Modern transformers used in large language models (LLMs) store a vast amount of information in their training data within a 'latent space,' enabling them to recall details and related contexts more accurately. However, this capability is not directly measured in evaluations.
The chart will appear after repeat observations. The current metric comes from the source.
17 source pointsReal observations only. History before source connection is not reconstructed.
The study presents AgentSWE, a benchmark evaluating coding agents' ability to build agents aligned with user requirements. The research found that existing agents fail to accurately fulfill tasks, as there is a significant gap between demonstrated and actual performance.
5 days agoDaily PapersThe 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.
4 days agoDaily PapersThe paper introduces TerminalHorizon, an automated data synthesis engine for long-horizon terminal agents. It constructs 1.5K task environments and fine-tunes Qwen3.5-27B to create TerminalHorizon-27B, achieving performance gains across multiple benchmarks. The model demonstrates improved ability to experiment, refine methods, and retain effective improvements in final results.
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.
6 days agoDaily PapersMulti-Stream Attention Residuals (MSAR) extends the AttnRes method by generalizing the attention mechanism to multi-head depth, achieving significant performance improvements with minimal additional overhead. MSAR leverages multi-scale causal convolutions to draw stream diversity from sequence context, allowing different heads to retrieve historical representations through distinct contextual views.
5 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.
6 days ago