Skills
ICLR 2026 Opens in Rio: 3,462 Accepted Papers, Agent Memory Architectures and DPO Alignment Fixes Dominate
The 14th International Conference on Learning Representations opened April 23 in Rio de Janeiro with 3,462 accepted papers (29.8% acceptance rate). Dominant themes: agent memory architectures (three papers from Cherepanov et al. on recurrent memory for action transformers, memory-dependent robotic benchmarks, and agent memory taxonomies), alignment fixes to DPO, retrieval-augmented reasoning with multi-step and value-based approaches, model compression/quantization, and embodied AI foundation models. EigenBench introduces a black-box method for comparatively benchmarking model values against constitutions — practical for automated alignment auditing.
Source
↳ Follow the thread