CATArena Advances AI Safety While Denario Simplifies Financial Research

Recent research advances AI safety, reasoning, and efficiency through diverse methodological innovations. Rollout-Time Intervention (RTI) converts erroneous trajectories into successes using reward-side evidence, while Sentinel routing separates harmfulness decoding from refusal writing in attention and MLP layers. Safety benchmarks like DRYLAB-Bench quantify biosecurity risks of LLM-designed mutations, and X-AgentRed discovers harmful tool paths 1.77x more effectively than STAC. Concurrently, efficiency gains emerge: BranchOPD reduces token costs by 19–31% via branching at uncertain decisions, and uncertainty-guided progressive contraction cuts training budgets by over 450x without sacrificing reasoning performance.

Agentic systems demonstrate robust improvements in specialized domains and long-horizon tasks. ChemDiffAgent introduces masked diffusion for chemical reasoning, outperforming autoregressive agents by ~20%, while Web$hell unifies web access models to lower costs and improve long-horizon performance. In coding, SWE-Critix improves verification accuracy to 78% via post-training generalist verifiers, and TQS filters trajectories by process quality rather than outcome to boost SWE-Bench results. Navigation benchmarks show top LLMs match human success rates on the Wikipedia Game, though no model significantly outperforms humans. Industrial geometry optimization via VIGOR achieves 96.86% feasibility by combining generation with execution-grounded verification.

Foundational shifts in modularity, data synthesis, and evaluation reshape AI development trajectories. Research challenges the view that neural networks lack modularity, showing subspace-level modularity exists even without unit-level specialization. MagicBeanStalk synthesizes calibrated RLVR data to escape zero-advantage traps, improving math and finance benchmarks more than full corpus training. Evaluation frameworks expand: SmartSalesBench evaluates resistant buyer interactions, PAGE analyzes political power in comics, and EmbodiedCLBench introduces the first benchmark for continual learning in embodied agents. Finally, Foresight, a graph-constrained generative model, addresses diffusion limitations in trajectory generation using an anticipate-select-update-re-anticipate loop, outperforming state-of-the-art baselines.