Tag
This paper investigates how restricting evidence visibility in multi-module neural systems composed of language models can improve compositional generalization, showing that restricted-visibility societies outperform globally visible ones by over 20 percentage points in a relay communication setup.
This paper introduces a calibrated test of internal action maps in language models, showing that state signals can be decodable and causally usable without global affine closure, using an evidence lattice framework validated on finite worlds and the Qwen3-4B model.
This blog post argues that better generalization in language models should come from the 'harness' — the interface program — rather than just scaling training data. Experiments show that a Recursive Language Model harness enables length and domain generalization far beyond what the base Transformer achieves.
InternVLA-A1.5 integrates pretrained vision-language models with future prediction in latent space to enable efficient robot manipulation with compositional generalization and long-horizon execution, achieving state-of-the-art results on simulation benchmarks.
This paper investigates object-driven shortcuts that hinder compositional generalization in zero-shot compositional action recognition, proposing RCORE to mitigate verb-collapse and improve unseen composition generalization.
This paper theoretically analyzes how curriculum learning, by decomposing complex problems into simpler sub-problems and composing solutions, can dramatically reduce the sample complexity of learning to simulate sequential computations (semiautomata) compared to direct methods, achieving subpolynomial supervision requirements in supervised fine-tuning and exponentially weaker coverage conditions in reinforcement learning with verifiable rewards.
This paper formalizes 'compositional behavioral leakage' (CBL), a failure mode in prompt-composed agentic systems where editing one prompt module silently shifts the behavior of others due to transformer self-attention lacking module-level isolation. It presents an operational definition, a reusable three-channel protocol, and empirical evidence from 144 trials on a Claude Sonnet 4.6 agent, finding sub-threshold interference that could compound across thousands of decisions.
This paper proposes Adversarial Concept Search, a method that uses the representational geometry of large language models to predict compositional failures without evaluating specific inputs. The approach identifies high-risk scenarios by measuring interference between salient features.