From General to Granular: Post-Training and Multimodal Scaling Redefine AI Deployment

Post-training research intensified this week: TACT introduces a teaching-adaptive framework that embeds vertical scenarios like English tutoring into post-training, while CoPES designs a co-evolution strategy for resource-constrained agent models. RP-OPSD addresses multilingual reasoning transfer via reasoning-pivot-guided online self-distillation, and failure-driven image self-enhancement offers an alternative self-improvement path for multimodal LLMs. Together, these works show post-training moving from generic reinforcement toward deep customization of domain knowledge, data distribution, and reasoning strategies.

• TACT: Taxonomy-Aligned Post-Training for Pedagogically Adaptive English Tutoring
• CoPES: Cooperative Coevolution for Resource-Constrained Agentic LLM Post-Training
• RP-OPSD: Reasoning-Pivot-Guided On-Policy Self-Distillation for Multilingual Reasoning Transfer
• Failure-Informed Image Self-Augmentation: A New Framework for Multimodal LLM Self-Improvement

Multimodal LLMs face dual challenges in scaling: ParVL proposes a parallel scaling and allocatable computation framework to relieve compute pressure, while Pattern over Pixels measures pattern-completion bias in code generation—models relying on structural patterns rather than pixel-level understanding. The two works address the tension between scale and reliability from compute allocation and behavioral analysis perspectives respectively.

• ParVL: Parallel Scaling and Expandable Compute Allocation for Multimodal LLMs
• Pattern over Pixels: Measuring Pattern Completion Bias in Multimodal Code Generation

AI is accelerating from algorithmic innovation into the intersection of scientific discovery and hardware infrastructure. DASyR-LLM injects LLM knowledge into chemical kinetic model discovery, and OPERA uses operator residual feedback to build reliable autonomous optical experiment agents. A unified analysis of chip design security examines compute risks in the LLM era, while the connection between reinforcement learning and control theory provides a new theoretical anchor for these applications, signaling simultaneous acceleration of fundamental research and applied deployment.

• DASyR-LLM: Domain-Aware Symbolic Regression with LLMs for Kinetic Model Discovery
• OPERA: Operator-Residual Feedback for Reliable Autonomous Optical Experiments with Language-Model Agents
• Hardware Design and Security in the Era of Chiplets and LLMs: A Unified Analysis
• Foundations of Reinforcement Learning and Control: Connections and New Perspectives

Evaluation of autonomous agents and engineering safety formed another main line this week. A study on skill usage examines how well LLMs truly employ skills within agent frameworks, and PredAct-Bench evaluates tool-augmented dialogue under controlled tool noise. Datasette's SQL injection fix in the mixed public/private table scenario echoes the 'release early, iterate' engineering philosophy, reminding us that technology dissemination and security fixes must advance in tandem.

• Skill-Use: Can LLMs Actually Use Skills in Agentic Harnesses?
• PredAct-Bench: Benchmarking Tool-Augmented Dialogue under Controlled Tool Noise
• Datasette 1.0a38 Fixes SQL Injection in Mixed Public/Private Table Setups