ARIA: Runtime Inspection and Amendment for Multi-Agent Systems
ARIA targets a nasty multi-agent failure mode – a small early mistake gets baked into later steps and turns a potentially repairable run into a dead end. The framework learns from 24,000+ trajectories by converting execution-verified repairs into supervision, then uses a trained model to decide when and how to amend ongoing runs without repeatedly re-searching alternatives.
Meta removes language that may have allowed it to scrape news sites
Meta updated its commercial terms again after backlash, removing wording interpreted as permission to use content from certain domains and URLs – including news sites – for training AI systems. The Verge frames the change as a response to legal scrutiny and publisher concerns about how AI data collection works in practice.
AlphaDiana: A Measurement System for Agents in Verifiable Reasoning Tasks
AlphaDiana provides a harness-aware way to measure agent performance by separating model correctness from how tools, skills, and memory are orchestrated during execution. Across multiple open-source models and harnesses, the study finds harness changes do not reliably improve accuracy, and it warns that resource usage and output “confidence signals” can mislead without correctness-anchored trace metrics.
Evaluating Model Capabilities After Benchmark Saturation
When benchmarks saturate, accuracy scores stop distinguishing frontier models, but this work argues the transcripts still carry capability signal. Using retrospective saturation experiments, the authors show that extracting information from full solution logs – especially with pairwise judging – can predict held-out capability better than relying on saturated scorelines alone.
Websites are closing their doors to AI agents
Websites are increasingly blocking agent-driven browsing and task execution, forcing AI assistants to run into paywalls, bot restrictions, or “agent-only” licensing gates. The Verge reports on emerging restrictions across major services – a sign that the agent era is colliding with access policies, not just technical reliability.
When Reality Outpaces Cognition: A New Scaling Principle for Reliable Intelligence
This paper proposes a “speed of coherent cognition” benchmark that asks how long an AI can keep meeting reliability criteria as the observation-to-reality gap grows. In fixed-resource experiments, a cellular-inspired Co^4 architecture holds up substantially better than simpler transformer and gated-MLP baselines, pointing toward a measurable scaling principle for reliability in fast-changing worlds.
L2G-SR: Learning to Guide LLM-based Symbolic Regression
L2G-SR tackles symbolic regression failures where numerical error alone steers the search toward the wrong structure. It adds a structural verifier trained with tree edit distance supervision (including commutative/operator-equivalence handling) so search can be guided by learned structural similarity, improving equation-level accuracy alongside numerical fit.
IneqAgent: Agentic Inequality Proving in Lean with Symbolic Computation
IneqAgent combines LLM reasoning with a symbolic computation pipeline to generate and verify exact inequality proofs inside Lean4. Using lemma-level Monte Carlo tree search plus specialized certificate backends, it reports very high success rates on math benchmarks and demonstrates Lean-checked results including strengthenings of known conjectures.
Value-Informed Schedule Trimming for Accelerated Sequential Monte Carlo Guidance in Discrete Diffusion
Instead of paying the heavy cost of full guidance at every denoising step, this work introduces VISTA – a training-free method that selects which timesteps to guide under a fixed budget. By approximating the ideal full-guidance probability path and optimizing an empirical surrogate via dynamic programming, VISTA finds task-dependent schedules that cut latency dramatically while keeping strong output quality.