The study investigates where a large audio language model (LALM) derives its refusal responses to harmful spoken requests. By applying causal interventions at the audio-to-language-model interface and within the language-model residual layers, the authors find that the primary influence on the refusal margin comes from mid-to-late layers of the text language model rather than the audio front end. Ablations of the audio interface have minimal effect, while zeroing the encoder output still reduces the margin, indicating the audio pathway remains active but is not the main source of refusal decisions.
arXiv:2608. 19211v1 Announce Type: cross Abstract: Human speech is richly expressive, with prosody carrying linguistic and emotional information beyond the lexical content.
arXiv:2609.05871v1 Announce Type: cross Abstract: Audio-conditioned language models often underuse acoustic cues such as prosody, emotion, and non-speech sounds, raising the question of whether ASR-s...
The study investigates how safety alignment in large language models, trained mainly in English, transfers to other languages. While models show near-perfect harmfulness detection (AUROC > 0.98) using unrelated harmless prompts (easy negatives), performance drops sharply in low‑resource languages when using surface‑similar benign prompts (hard negatives). This degradation persists across multiple languages and models, indicating that easy‑negative evaluation alone cannot confirm cross‑lingual harmfulness representation quality.
arXiv:2607. 14147v1 Announce Type: cross Abstract: Aligned language models refuse harmful requests, but a one-line prefill ("Sure, here is") strips the refusal.
The study investigates the internal workings of an audio language model (Qwen3-Omni) by applying a logit lens to its middle layers. It finds that the model’s reasoning about spoken questions becomes legible in words before any token is emitted, revealing language‑agnostic, paralinguistic, and temporally distinct signals that are causally used in the network’s decision process. The authors demonstrate that these signals can be isolated and mapped to specific layers, providing a qualitative account of how the model processes audio input.
The paper presents the first music‑specific, layer‑wise empirical study of hallucination in audio‑language models, framing it as a hierarchical perceptual grounding failure across five layers: sound events, temporal properties, tonal attributes, style, and emotion. It introduces MuseDiag, a diagnostic framework that evaluates nine models and finds universal vocal misperception, significant tonal perception differences, and identifies Audio‑Flamingo‑3 as the most stable model. The study also proposes two training‑free mitigation methods, ADD‑M and TPA, which reduce hallucination in probing but show variable effectiveness in free‑form generation, highlighting the need for multi‑paradigm evaluation.
arXiv:2606. 13720v1 Announce Type: new Abstract: Arditi et al.
arXiv:2604. 04385v5 Announce Type: replace-cross Abstract: We localize the policy routing mechanism in alignment-trained language models.
arXiv:2609.22219v1 Announce Type: cross Abstract: Evaluating an aligned language model by reading its answers assumes the answers carry the distinction the evaluator cares about. We introduce LLM end...
arXiv:2602. 06941v2 Announce Type: replace-cross Abstract: Large language models can recover mid-generation from task-misaligned activation steering, producing explicit verbal restarts (e.
arXiv:2608. 12334v1 Announce Type: cross Abstract: Despite the impressive multilingual capabilities of Large Language Models, the latent dynamics dictating language selection remain poorly understood.
The paper investigates whether the layer that yields the highest probing accuracy in omni‑modal large language models is also the most effective for steering interventions. Across three independently developed models, the authors find that the best probing layers differ widely, whereas the most steerable layers consistently lie in a narrow mid‑to‑late range of the network. Using emotion as a testbed, they demonstrate a significant causal gap between probing and steering, and propose a two‑factor account linking readability and downstream plasticity to steering effectiveness.