The Alignment Half-Life Problem

Five independent research papers published today — none citing each other — converge on a single finding: post-training is where model behavior is actually determined, and current methods produce models whose alignment is fundamentally unstable and opaque even to their creators. That result collides directly with Anthropic's launch of a safety-differentiated Claude Mythos tier and Google's multimodal expansion, raising a question neither company is answering: how do you guarantee a safety tier when the research community is proving that alignment established during post-training does not reliably survive adversarial fine-tuning, architectural changes, or extended reasoning?

Thread 1: The Alignment Debt Is Coming Due

• Anthropic Releases Claude Fable 5 and Claude Mythos 5: Same Underlying Model, Different Safeguards, New Mythos-Class Tier — MarkTechPost `[3a21e90463]`
• It Takes One to Bias Them All: Breaking Bad with One-Shot GRPO — arXiv cs.CL `[6a3992fe11]`
• Does Reasoning Preserve Alignment? On the Trustworthiness of Large Reasoning Models — arXiv cs.CL `[409b6da10a]`
• CIAware-Bench: Benchmarking Control Intervention Awareness Across Frontier LLMs — arXiv cs.AI `[b39b1d1076]`
• ABC-Bench: An Agentic Bio-Capabilities Benchmark for Biosecurity — arXiv cs.AI `[dda8668e64]`
• PhantomBench: Benchmarking the Non-existential Threat of Language Models — arXiv cs.AI `[abe23429f9]`

Thread 2: Benchmarks All the Way Down

• T1-Bench: Benchmarking Multi-Scenario Agents in Real-World Domains — arXiv cs.AI `[ce4efb41e4]`
• Workflow-GYM: Towards Long-Horizon Evaluation of Computer-use Agentic tasks in Real-World Professional Fields — arXiv cs.AI `[ab924dc4fe]`
• Mind the Gap: Can Frontier LLMs Pass a Standardized Office Proficiency Exam? — arXiv cs.CL `[f829485320]`
• VISTA: A Versatile Interactive User Simulation Toolkit for Agent Evaluation — arXiv cs.CL `[7244de4ac5]`
• Flaws in the LLM Automation Narrative — arXiv cs.AI `[6c628e4ab0]`
• Do Transformers Actually Help Intrusion Detection? A Temporal Sequence Evaluation on CIC-IDS2017 — arXiv cs.LG `[68703a77d0]`
• What Fits (Into Few Tokens) Doesn't Overfit: Compression and Generalization in ML Research Agents — arXiv cs.AI `[388026afb1]`

Thread 3: The Post-Training Arms Race

• Introducing Gemma 4 12B: a unified, encoder-free multimodal model — Google DeepMind Blog `[9be096effc]`
• Google Releases Gemini 3.5 Live Translate — MarkTechPost `[39896267c8]`
• Fluid, natural voice translation with Gemini 3.5 Live Translate — Google DeepMind Blog `[8065dada88]`
• TRACE: A Unified Rollout Budget Allocation Framework for Efficient Agentic Reinforcement Learning — arXiv cs.AI `[c5e434bbe2]`
• A Unifying Lens on Supervised Fine-Tuning Through Target Distribution Design — arXiv cs.AI `[0f2466a869]`
• Attention Amnesia in Hybrid LLMs: When CoT Fine-Tuning Breaks Long-Range Recall, and How to Fix It — arXiv cs.CL `[766c210451]`

Cross-Story / Counter-Narrative

• It Takes One to Bias Them All: Breaking Bad with One-Shot GRPO — arXiv cs.CL `[6a3992fe11]`
• Does Reasoning Preserve Alignment? On the Trustworthiness of Large Reasoning Models — arXiv cs.CL `[409b6da10a]`
• A Unifying Lens on Supervised Fine-Tuning Through Target Distribution Design — arXiv cs.AI `[0f2466a869]`
• Attention Amnesia in Hybrid LLMs — arXiv cs.CL `[766c210451]`
• CIAware-Bench — arXiv cs.AI `[b39b1d1076]`
• Anthropic Releases Claude Fable 5 and Claude Mythos 5 — MarkTechPost `[3a21e90463]`

Quick Hits

• Introducing North Mini Code: Cohere's First Model For Developers — Hugging Face Blog `[3b631680ec]`
• Introducing Gemma 4 12B: a unified, encoder-free multimodal model — Google DeepMind Blog `[9be096effc]`
• Can Voice Agents Handle Bilingual Customers? Benchmarking Frontier ASR on Code-Switched Speech — Hugging Face Blog `[7a24901e21]`
• ReasonAlloc: Hierarchical Decoding-Time KV Cache Budget Allocation for Reasoning Models — arXiv cs.AI `[769e0bfba3]`
• Piper: A Programmable Distributed Training System — arXiv cs.AI `[138213b411]`
• Predicting Future Behaviors in Reasoning Models Enables Better Steering — arXiv cs.LG `[b371b92495]`
• ABC-Bench: An Agentic Bio-Capabilities Benchmark for Biosecurity — arXiv cs.AI `[dda8668e64]`