Cheaper Thinking, Web Traps, Denoised AgentsThree new papers tackle reasoning efficiency, agent vulnerability to web misinformation, and error correction in multi-step AI workflows.