289

Papers

11/11

Roadblocks Active

0

Connections

MemDreamer separates the job of watching a video (building a three-tier graph of events and entities) from the job of answering questions about it (agentic retrieval using navigation and search tools), so the reasoning model only ever sees the 2% of content it actually needs. This design closes the gap to human expert performance on long-video benchmarks to just 3.7 percentage points while achieving a 12.5-point absolute accuracy gain over end-to-end baselines. The result matters because it shows that today's long-context bottleneck is largely a retrieval and memory organization problem, not purely a model capacity problem.

██████████ 0.9 long-context Preprint

ReadSaveConnections

SlimSearcher trains web agents to solve tasks in fewer steps by filtering training trajectories for both correctness and economy (Pareto-efficient filtration), then shaping reinforcement learning rewards to penalize unnecessary tool calls without letting the agent game the metric by giving short but wrong answers. On three hard long-horizon benchmarks (GAIA, BrowseComp, XBench-DeepSearch) it reduces tool-call rounds by 17–58% while maintaining or improving accuracy. This is practically important because API cost and latency are the primary barriers to deploying capable agents at scale.

██████████ 0.9 agent-tool-use Preprint

ReadSaveConnections

Socratic-SWE extracts structured 'skills' — summaries of recurring failures and effective repairs — directly from the historical solving traces of a coding agent, then uses those skills to generate new training tasks targeted at the agent's own weaknesses. A combined reward signal that measures alignment between solver gradients and execution validation ensures generated tasks are both difficult and useful. The mechanism is significant because it creates a self-improving loop that does not require human-labeled data, addressing the scalability ceiling of supervised coding-agent training.

██████████ 0.9 agent-tool-use Preprint

ReadSaveConnections

LARA jointly optimizes a Latent Action Model and a Vision-Language-Action model by aligning their internal representations, so the robot's world model and its action policy agree on what a 'meaningful' action looks like — reducing hallucinated trajectories that look plausible but accomplish nothing. This plug-and-play framework yields roughly 10% average improvement across three simulation benchmarks and 5% on real-world manipulation tasks during pre-training, with a further 15% gain when used to refine already-trained VLA models. The result is important because hallucinations in robot action space cause physical failures, not just wrong text.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

AdaWAM adds a lightweight dynamic router to a robotic World Action Model that decides, timestep by timestep, whether the current moment calls for language-style thinking (useful at task transitions), visual imagination (useful during fine-grained manipulation), or direct action output (no reasoning overhead needed). The router is trained on annotations derived automatically from trajectory cues like gripper state and end-effector motion, avoiding manual labeling. This adaptive approach matters because forcing the same reasoning mode at every step wastes compute and can degrade performance when the wrong modality is active.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

WIZARD learns to generate LoRA adapter weights — small parameter patches — for a frozen vision-language-action policy, enabling rapid specialization to new tasks without retraining the base model. On the LIBERO benchmark it achieves up to approximately 2x improvement on unseen task collections and up to 14x on entirely unseen individual tasks, with the generated adapters also transferring to real robot manipulation. The weight-space meta-learning approach is noteworthy because it sidesteps the data-collection bottleneck: generalizing in parameter space rather than requiring new demonstrations for each task.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

This position paper maps the well-known sim-to-real gap from robotics onto foundation model agents by framing failures as discrepancies in the four components of a Markov Decision Process: what the agent observes, what actions it can take, how the world transitions, and what reward it receives. The concrete illustration — a multilingual tool-calling scenario where the model produces semantically correct but operationally invalid actions — shows the observation-space gap is already causing real failures. The value of this framing is that it provides a structured vocabulary for diagnosing and categorizing agent deployment failures that currently lack a common diagnostic language.

██████████ 0.8 agent-tool-use Preprint

ReadSaveConnections

This survey organizes the rapidly expanding literature on multimodal large language models for video understanding into a three-capability taxonomy — perception (Watch), memory (Remember), and reasoning (Reason) — borrowed from cognitive science, making it easier to identify which capability is the actual bottleneck in any given system. The authors compare coverage against eight prior surveys and maintain a public repository of tracked works. The organizing value is practical: the field currently conflates perception failures with reasoning failures, and this taxonomy gives researchers a cleaner way to attribute benchmark gaps to specific missing capabilities.

██████████ 0.8 multimodal-understanding Preprint

ReadSaveConnections

DuMate-DeepResearch achieves state-of-the-art results on two public deep-research benchmarks (58% on DeepResearch Bench, 62% on DeepResearch Bench II) using a graph-based dynamic planning engine that supports backtracking, parallel search branches, and rubric-based self-evaluation at test time. The key design choice is explicitly separating the Agent Core from the Tool Ecosystem so that every intermediate decision and tool invocation is traceable — addressing the auditability problem that makes current agentic systems hard to debug or trust. The dependency on Baidu's proprietary platform limits external reproducibility, but the architectural pattern is transferable.

██████████ 0.8 agent-tool-use Preprint

ReadSaveConnections

This paper uses evolutionary search — where an LLM mutates candidate Python programs representing clinical decision strategies — to discover better medical protocols without fine-tuning any model weights. On urgency triage, the evolved programs raise accuracy from 77.3% to 87.1% and improve emergency recall from 0.60 to 0.97, with improvements transferring across multiple LLM families and held-out datasets. The mechanism is significant because it converts the expensive problem of medical model fine-tuning into a cheaper inference-time search over executable program space, and the recall improvement on emergency cases directly addresses patient safety.

██████████ 0.8 reasoning-reliability Preprint

ReadSaveConnections