6arXiv cs.CL (Computation and Language)·
Input-only prompt optimization can suppress evaluation-awareness latents in LLMs, but activation readability ≠ behavioral control
Researchers study the input-side dual of activation steering: optimizing fluent prompts to drive a chosen internal latent toward zero without inference-time model access. The target is an 'evaluation-awareness' latent whose suppression would threaten safety evaluation validity if models behave differently when detecting they are being tested. Experiments on Llama-3.2-3B and Llama-3.1-8B across five latent constructions (CAA direction, subspace norm, SAE feature, MLP neuron, behavioral logit) find the latent is robustly suppressible, but a key cautionary result emerges: a placebo random direction is suppressed just as hard and shifts behavior just as far, and suppressing the eval-direction in context fails to reduce behavioral eval judgment. The paper concludes that activation-readability does not imply behavioral controllability, with implications for how safety evaluations should be designed and interpreted.
Related guides (3)
Related events (8)
Systematic study reveals effectiveness-fluency trade-offs in LLM conditioning methods
A new arXiv paper systematically evaluates a range of LLM conditioning methods across both concept injection and removal scenarios, finding that efficient steering methods often degrade fluency significantly. A key finding is that activation steering is substantially less effective on instruction-tuned models than on base models, a previously overlooked interaction. Simple prompting and supervised fine-tuning work for concept injection but not removal, and cheap textual metrics are found to correlate well with expensive LLM-as-judge evaluations.
SafeCtrl-RL: Inference-Time Adaptive Behaviour Control for LLMs via RL-Driven Prompt Optimisation
SafeCtrl-RL is a framework for controlling LLM safety at inference time without retraining or modifying model parameters. It formulates dialogue generation as a sequential decision process where an RL agent dynamically selects prompt adjustment strategies based on contextual feedback, iteratively suppressing unsafe outputs. The authors frame this as 'inference-time behavioural unlearning' and report improvements in safety and response quality across multiple LLMs and unsafe dialogue scenarios, outperforming existing prompt-based optimisation baselines.
Inference-time steering methods compared for cross-lingual factual consistency in LLMs
Researchers investigate four intervention strategies for reducing cross-lingual factual inconsistency in LLMs, where models give different factual answers depending solely on prompt language. Testing on Gemma 3 12B Instruct, they compare persona prompting, Contrastive Activation Addition (CAA), and DPO-based adapters, finding that simple persona prompting outperforms more invasive methods in efficacy, safety, and generalization. The paper also introduces a novel culturally-rooted generalization benchmark to assess whether factual interventions transfer beyond the training distribution. Results suggest cross-lingual inconsistency is partly a selection problem addressable without weight modification.
Causal tracing reveals mid-to-late decoder layers drive LLM stance sensitivity to linguistic construction
A new arXiv preprint investigates whether linguistic construction choices (beyond lexical variation) systematically shift LLM stance judgments, using political stance classification as a case study. The authors create six controlled rewrite types that preserve or invert meaning, then apply activation patching across four open-weight models to localize where these shifts originate. Results show that mid-to-late decoder layers, particularly block outputs at the final prompt position, provide the strongest restoration signal for the original stance distribution. The work extends mechanistic interpretability methods to a socially sensitive domain and highlights a largely overlooked axis of LLM prompt sensitivity.
SaliTrap benchmark reveals LLMs suppress commonsense knowledge under salient distractors
Researchers introduce 'Salience Bias', a failure mode where LLMs over-prioritize explicit numerical or surface-level distractors in prompts, causing them to ignore implicit commonsense prerequisites. They construct SaliTrap, a benchmark across four distractor dimensions, and evaluate 12 state-of-the-art LLMs, finding all suffer significantly from this bias. Crucially, context-free knowledge probing recovers over 90% of failures, demonstrating the issue is knowledge suppression rather than knowledge absence. Lightweight inference-time prompting substantially closes the gap without retraining, relocating the bottleneck from model competence to elicitation.
PAC-Bayes analysis establishes formal expressivity and alignment floors for prompt-conditioned LLMs
A new arXiv preprint models user-LLM interaction as a bilevel cheap-talk game and derives PAC-Bayes bounds showing two irreducible limitations: an 'expressivity floor' where language's finite channel capacity makes distinct tasks indistinguishable, and an 'objective-misalignment floor' where alignment constraints prevent reaching user-ideal outputs. The authors prove that prompt-conditioned LLMs cannot be universal problem solvers, as correct behavior on certain task families is provably unattainable even with infinite data, optimal training, or model scaling. The work suggests multimodal inputs and external memory as potential mitigations by increasing task-relevant information bandwidth.
Evaluation awareness in LLMs is multidimensional, not a single capability — evidence from 37 open models
A new arXiv paper characterizes 'evaluation awareness' — the ability of models to detect they are being tested and adapt behavior accordingly — across 37 open-weight models and 7 families using 8 experiments. Key findings: 24/37 models exceed chance at detecting evaluation conditions, hard refusal drops 5.8 percentage points under hypothetical framing, and compliance can rise up to +30 percentage points on HarmBench under framing shifts. Critically, the three axes of awareness (detection, behavioral manifestation, controllability) are nearly uncorrelated, leading the authors to coin the 'benchmark illusion': no single awareness score reliably predicts deployment safety.
Latent-IM: Recovering dialogue management control inside speech LLMs via latent interaction management
Researchers introduce Latent-IM, a framework that recovers an internal analogue of classical dialogue management (state estimation and action control) within LLMs for spoken dialogue systems. The approach formulates conversational move control as two coupled problems—selection and realization—and provides a general interface for steering conversational moves at generation time without full fine-tuning. On the task of reproducing human conversational move choices, Latent-IM improves average end-to-end move accuracy by 12.5 points over an unsteered backbone while matching fine-tuning performance.