Continual post-training faces challenges of acquiring new capabilities while retaining prior knowledge. A method called Synthetic Self-distillation is introduced, providing high-quality supervision for small models without disrupting their original distribution. Experiments show the method outperforms six other approaches.
This source has not updated recently. Its observation time is shown above.
The chart will appear after repeat observations. The current metric comes from the source.
17 source votesReal observations only. History before source connection is not reconstructed.
The method can be useful for developing small models requiring low latency and on-device operation.
The paper introduces S-Space, a linear subspace in multimodal language models where object-token representations encode continuous spatial coordinates. It demonstrates that S-Space serves as an internal workspace for spatial reasoning, enabling accurate spatial judgments through perception, internal reasoning, and selective verbal access, while highlighting limitations in spatial transformations across viewpoints.
3 days agoDaily PapersThe paper investigates on-policy distillation (OPD) in cross-tokenizer settings, focusing on alignment coverage and supervision reliability. It finds that strict 1:1 token alignment covers most student-generated tokens despite vocabulary mismatches, and that restricting reverse KL to a top-16 subset of shared vocabulary achieves comparable accuracy to full shared-vocabulary OPD, while adding span supervision reduces accuracy.
6 days agoDaily PapersThe paper introduces Memento 3, a system enabling frozen LLM agents to continuously learn explicit world models through external memory. The agent maintains a natural-language rulebook as persistent semantic memory, which is compiled into executable code for prediction and planning. Through a loop of observation, reflection, and verification, the agent refines its model and uses verified updates to guide interaction. On ARC-AGI-3, the single-model agent achieves 100.0 mean Relative Human Action Efficiency and uses 44% of the human action count. In an Atari Pong case study, a learned feedback controller wins 21:0 in each of three evaluated episodes.
4 days agoDaily PapersCheckerBench is a benchmark for evaluating coding agents' ability to synthesize static-analysis checkers from defect specifications, including tasks derived from 300 CVEs across multiple repositories and language ecosystems. CheckerLab, an evaluation framework, measures diagnostic contrast, patch localization, and tool use, with results showing that current agents struggle to develop reliable checkers, achieving a maximum Pass@1 of 45.33%.
6 days agoDaily PapersTetris3D is a generative framework for 3D scene reconstruction from a single image, ensuring objects are geometrically and physically coherent. It uses explicit conditioning on surrounding objects' geometry and physical relationships, along with a physics-based dataset called ComOb for training and evaluation.
5 days agoDaily PapersThe study explores Adversarial Heuristic Learning (AHL), where AI agents refine game strategies with fixed model weights. The AAArena benchmark includes 12 games and 1,920 archived programs, evaluated through protocols mimicking real competitions. Results show agents can improve strategies but face challenges in complex games.
4 days ago