A practical reading of TRACE and related vision-language reward research, showing why reliable multimodal agents need stepwise diagnostic feedback before they deserve stronger incentives.
A practical reading of AI political agents: not as robo-politicians, but as a speculative architecture for lowering the cost of democratic participation while raising new risks around power, privacy, and institutional capture.
Agentic Episodic Control shows how language-grounded memory can make reinforcement-learning agents learn faster, generalize better, and spend expensive reasoning only where it matters.
ProgRM shows that GUI agents may learn more from measuring partial progress than from waiting for a final pass/fail signal.
A practical reading of MOSAAIC, a framework for designing co-creative AI systems where control is distributed across autonomy, initiative, and authority.
ModelingAgent shows that real-world AI problem solving improves less from raw tool access than from structured agent roles, shared memory, and critic-driven refinement.
ContextAgent shows that proactive AI assistants are less about speaking first and more about sensing, scoring, and acting only when context makes interruption worthwhile.
A profit-aware DRL framework for injection molding shows why real-time manufacturing optimisation is less about beating genetic algorithms and more about making margin-sensitive decisions fast enough to use.
Pre-Act shows that enterprise agents need explicit planning state, not just one-step tool reasoning, if they are expected to survive real workflows.
A practical reading of why reflection, planning, and heuristic prompts help LLM agents only when the task, model, and feedback loop are aligned.