Vision-language models (VLMs) often generate fluent but visually unsupported descriptions, especially by mentioning objects absent from the image. We propose QK Product Steering, a data-free, training-free, and zero-inference-cost weight edit for reducing object hallucination. The method directly edits the per-head query-key product, the operator that produces pre-softmax attention logits, by suppressing a small number of dominant singular modes in selected middle layers. The edited product is then mapped back to the query weights through a closed-form query-only update while keeping shared key weights fixed, making the edit compatible with grouped-query attention. We further decompose the QK product into symmetric and antisymmetric components to distinguish mutual content-similarity patterns from directional attention patterns. Across three GQA-based VLMs, QK Product Steering achieves an average relative CHAIR$_s$ reduction of $4.0\%$, while matched random-mode controls show negligible change. Interpretability ablations show that the hallucination signal is specific to dominant QK modes and is primarily localized to the symmetric mutual-attention channel. Overall, QK Product Steering offers a simple alternative to decoding-time mitigation, requiring no additional data, fine-tuning, or inference-time overhead while largely preserving general multimodal capability.
Aristotelis Lazaridis, Aman Sharma, Dylan Bates +3cs.LG cs.AI
The Gemma 4 instruction-tuned models share a reproducible failure: on long factual enumeration prompts, such as TV episodes, the 88 IAU constellations, or the 151 original Pokemon, they collapse into repetition, either a tight verbatim loop or a list whose entries decay onto one answer. These loops reach 87.5% (7/8 generations) and survive prompt rewording and most sampling adjustments. In this paper, we explore whether edits to a few internal model components can directly reduce this failure, without relying on repetition penalties, which can distort valid repetition and degrade task performance. To locate such targets, we combine per-layer ablation with per-neuron or routed-expert attribution, then evaluate weight edits over complete generations. We find that these edits substantially reduce detected loops on the prompts and seeds used to select them; in Gemma 4 E2B, for example, one sign-inverted neuron suffices. Across all four Gemma models, detected loops fall from 46/384 to 12/384 on frozen held-out prompts and seeds, driven mainly by E4B and 31B, while general-purpose benchmarks show no statistically detectable regressions. Our attribution methodology identifies useful candidates, but rankings vary across examples. At longer generation budgets, edits remain effective for E2B and E4B, whereas remaining failures in 26B and 31B shift toward doom looping: non-convergent self-correction over facts the model cannot recall. In exploratory experiments on Qwen3.5 and LFM2.5, sparse edits also reduce repetition, providing preliminary cross-family evidence, although effect strength and selectivity vary. Overall, our results show the promise and limits of targeted, small-scale weight editing: it can suppress specific repetition failures and provide a training-free causal intervention, but does not reveal a universal loop circuit, guarantee clean termination, or supply missing knowledge.
Abliteration, the removal of refusal capabilities from large language models by projecting weight matrices orthogonal to an extracted refusal direction, has emerged as a prominent safety concern through its ability to bypass post-training alignment using only a small set of contrastive prompts. We find that existing defenses commonly overlook the cause of abliteration; that is, how easily the refusal direction can be extracted. To hinder this process, we introduce a weight-editing method that obscures the refusal signal by applying rank-$k$ updates to residual stream writer matrices while replacing refusal-inducing activations with random aliases and correcting downstream reader matrices to preserve the model's original behavior. On Llama-3-8B, AMRA improves post-abliteration refusal scores by $2.16$ points over the undefended baseline with less than $0.5$ percentage points of MMLU degradation. On Gemma-2-9B, it improves the post-abliteration refusal by $14.70$ points over the baseline while keeping harmful output rates similar to the baseline, albeit at a greater utility cost.