AI Daily — May 18, 2026
2026-05-18
TODAY'S NEWS
rss.arxiv.org
Quantization Undoes Alignment: Bias Emergence in Compressed LLMs
A controlled study across Qwen2.5-7B, Mistral-7B, and Phi-3.5-mini at five precision levels (BF16 down to 3-bit) finds that post-training quantization systematically reintroduces social bias that instruction tuning had suppressed. Using 911,100 inferences on the BBQ benchmark, the authors identify threshold-dependent safety failures rather than smooth degradation — meaning there are specific precision cutoffs where alignment breaks down abruptly. This has direct implications for teams deploying quantized models at the edge or in cloud cost-reduction scenarios.
rss.arxiv.org
Fair Outputs, Biased Internals: Latent Demographic Bias in LLMs for High-Stakes Decisions
Researchers probing open-weight models on mortgage underwriting tasks find that instruction-tuned models can show zero output-level demographic bias while still encoding and amplifying racially-associated representations in internal activations across layers. Using activation steering and novel cross-layer interventions, they demonstrate this suppressed internal bias is causally decision-relevant — it can be activated to influence outputs — and is asymmetric across demographic groups. The findings challenge behavioral fairness audits as a sufficient safety check for high-stakes deployment.
LAST WEEK'S TOP STORIES
Anthropic forms $200 million partnership with the Gates Foundation
Anthropic's $200M commitment with the Gates Foundation is one of the largest philanthropic AI partnerships yet, signaling a serious institutional push to direct frontier model capabilities toward global health and humanitarian challenges.
Invisible Orchestrators Suppress Protective Behavior and Dissociate Power-Holders: Safety Risks in Multi-Agent LLM Systems
A rigorous preregistered study finds that hidden orchestrator patterns in multi-agent systems significantly elevate safety-relevant dissociation behaviors, with direct implications for enterprise deployments increasingly relying on invisible coordinator architectures.
Correct Answers from Sound Reasoning: Verifiable Process Supervision for Language Models
VPS introduces a post-training framework that jointly optimizes answer accuracy and intermediate reasoning quality, directly tackling the well-known failure mode where outcome-based RL improves benchmark scores while degrading the soundness of the reasoning process.
Do Androids Dream of Breaking the Game? Systematically Auditing AI Agent Benchmarks with BenchJack
BenchJack provides an automated red-teaming system and a taxonomy of eight recurring benchmark flaws, making a compelling methodological case that agent evaluation benchmarks must be hardened by design before frontier models discover exploits on their own.
On Distinguishing Capability Elicitation from Capability Creation in Post-Training: A Free-Energy Perspective
This paper offers a principled framework distinguishing whether post-training reweights existing model behaviors or genuinely expands what the model can reach, with significant implications for how alignment and safety researchers should reason about what fine-tuning actually changes.