Search papers, labs, and topics across Lattice.
This paper critiques the conventional understanding of large language models (LLMs) by highlighting the significant role of undisclosed inference-time steering in shaping their outputs, which is often overlooked in evaluations. It introduces three key concepts鈥擨nference Attribution Problem, Probability Placement, and Inference Policy Transparency鈥攖hat elucidate how external interventions can systematically bias generated language towards specific frames, complicating the attribution of behavior solely to model weights. The findings emphasize the need for greater transparency and governance in LLM deployments, particularly in light of regulatory frameworks like the EU AI Act.
Undisclosed inference-time steering can systematically bias LLM outputs, challenging the assumption that model weights alone dictate behavior.
Large language models (LLMs) are commonly evaluated under the assumption that their observable behavior is primarily determined by model weights, training data, alignment procedures, and user prompts. This view is incomplete. Modern inference pipelines may systematically modify the probability distribution produced by a model immediately before token selection, creating an additional layer of control between frozen weights and observed text. While controlled generation (e.g., PPLM, GeDi, DExperts, FUDGE) and text-watermarking systems (e.g., SynthID-Text) demonstrate the technical maturity of decoding- and logit-level interventions, the governance, security, and economic implications of an undisclosed inference policy remain comparatively underexplored. This paper examines the emergence of inference-time framing bias: the systematic modification of generated language toward political, ideological, institutional, or commercial frames via interventions applied after model inference but before token sampling. We formalize the operational reality Model != Deployed System and introduce three concepts: (1) the Inference Attribution Problem, characterizing why observed behavioral bias cannot generally be causally attributed to model weights alone under limited observability; (2) Probability Placement, defining a hypothetical advertising primitive in which commercial influence is implemented through systematic shifts in generation probabilities rather than explicit product insertions; and (3) Inference Policy Transparency, a governance principle for making deployment-layer interventions auditable. We examine these concepts in relation to Article 5 of the EU AI Act, the EU Digital Services Act, and FTC doctrines.
