acceptodds
Under review as a conference paper at ICLR 2027

System Attention Bias: Zero-Tax Policy Alignment via Inference-Time Logit Modulation

Abstract

Large language models follow system-prompt policies unreliably, and improving compliance usually requires fine-tuning. We introduce System Attention Bias (SAB), a training-free intervention that adds log β to the pre-softmax attention logits wherever a non-system query attends to a system key, multiplying those keys’ unnormalised attention scores by β. SAB adds no parameters, auxiliary model or forward pass, and β can be changed per request on a loaded model. We pair it with a calibration protocol that selects β under a capability budget and evaluate on five instruction-tuned models (0.5B–8B, four families). With β selected on held-out AdvBench splits, SAB gives a lower AdvBench harmful-compliance rate than the safety prompt for Qwen-0.5B and SmolLM2-1.7B (15.3% → 10.1%) at 93.9– 96.6% MMLU retention; the other models are at the floor, transfer to HarmBench is not significant for any model, and TinyLlama-1.1B’s drop from 87.5% to 8.8% is an artefact: it consists of degenerate responses that restate the system prompt, and a control shows that it comes from our implementation also biasing the beginning- of-sequence token and vanishes when only the system block is biased. Compared under the HarmBench classifier with prompt repetition and classifier-free guidance, SAB is not the strongest method: all three lower harmful compliance only by refusing more safe requests (XSTest), and at matched over-refusal they coincide on AdvBench. SAB reaches that operating point in one forward pass (0.2–1.7% overhead per token) where guidance needs two, and, unlike repetition, offers a continuous dial. On role-specific policies (CoSApien), SAB raises Llama-3.1- 8B’s CoSA-Score from 0.423 to 0.457 under greedy decoding, a gain that does not persist under sampling. Over-biasing degrades both safety and capability, so calibration is necessary. The reference implementation uses a dense O(n2 ) mask; the same bias can also be computed inside a fused FlexAttention kernel without materialising it.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
System Prompt Compliance, Inference-Time Intervention, Attention Steering, LLM Alignment & Safety, Training-Free Policy Enforcement

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_WEyTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes System Attention Bias (SAB), a training-free inference-time intervention that adds +logβ to the attention logits from non-system queries to system-role keys. The authors calibrate β under an MMLU capability budget and evaluate safety, over-refusal, role-specific policy compliance, attention dilution, and implementation overhead across several instruction-tuned LLMs. The paper reports safety gains for Qwen-0.5B and SmolLM2-1.7B under held-out AdvBench calibration, while also documenting important limitations including over-refusal, weak HarmBench transfer, non-monotonicity, and a substantial TinyLlama implementation artefact.

Critical strengths
  1. The mechanism is simple, well motivated, and computationally attractive. Equation (1) precisely defines a role-conditioned additive attention mask, and Equation (2) correctly makes explicit that the intervention multiplies the unnormalised attention score of each selected system key by β.

  2. The paper includes useful practical analyses rather than treating the intervention as free by assertion. The per-token comparison in Table 16 isolates the bias cost from differences in generation loops and reports a modest 0.2%to 1.7% overhead. The authors also appropriately distinguish this from the misleading end-to-end “speedups” in Table 15, which are caused by response-length and generation-loop differences.

Critical weaknesses
  1. No head-to-head against the closest work.The paper calls Spotlight "the closest mechanism" but never runs it (or PASTA). Table 5 is just a structural checklist with author-defined categories — it says nothing about real trade-offs.

  2. The biased span is wrong for 3 of 5 models. Exact only for Qwen and SmolLM2. TinyLlama's effect is a BOS artefact, Phi-3.5 leaks a user tag, Llama-3.1 includes BOS and drops a delimiter. So Eq. (1) isn't actually role-conditioned there, and those results need recomputing.

  3. The "arbitrary policy" claim is one cherry-picked point.One model, 200 prompts, greedy decoding, β picked from an 8-value sweep on the same benchmark with no held-out split. The curve is non-monotonic, and the gain reverses under sampling (0.461 vs 0.434).

Questions for authors

Same as weakness.

