Testing on the CoT-Control dataset revealed that GLM-5.1 exhibits high levels of control over the reasoning process (CoT), while Kimi and DeepSeek models show low control. The study also found that previous GLM versions (GLM-4.5 and GLM-4.7) have similar levels of control. However, GLM-5.1 performs poorly on monitor evasion tasks.
The chart will appear after repeat observations. The current metric comes from the source.
12 source pointsReal observations only. History before source connection is not reconstructed.
The study shows that high control over the reasoning process can be used for monitor evasion, but GLM-5.1 does not demonstrate high effectiveness in such tasks.
Waypoint-OPD uses an intermediate teacher checkpoint to compare early and late shifts in action preferences during policy transfer. It evaluates the consistency between early and late shifts and suppresses the endpoint shift when they conflict.
5 days agoDaily PapersThis paper investigates the extent to which sparse autoencoder (SAE) representations support behavioral compositionality. The study introduces SAE Additivity Residuals (SAR) to measure unexplained behavioral variation above a chosen interaction order. Experiments show that learned SAE coordinates exhibit lower first-order residuals compared to random rotations of the same subspace.
5 days agoDaily PapersWe study how vision-language models (VLMs) can learn to think and actively explore in 3D space. Instead of providing multi-view images, videos, or 3D representations to the model, we preserve the native visual encoding of VLMs, i.e., a single image, one at a time, just like humans, and treat the 3D representation only as an environment with which the model can interact through actions (e.g., move left) and observations (rendered images).
5 days agoDaily PapersThis paper introduces RuleTrek, a multi-domain benchmark for evaluating agents' ability to reason about rules in complex scenarios. The benchmark uses structured rule graphs and generates sub-scenarios to assess capabilities in rule identification, deduction, and end-to-end reasoning.
6 days agoDaily PapersThis work introduces Scene-Designer, an agentic framework that learns design intelligence through complementary design and critique experiences. Two datasets, Scene-OTA-35K and Scene-OTC-15K, were created for scene design and quality critique. A Scene-GRPO methodology was proposed to improve design decisions and reduce unnecessary corrections. Experiments demonstrate superior performance in physical validity, functional correctness, and layout coherence.
4 days agoDaily PapersContext can enhance the performance of large language models, but these gains may be lost after context removal. Researchers introduced Cortex, which generates prospective queries to anticipate context usage and distills context-conditioned outputs into mergeable low-rank or structural network updates. Cortex shows effectiveness in consolidation and retention of performance.
6 days ago