MatanBT/tropt-optbench-emb-triggers

TROPT Optimizer Benchmark — Corpus Poisoning (Embedding) Triggers

Optimized adversarial passages from ranking 15 discrete optimizers on concept-specific corpus poisoning, and their retrieval evaluation against MatanBT/msmarco-concepts.

Default FLOP budget per model

Every optimizer within a model runs under the same budget. Rows at this budget carry budget_label = "full".

model	encoder	params	default budget (FLOPs)
minilm	sentence-transformers/all-MiniLM-L6-v2	0.022B	3.6e+14
e5	intfloat/e5-base-v2	0.11B	1.3e+16
qwen3emb	Qwen/Qwen3-Embedding-0.6B	0.6B	1.3e+16
qwen3emb8b	Qwen/Qwen3-Embedding-8B	7.57B	2.1e+17

Grid

4 models x 15 optimizers x 8 concepts x 3 seeds x 2 trigger lengths (30, 100).

Optimizers: adv_decoding, arca, autoprompt, beast, gaslite, gbda, gcg, hotflip, mac, mcpal, pal, pez, qcg, ral, random_search.

Files

file	one row per
triggers.parquet	(run, budget_label)
eval_triggered_messages/<model>.parquet	trigger

budget_label is one of 1pct, 2pct, 5pct, 10pct, 20pct, 30pct, 42pct, 60pct, 80pct, full — snapshots of the same run truncated at that share of its budget. budget_flops gives the absolute cap.

Key columns

triggers.parquet — uid, model_short, optimizer_name, concept, seed, trigger_len, budget_label, budget_flops, best_trigger_str, best_loss, best_cos_sim, trigger_flops, n_steps_within_budget, optimized_instruction (template, with {{OPTIMIZED_TRIGGER}}), adv_passage (resolved), mal_info, heldin_queries, run_id.

eval_triggered_messages/<model>.parquet — trigger_uid, heldin_cos_sim, heldout_cos_sim, heldin_mean_cos_sim, heldout_mean_cos_sim, heldin_appeared@10, heldout_appeared@10, heldin_ranks, heldout_ranks, n_heldin_eval, n_heldout_eval, n_corpus.

heldin_cos_sim / heldout_cos_sim are computed exactly as the training loss: cosine similarity to the centroid of the respective query split.

Equal compute, and where it does not hold

The budget is what stops most optimizers: gcg, gaslite, mac, mcpal, pal, ral, qcg, arca, autoprompt, gbda and random_search all finish at ~100% of it, so comparing them at budget_label = "full" is a straight equal-compute comparison.

Four do not, because a step cap or their own termination stops them first — median budget actually consumed, per model (minilm / e5 / qwen3emb / qwen3emb8b):

optimizer	stopped by	budget consumed
hotflip	1,500-step cap	2.5% / 0.8% / 3.8% / 3.8%
beast	own termination	65% / 3.5% / 8.0% / 5.4%
adv_decoding	one step per trigger token	101% / 16% / 37% / 23%
pez	60,000-step cap	88% / 23% / 91% / 90%

pez's cap is set where it has provably converged: over the uncapped runs the step at which best_loss first comes within 0.005 cos-sim of its final value is at worst 45,608, and capped runs reproduce the uncapped results exactly on the model spot-checked. The cap buys a 3.5x wall-clock saving on e5, which is otherwise the arm's dominant cost.

Companion

MatanBT/tropt-optbench-triggers — the same benchmark in the LLM-jailbreak domain.