Reviewer_bi3ROpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents System Attention Bias (SAB), a training-free intervention that adds a constant bias to pre-softmax attention logits from non-system queries to system keys. A calibration protocol is presented to choose the constant bias that maximizes task-specific adherence score while ensuring that capability degradation on unrelated tasks is under a fixed budget. The empirical evaluation includes five instruction-tuned models and three benchmark domains (safety, role-specific compliance, and capability). SAB outperforms a generic prompt and a safety prompt without bias on AdvBench, while matching or outperforming these baselines on 3 / 5 models for HarmBench. SAB outperforms 2 published baselines on CoSApien.

Critical strengths
  1. The paper is written well and details of SAB and the empirical evaluation are easy to follow.

  2. The use of the capability threshold in the proposed calibration protocol is novel to the best of my knowledge and the associated empirical analysis is thorough.

Critical weaknesses
  1. I find it difficult to appreciate the novelty of the contributions of this work as it appears identical to other attention steering techniques from literature. The proposed additive bias to pre-softmax logits across all attention heads appears equivalent to InstABoost [1]. The identification of the boosted attention span is trivial given the system role tokens. I acknowledge that the proposed calibration protocol is different from that used by InstABoost but the capability threshold is not a sufficiently novel methodological contribution in my opinion.

  2. The empirical evaluation does not include other attention steering methods, such as PASTA and Spotlight mentioned in related work.

[1] Guardieiro, Vitoria, et al. "Instruction Following by Principled Boosting Attention of Large Language Models." arXiv preprint arXiv:2506.13734 (2025).

Questions for authors
  1. Could you shed some light on how SAB is mathematically different from InstABoost? Please see weakness 1 for more details.

  2. Could you comment on how SAB would compare to other attention steering methods (PASTA, Spotlight) on the benchmarks evaluated here?

Reviewer_gb7XOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SAB is a training-free method that boosts a model’s attention to system prompts using a single adjustable parameter. Tests across five models show modest safety gains on AdvBench for two smaller models, with little transfer to HarmBench. Its biggest reported gain is traced to an implementation artefact, and it does not outperform prompt repetition or CFG overall. SAB’s main practical advantage is that it is lightweight and adjustable at inference time; its CoSApien improvement appears only with greedy decoding.

Critical strengths
  1. Exceptional empirical honesty and rigour in self-auditing. The authors identify and report their own implementation defect (BOS biasing), run span controls, quantify cross-stack reproducibility, and explicitly present SAB as not dominating baselines on the safety/over-refusal frontier. This materially raises confidence in every remaining positive claim and is a model of responsible reporting.
  2. A clean, deployable mechanism with a principled calibration protocol. SAB needs no training, no auxiliary model, one forward pass (0.2–1.7% per-token overhead), and a runtime-adjustable scalar applicable to arbitrary system-prompt content—distinct from PASTA/Spotlight (per-input annotation) and CFG (two passes). The Pareto-frontier calibration (Def. 1/Eq. 3) is a reusable contribution beyond this specific intervention.
Critical weaknesses
  1. Limited demonstrated efficacy. Under the unified HarmBench classifier, SAB significantly reduces AdvBench HCR over the safety prompt for only two small models; HarmBench transfer is non-significant for all models; and at matched over-refusal SAB coincides with prompt repetition and CFG. The practical case therefore rests on efficiency (one pass, continuous dial) rather than on safety gains—real, but modest.
  2. Mechanism validation rests on a narrow, partly flawed empirical base. The strongest observed effects are either artefacts (TinyLlama) or confined to two ≤1.7B models with exact spans; the role-policy gain vanishes under sampling; and scale (70B+), RAG/long-context interaction, and adversarial multi-turn attacks are untested.
Questions for authors
  1. For Llama-3.1-8B and Phi-3.5, the biased span includes BOS/<|end|> tokens. Have you re-run the main Table 1/2 results with the exact system block for these models, and do any conclusions change beyond the 2–7 point shifts in Table 25?
  2. Given HarmBench non-transfer, do you believe β calibrated on AdvBench generalizes to any policy distribution other than AdvBench-style requests? What evidence supports deployability on a different policy type?
  3. Why does over-biasing increase HCR for Llama-3.1-8B (4.6% at β=10 vs. 0.4% baseline)? You state the mechanism is not isolated—can you offer even a hypothesis, since this bounds safe deployment?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.