Feed

528 materials

Sort by
News1 October 2026

Curvature in Attention Graphs Reveals Context Failures That Lead to LLM Hallucinations

The article proposes a one-step detector of unreliable answers based on the topology of information flows in attention graphs and Forman–Ricci curvature. The authors show that disruptions in context transmission between tokens during autoregressive generation are reflected in the structure of connections and make it possible to distinguish hallucinations across several models and benchmarks.

Curvature in Attention Graphs Reveals Context Failures That Lead to LLM Hallucinations
Read
News1 October 2026

CaLR: Latent Optimization with Causal Constraints for Robust Diffusion Model Inference

The paper introduces CaLR, an approach that casts reasoning as constrained optimization in latent space. Its structure comes from a causal topology matrix derived from an expert model, while training relies on implicit differentiation. The method improves consistency and enables intermediate results to be adjusted during parallel generation, leading to better performance on challenging datasets and tasks with strict constraints, including Sudoku.

CaLR: Latent Optimization with Causal Constraints for Robust Diffusion Model Inference
Read
News29 September 2026

Simulation for Testing Causal Inferences About Advertising Failures

In TRACE, hidden interventions in the digital advertising simulator make it possible to automatically assess whether an agent correctly identified the source of a failure and the affected segment, even though it must analyze noisy data using Python and SQL. After reinforcement learning on synthesized signals, the Qwen3.5-35B-A3B model scored 0.757 on FullAttr@1 across 235 test episodes—higher than all tested prompted variants, while using fewer tool calls than the baseline model.

Simulation for Testing Causal Inferences About Advertising Failures
Read