Decoy Direction Optimization: A Post-Hoc Defense Against LLM Abliteration · Research brief

Request deep report Read original paper ↗View page data (JSON)Download research skill ↓Discuss this paper

For solo builders

A possible application for this research area. Paper-specific feasibility still needs a full-paper review.

An agent reads external documents or tool responses that may contain malicious instructions.

What you would change

Replay attacks and legitimate lookalikes in an isolated copy of your agent workflow.

Try the idea on your product

A proposed first test. Set your budget and success criteria before running it.

Replay attacks and legitimate lookalikes in an isolated copy of your agent workflow.

Compare against

Your existing checks with the same tasks, permissions, and tool responses.

Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.
Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.
Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.
Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test.

What to measure

Successful attacks missed, legitimate actions blocked, detection delay, and review cost.

When to keep it

Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation.

Download test plan ↓

abstract indexedPaperraft next step

Do not allocate compute yet.

This page uses the abstract and public adoption signal only. Request the deep report before relying on its recommendation.

Abstract-level screening only. Risks and infrastructure may not have been evaluated. This recommendation is not a reproduced result.

minimum test setup: not reported
published evidence setup: not reported
source-linked claims: 0 of 2
exposure checks: 0 of 8

Why it was shortlisted

The technical change, research area, and reason it entered the archive.

recommendation: test
research area: safety and guardrails
technical change: Decoy Direction Optimization against refusal ablation
published: 2026-09-14

The defense is cheap and post-hoc, making it practical for a small team shipping open-weight models, but the 65% worst-case ASR under adaptive attacks means it must be validated against the reader's specific threat model before adoption.

Central equations

Original equations, typeset for reading. Paperraft explanations connect the mechanism, its aim, and the evidence.

DDO optimizes decoy directions ${𝐮_{i}}$ via gradient descent through the defended model, while all base-model weights remain frozen. Scalar gains $(β, s)$ are treated as hyperparameters and tuned via Bayesian hyperparameter search (Optuna; Appendix F.6). The optimization minimizes a composite loss over 128 harmful and 128 safe probes:

ℒ = λ_{ref} ​ ℒ_{refuse} + λ_{ret} ​ ℒ_{retain} + λ_{conf} ​ ℒ_{confusion} + λ_{rs} ​ ℒ_{refusal-score},

equation (3) · §3 Method: Decoy Direction Optimization (DDO) · loss function

where $𝐮_{i} ⟂ {\hat{𝐫}}_{ℓ}$ is a decoy output direction and $β, s > 0$ are scalar gains. For SwiGLU in replace mode, writing $a_{ℓ} ≜ ⟨ {\hat{𝐫}}_{ℓ}, {\tilde{𝐱}}_{ℓ} ⟩$ for the neuron’s refusal coordinate, its decoy contribution is

δ_{i} ​ ({\tilde{𝐱}}_{ℓ}) = s ​ g_{β} ​ (a_{ℓ}) ​ 𝐮_{i}, g_{β} ​ (a) ≜ a ​ silu (β ​ a) = β ​ a^{2} ​ σ ​ (β ​ a),

equation (2) · §3 Method: Decoy Direction Optimization (DDO) · proposed method

DDO repurposes a small set of SwiGLU (or GeGLU) units to implement a gated read–write map: each unit reads the refusal coordinate and writes into an orthogonal decoy direction. To minimize utility loss, we choose $n$ low-impact units per layer by selecting the $n$ smallest-norm columns of $𝐖_{down}$ (units whose down-projection contributes weakly to the residual update). In each targeted layer $ℓ$ and for each selected unit $i$ , we write a refusal-aligned trigger into row $i$ of $𝐖_{up}$ and $𝐖_{gate}$ (denoted $𝐰_{up, i}$ and $𝐰_{gate, i}$ ), and a decoy write vector into column $i$ of $𝐖_{down}$ (denoted $𝐰_{down, i}$ ):

𝐰_{up, i} \leftarrow {\hat{𝐫}}_{ℓ}^{⊤}, 𝐰_{gate, i} \leftarrow β ​ {\hat{𝐫}}_{ℓ}^{⊤}, 𝐰_{down, i} \leftarrow s \cdot 𝐮_{i},

equation (1) · §3 Method: Decoy Direction Optimization (DDO) · proposed method

equations from arXiv HTML · fetched 2026-10-11

Claims and evidence

A claim is source-linked only when the PDF page and matching excerpt are available.

claim 01Paperraft inference
Read the brief interpretation
DDO replaces per-checkpoint safety finetuning against Refusal Feature Ablation with a post-hoc weight edit that injects a high-magnitude nonlinear decoy signal into MLP neurons, corrupting the attacker's contrastive estimator of the refusal direction. It requires no base-model finetuning and runs at 30 to 450 times lower optimization cost than trained baselines, so it is feasible on a single 24 GB GPU. It can fail under adaptive multi-phase attacks, where worst-case ASR still reaches 65%, and it only defends against ablation-style attacks, not finetuning-based removal of safety behavior or other jailbreak classes.
This is a Paperraft inference. It is not pinned to a verified PDF passage.
Inspect the source
claim 02Paperraft inference
Reduces ASR under standard RFA to below 10% across six model families and cuts Heretic attack ASR from 88.7% to 18% on Llama-3-8B-Instruct, at 30 to 450 times lower optimization cost per configuration than trained defense baselines; under adaptive multi-phase attacks it remains comparable to trained defenses (65% vs. 58% worst-case ASR).
This is a Paperraft inference. It is not pinned to a verified PDF passage.
Inspect the source

Exposure map

Author findings, Paperraft interpretations, and the next check. Missing analysis never counts as evidence of safety.

quality

Still unknown

Which tasks improve, and which regress against the baseline?

compute and memory

Still unknown

What memory, hardware, and total compute does the complete method require?

latency

Still unknown

Are startup time, typical response time, and slow requests measured separately?

operations

Still unknown

What must be monitored, maintained, and rolled back?

software compatibility

Still unknown

Does the implementation support your model, runtime, and hardware?

security

Still unknown

What new data access, permissions, or trust boundaries does it introduce?

data and training

Still unknown

What data preparation or training is required, and is it included in the cost?

reproducibility

Still unknown

Are code, versions, inputs, and evaluation scripts available?

Continue the investigation

A reading guide for safety and guardrails.

Editorial prompts, not findings from this paper or a validated test plan. Detailed analysis is still pending for the topics below.

What to inspect

Which assumptions does the proposed method need, and where might your workload violate them?

Plan a fair comparison

Identify the baseline, metric, dataset, and experimental settings in the paper. Which can you match?

Questions for adoption

What implementation and evaluation artifacts are available, and what would you need to supply?

Read the paper with these questions

Inside the paper

Preview the first 3 pages of the original PDF. Click the stack to turn the page.

The downloadable skill includes SKILL.md and evidence.json, with recorded findings and unanswered checks. TRY-IT.md adds a product test plan and a blank results sheet.

A reading aid for your agent. Findings still need validation.

Page 1 of 3

Open this page in the PDF ↗

Provisional research brief updated 2026-10-18. Paperraft has not reproduced this experiment.