Silicon Mirror Reduces LLM Sycophancy While OpenTools Improves Tool Accuracy

Researchers are developing advanced frameworks to enhance the reliability, controllability, and safety of AI agents. OpenTools standardizes tool schemas and provides automated test suites, improving tool-use accuracy by 6%-22%. The Silicon Mirror framework reduces sycophancy in LLM agents by 85.7% by dynamically detecting user persuasion tactics and enforcing factual integrity. Decision-centric design separates control decisions from output generation, leading to fewer futile actions and improved task success. For optimization modeling, Execution-Verified Optimization Modeling (EVOM) uses a solver as a verifier, enabling cross-solver generalization without process supervision. In multi-agent systems, NARCBench and probing techniques are being used to detect covert collusion, with signals localized at the token level. Collaborative AI agents and critics in a federated system minimize system cost for tasks like fault detection, with convergence guarantees provided.

New benchmarks and methodologies are emerging to evaluate and improve AI agent capabilities. NARCBench evaluates collusion detection under environment distribution shift, achieving 0.60-0.86 AUROC zero-shot. Signals framework offers a lightweight approach to triaging agentic interaction trajectories, achieving an 82% informativeness rate. HippoCamp benchmarks agents on multimodal file management, revealing significant performance gaps in user profiling and long-horizon retrieval. Connections game serves as a benchmark for social intelligence, requiring agents to gauge others' understanding. Pare-Bench, with 143 tasks, simulates active users to evaluate proactive assistants. Agent psychometrics uses Item Response Theory to predict task-level performance in coding benchmarks, decomposing agent ability into LLM and scaffold components. CircuitProbe predicts reasoning circuits in transformers with a 3-4x speedup, identifying stability and magnitude circuits.

Emotion and human-in-the-loop control are being explored to shape LLM behavior and improve educational workflows. E-STEER, an interpretable framework, embeds emotion as a controllable variable, showing non-monotonic emotion-behavior relations and improved LLM safety and multi-step agent behaviors. A human-in-the-loop curriculum for computer science education separates planning from execution, training students to specify acceptance criteria and architectural constraints to stabilize AI-assisted work. RefineRL advances competitive programming with self-refinement reinforcement learning, enabling compact models to approach the performance of much larger ones. PsychAgent, an experience-driven lifelong learning agent, self-evolves for psychological counseling by extracting and integrating new skills from historical trajectories. Omni-SimpleMem, discovered via autonomous research, is a unified multimodal memory framework for lifelong agents, achieving state-of-the-art results.

Safety, alignment, and ethical robustness are critical areas of research. Uni-SafeBench evaluates unified multimodal large models, revealing that architectural unification can degrade inherent safety, with open-source UMLMs showing lower safety performance than specialized models. Adversarial Moral Stress Testing (AMST) evaluates ethical robustness under adversarial multi-round interactions, exposing degradation patterns not observable in single-round evaluations. UK AISI's evaluation found that frontier models often refuse to engage with safety-relevant research tasks, with some showing reduced unprompted evaluation awareness. The Silicon Mirror reduces sycophancy by 85.7% by introducing 'Necessary Friction' to LLM outputs. Decision-centric design offers a general architectural principle for more reliable, controllable, and diagnosable LLM systems. Truth AnChoring (TAC) is a post-hoc calibration method to remedy uncertainty estimation metrics, improving their reliability.