286

Papers

11/11

Roadblocks Active

0

Connections

EnvFactory builds a pipeline of three AI agents that autonomously discover, verify, and package real software tools into executable training environments — replacing LLM-hallucinated simulators that teach agents wrong tool behavior. Using only 85 verified environments, it generates enough training data to match or beat systems trained on five times more synthetic data. This matters because the dominant bottleneck for tool-using AI agents has been the cost and unreliability of training data, and executable verification directly addresses both.

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

ReadSaveConnections

LongMINT is a benchmark of 15,600 questions over contexts averaging 138,800 tokens, where information is frequently updated and interleaved, forcing agents to track multiple changing facts simultaneously. Seven representative systems — including RAG, vector memory, and long-context LLMs — averaged only 27.9% accuracy, with the sharpest failures on tasks requiring aggregation across multiple evidence fragments. The result quantifies a specific and practically important gap: current memory systems are adequate for simple lookups but collapse under the kind of multi-threaded reasoning real tasks require.

██████████ 0.9 long-context Preprint

ReadSaveConnections

SP-CoR lets multiple robots pool their individual camera views to answer spatial questions that no single robot could answer alone, using physics-informed fusion during training and distilling pose knowledge into prompt tokens so robots need no special sensors at test time. Evaluated on a 114,227-question benchmark across Habitat and iGibson simulators plus real quadruped robots, it improves over the strongest baseline by 3.9% on Habitat and 7.1% on iGibson. The practical implication is a plausible path to multi-robot spatial awareness without requiring expensive shared sensing infrastructure.

██████████ 0.9 multimodal-understanding Preprint

ReadSaveConnections

ESI-Bench tests AI agents on 3,081 spatial reasoning tasks derived from Spelke's core knowledge systems, comparing agents that actively choose what to look at versus those given fixed or random extra views. Active exploration substantially outperformed passive observation, while random multi-view acquisition actually hurt performance by adding visual noise. This quantifies something intuitive but previously unmeasured: purposeful looking is not interchangeable with simply seeing more.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

CrossView Suite introduces a dataset, fine-tuned model, and benchmark specifically for training multimodal language models to reason consistently about the same objects seen from different angles — a capability current models handle poorly. The core finding is that a three-stage training progression (perception, then cross-view alignment, then reasoning) is necessary; skipping any stage degrades results. This provides a concrete curriculum for the underexplored problem of viewpoint-invariant spatial understanding in vision-language models.

██████████ 0.9 multimodal-understanding Preprint

ReadSaveConnections

Key-Gram adds an external memory of language-derived world knowledge to robot manipulation models, retrievable in O(1) time via deterministic hashing, without retraining the underlying model. By separating what language instructions imply from what the robot currently sees, it achieves a 29.5% relative improvement on the π₀ backbone on RoboTwin2.0. The modular design means this knowledge store can be updated or extended for new tasks without touching the core model — important for deploying robots across changing environments.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

Bio-Harness combines LLM-based planning with deterministic template compilers and strict tool-configuration contracts that prevent the LLM from ever hallucinating a tool name or setting a wrong parameter — the agent reasons, but execution is locked to validated templates. Across 144 test cases it recorded zero hallucination events, zero fallbacks, and zero fail-open failures. The architecture demonstrates that separating LLM reasoning from execution enforcement is a viable and auditable solution to tool-use reliability in regulated scientific domains.

██████████ 0.9 hallucination-grounding Peer-reviewed

Read

Robo-Cortex addresses the problem that embodied navigation agents forget useful lessons between episodes by building a two-tier memory: a short-term reflective store for current-task observations, and a long-term library of distilled heuristics (guiding rules and cautionary patterns) extracted from past trajectories by an Autonomous Knowledge Induction mechanism. An adaptive variant updates this heuristic library during inference, not just training. Evaluated on three navigation benchmarks, the system shows that structured knowledge distillation from experience outperforms raw trajectory replay.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

MementoGUI adds a learned memory controller to GUI-operating AI agents that selectively compresses and stores task-relevant interface events — including region-of-interest visual evidence and text summaries — rather than replaying full interaction histories. The system works as a plug-in without retraining the underlying vision-language model backbone, and outperforms both no-history and text-only memory baselines on GUI-Odyssey and MM-Mind2Web. For anyone building AI assistants that operate computers over many steps, this points toward selective memory as preferable to raw context growth.

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

ReadSaveConnections

Using real design preference data from a UX research platform, this study tested multiple GPT configurations — varying reasoning mode, sampling temperature, and persona specificity — against actual human responses and found systematic misalignment that no configuration corrected. LLM-generated justifications for design choices were found to lack the nuance and variability of genuine human reasoning, even when they appeared plausible. The practical implication is that using AI-simulated users as a substitute for human testing in product design carries a measurable risk of steering decisions in the wrong direction.

██████████ 0.8 alignment-safety Preprint

ReadSaveConnections