Diffusion safety, Language model alignment, and Artificial Cognition

Modular, decode-time alignment without retraining the backbone.Repurposes speculative decoding's draft/verify split into an alignment mechanism: a fast Drafter proposes candidate reasoning steps, and swappable "Blade" auditor models — each trained on a different objective — score and select between them at inference time.

• Mode B, the proposed strategy: candidate steps compete in a probabilistic Elo tournament scored by a Blade reward model, and the champion is accepted unconditionally — no Verifier forward pass, no rejection sampling.
• Tournament respects score uncertainty (Thurstonian win probability), letting a lower-scoring but more confident candidate beat a higher-scoring but noisier one.
• Alignment objectives live in independently swappable Blade heads rather than the base model — change what's being optimized for without retraining the generator.
• Benchmarked against ARGS, DeAL, MOD, Best-of-N, and RRM on HH-RLHF harmlessness/helpfulness and TruthfulQA.
• Mitigates degenerate decode-time failure modes — refuse-always, templated boilerplate — that simpler preference-selection strategies fall into.

Speculative safety decoding for text-to-image generation. A multi-stage system that audits diffusion models mid-generation and intervenes directly on the latent trajectory — rather than filtering or blurring completed images after the fact. This framing links safety and alignment to inference-time control under uncertainty.

• Audits unsafe image content at the latent level during denoising instead of post-hoc correction.
• Pairs a learned safety signal with a policy-guided intervention scheme to steer generation toward safer, more aligned outputs.
• Frames alignment as structured control under adversarial and distribution-shift conditions, with a clear connection to robust perception and trustworthy generation.

An argument against deterministic LLM inference. Pushes back on the recent trend toward bitwise-deterministic decoding, showing that collapsing a model's output distribution to a single canonical completion hides exactly the properties — uncertainty, emergent capability, multi-path reasoning, and honest safety behavior — that make a language model worth studying in the first place.

• Separates three distinct stability goals for LLM inference — bitwise determinism, distributional reproducibility, and semantic stability — and argues the field has been optimizing for the wrong one.
• Robustness benchmark across MNLI/QQP/QNLI/SST-2 comparing deterministic vs. majority-vote stochastic decoding under paraphrase, perturbation, and adversarial rewrites.
• Shows phase-transition-like emergent abilities that are visible under stochastic evaluation effectively disappear when measured only with greedy decoding.
• "Deterministic Safety Illusion" experiment: worst-case aggregation over k stochastic samples reveals unsafe completions that a single deterministic run never surfaces.
• Reasoning-graph benchmark (gsm8k) showing self-consistency and tree-of-thought degrade sharply once forced onto a deterministic backbone.