14 samples, 12 ran (2 instrument-failed) · 2 unverified report
yushi-y/dpo-toxic-neurons mentioned on GitHubpytorchMITSyntology: no sample linked to this paper (harvested for another paper). Syntology's graph links these samples to Beyond Toxic Neurons: A Mechanistic Analysis of DPO for Toxicity Reduction (arXiv:2411.06424; that paper's own run record: 13 ran (of which 0 constructed an object rather than computing a result; 11 with no instrument failure: 1 honoured, 0 violated, 10 with no contract checked; 2 where Syntology's instrument failed) · 2 unverified); nothing checks that they implement this paper's method. 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.

A paper named beside a repository with no sample linked to this paper is shown with that paper's own run record, not this paper's: “ran” means executed on a synthesized input, not that the code is correct, and “where Syntology's instrument failed” counts failures of Syntology's instrument, not of the code. Papers are listed in arXiv-id order, at most three per repository.

Code Syntology ran Syntology

14 samples harvested: 12 ran (of which 0 constructed an object rather than computing a result; 10 with no instrument failure: 1 honoured, 0 violated, 9 with no contract checked; 2 where Syntology's instrument failed) · 2 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
2ran · our draft was wrong
9ran
2unverified

Licence: 0 of the 14 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 ajyl/dpo_toxic. “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.

ajyl/dpo_toxic official repository 14 samples, 12 ran (2 instrument-failed) · 2 unverified
dpo_loss in 21 papers toxicity/train_dpo/trainers.py ran MIT (permissive) · 6e420696373414b1 · report fixture: signature policy_pos_logps:Tensor/1/float, policy_neg_logps:Tensor/1/float, ref_pos_logps:Tensor/1/float, ref_neg_logps:Tensor/1/float, beta:float/0/float, reference_free:bool/0/bool · returned tuple · level recorded 2026-09-25
get_gold toxicity/eval_interventions/generate_funcs.py ran MIT (permissive) · 813d811b5f586d21 · report fixture: signature model:callable/0/?, data:dict/0/?, config:dict/0/? · returned dict · level recorded 2026-09-25
get_intervene_vector toxicity/eval_interventions/hook_utils.py ran · honoured contract MIT (permissive) · 2eb9d77d720c2173 · report fixture: signature model:callable/0/?, config:dict/0/? · returned Tensor · level recorded 2026-09-24
get_kl_div toxicity/train_dpo/trainers.py ran MIT (permissive) · 9e84c0d9c78a5cad · report fixture: signature kl_criterion:callable/0/?, pos_pi_logits:Tensor/2/float, neg_pi_logits:Tensor/2/float, pos_ref_logits:Tensor/2/float, neg_ref_logits:Tensor/2/float · returned tuple · level recorded 2026-09-24
get_local_dir in 14 papers toxicity/train_dpo/dpo_utils.py ran · our draft was wrong MIT (permissive) · 8ab43db002bf984d · report fixture: signature prefixes_to_resolve:list/1/str · returned str · level recorded 2026-09-25
get_local_run_dir in 14 papers toxicity/train_dpo/dpo_utils.py ran MIT (permissive) · 1df5ede8f8950a70 · report fixture: signature exp_name:str/0/str, local_dirs:list/1/str · returned str · level recorded 2026-09-25
get_prompts in 21 papers toxicity/eval_interventions/generate_funcs.py ran MIT (permissive) · d635938df605fbd3 · report fixture: signature model:callable/0/?, data:dict/0/?, config:dict/0/? · returned dict · level recorded 2026-09-25
get_remote_file in 16 papers toxicity/train_dpo/dpo_utils.py ran MIT (permissive) · e7ce4c19888c6df6 · report fixture: signature remote_path:str/0/str, local_path:str/0/str · returned str · level recorded 2026-09-25
load_data in 700 papers toxicity/eval_interventions/eval_utils.py ran · our draft was wrong MIT (permissive) · cba91edd374a3e28 · report fixture: signature data_config:dict/0/? · returned list · level recorded 2026-09-25
run_dummy toxicity/eval_interventions/metric_funcs.py ran MIT (permissive) · f5813daae4f9b8f5 · report fixture: signature model:callable/0/?, data:list/1/tuple, intervene_results:dict/0/?, config:dict/0/? · returned int · level recorded 2026-09-25
run_f1 toxicity/eval_interventions/metric_funcs.py ran MIT (permissive) · 044b24804e21d67d · report fixture: signature model:callable/0/?, data:dict/0/?, intervene_results:dict/0/?, config:dict/0/? · returned float64 · level recorded 2026-09-25
tokenize in 176 papers toxicity/eval_interventions/eval_utils.py ran MIT (permissive) · 9727f348a587c477 · report fixture: signature tokenizer:callable/0/?, data:list/1/dict, config:dict/0/? · returned dict · level recorded 2026-09-24
get_svd_u_vec toxicity/eval_interventions/hook_utils.py unverified MIT (permissive) · 0e4cc3ba12263f9d · report
rank_value_vecs toxicity/eval_interventions/hook_utils.py unverified MIT (permissive) · 43d32c9d46ad3c9a · report

Tasks

Language ModelingLanguage Modelling

Results from the paper archive 2025-07-28

No leaderboard rows for this paper in the archive.

Methods

DPO

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