Apple Proposes DLR-Lock to Protect Open-Weight Models; Releases DeepAmbigQA Benchmark

Apple's machine learning research group published two notable works this week. The first, DLR-Lock, addresses a growing tension in AI: how to release open-weight models without enabling unauthorized fine-tuning for harmful or off-policy uses. DLR-Lock replaces each pretrained MLP with a deep low-rank residual network, structurally resisting fine-tuning that would alter the model's behavior while preserving inference-time performance. It's a technical contribution to the open-weights safety debate that Dario Amodei and others have been vocal about — and that Imbue's Kanjun Qiu pushed back on, arguing 'information wants to be free' and control-based approaches are futile.

The second, DeepAmbigQA, is a 3,600-question benchmark for ambiguous multi-hop reasoning. Its striking finding: even GPT-5 struggles with questions that have incomplete or ambiguous answers, exposing a persistent weakness in frontier models' handling of uncertainty — they tend to confabulate complete answers rather than acknowledge ambiguity.

Apple also published research characterizing diffusion versus autoregressive language models across NLP tasks, and introduced Arbitrage, an advantage-aware speculative decoding method that extends draft-and-verify decoding to cut inference cost for long chain-of-thought reasoning.

The throughline is that Apple's public AI presence this week is entirely research-driven — no consumer Siri or Apple Intelligence model launches, consistent with reports that its upgraded Siri is 'coming later this year.' Competitively, Apple continues to signal deep ML research strength while lagging on shipped consumer AI products, a gap rivals like Google (despite its turmoil) and OpenAI keep widening. The research is credible and cited; whether it translates into competitive products remains Apple's central AI question.