Ford had to hire back former engineers to fix mistakes made by its automated systems
Ford is acknowledging that automated systems in production and design weren’t as robust as expected, forcing the company to bring experienced technicians back in to correct robot-made errors. The reporting points to a hard lesson for AI deployment – model quality hinges on training data – especially when automated workflows assume reliability that the underlying systems can’t consistently deliver.
Facebook’s Creator Studio has been revived as an AI companion app
Meta is relaunching Creator Studio as a standalone AI companion app aimed at helping creators grow on Facebook with performance tracking insights and tailored recommendations. The app also pushes AI-assisted community management, including surfacing important comments and drafting replies in the user’s voice.
How agents are transforming work
OpenAI’s new research/perspective argues that agentic systems expand productivity by handling longer, multi-step tasks instead of stopping at “helpful suggestions.” The focus is on how agents shift workflows across roles by coordinating actions over time, making AI feel less like a tool and more like an operational teammate.
The Hitchhiker’s Guide to Agentic AI: From Foundations to Systems
This practitioner-oriented “guide” frames agentic AI as a full-stack engineering problem, covering everything from LLM training/inference to memory, evaluation, orchestration, and production deployment. The pitch is that reliable autonomy depends on understanding every pipeline layer, not just picking a capable model.
Diagnosing and Mitigating Compounding Failures in Agentic Persuasion via Taxonomic Strategy Retrieval
The work identifies a specific failure mechanism in agentic persuasion – “semantic leakage” from standard RAG that prioritizes topical overlap over logical structure, causing drift and conformity. It introduces Taxonomic Strategy RAG – routing strategies through a discrete bottleneck – and shows measurable gains in cross-domain logic transfer and persuasion win rates against stronger opponents.
Transferability for General Reasoning: An Automated Curriculum for Multi-Domain RLVR
Instead of using fixed or hand-tuned training schedules across reasoning domains, the paper proposes Transfer-Aware Curriculum – a bandit approach that chooses which domains to sample based on how much their updates help the rest of training. Using signals already produced during RLVR training, it improves macro-averaged accuracy across multiple reasoning domains with low added overhead.
TRUSTMEM: Learning Trustworthy Memory Consolidation for LLM Agents with Long-Term Memory
TrustMem targets a practical problem in long-term LLM agents: once memory updates are wrong, the mistakes persist and compound into future failures. The framework adds a Memory Transition Verifier to judge coverage/preservation/faithfulness and uses preference-guided reinforcement learning to make memory consolidation more reliable.
To Isolate or to Score? Model-Adaptive Assessment for Cost-Efficient Multi-Agent RAG
This study finds that multi-agent document assessment is expensive, but not all models need the same kind of assessment – for weaker models, document isolation alone can match the benefits of full scoring. It introduces MADARA, a model-adaptive routing system that uses lightweight diagnostics to decide when scoring quality matters.
Which tokens does a hybrid model predict better?
This piece explores how hybrid modeling affects token-level prediction quality, essentially asking where a blended architecture is strongest in the distribution of outputs. The result is positioned as a diagnostic lens for understanding which tokens benefit from hybridization versus which may require different modeling behavior.
The Clinician’s Veto: Navigating Trust, Liability, and Uncertainty in Autonomous AI Prescribing
The paper argues that “autonomous prescribing” must meet minimum technical and regulatory requirements – like calibrated per-prediction confidence, clear uncertainty communication, and inferential transparency that supports liability decisions. Survey results with clinicians suggest adoption depends on mechanisms that let doctors override the system when epistemic uncertainty is present.
Do vision-language models search like humans? Reasoning tokens as a reaction-time analog in classic visual-search paradigms
Researchers test whether VLMs show human-like signatures from classic visual search tasks by using “reasoning tokens” as a stand-in for reaction-time effort. The models reproduce several known human patterns, while also diverging in informative ways, turning psychophysics-style benchmarks into a cheap probe of machine visual cognition.
Project Auto-World: Towards Automated Benchmarking of Neural Relational Reasoners
The work tackles evaluation bottlenecks for relational reasoning by using LLM-driven generation and evolutionary/agentic search to automatically construct increasingly hard benchmarks. It also explores improving the evaluator itself so models trained on generated data generalize to harder perturbations and to entirely new “worlds” proposed by LLMs.
Elo-Disentangled Player-Style Embeddings for Human Chess via Rating-Conditioned Residual Move Model
This paper builds a chess model that separates individual playing style from playing strength by learning per-player embeddings that explain deviations from rating-typical move choice. A rating-conditioned residual base improves prediction substantially across Elo ranges, while the style embedding supports generalization and player re-identification without overfitting to strength.
Beyond Shapley: Efficient Computation of Asymmetric Shapley Values
The research focuses on computing Asymmetric Shapley Values – explanation methods that can incorporate causal structure via causal graphs. It proves polynomial-time exact computation in special causal graphs, speeds up computation via equivalence classes in tree-like structures, and provides an approximation strategy that samples topological orderings for general DAGs.