284

Papers

10/10

Roadblocks Active

0

Connections

MemVenom shows that the external memory web agents use to remember past actions can be poisoned with malicious text-image content, causing the agent to follow attacker instructions on future matching tasks with up to 99% success. The attack works in a black-box setting — no access to model weights is needed, only the ability to write one entry into the agent's memory store. This matters because external memory is a standard component of production agents, and the attack transfers across multiple agent frameworks and frontier vision-language models.

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

ReadSaveConnections

This benchmark uses an 8-stage automated evaluation pipeline to test whether vision-language models can solve engineering problems the way an engineer would — reading a technical diagram, selecting the right physical principle, and checking that the answer is physically valid. Models that score well on general reasoning benchmarks consistently fail at interpreting technical diagrams and produce solutions that look plausible but violate basic physics. The stage-wise design pinpoints exactly where reasoning breaks down, which is more actionable than a single aggregate score.

██████████ 0.9 reasoning-reliability Preprint

ReadSaveConnections

PhysTool-Bench tests 13 vision-language models on 2,510 real-world queries involving physical tools from agriculture, manufacturing, and healthcare: the best model (Gemini-3.1-Pro) correctly identifies only 59% of tools in a scene and completes only 21% of tasks end-to-end. The study identifies two separate problems — poor visual recognition of domain-specific tools, and inability to translate what is seen into a sensible action plan — showing these are distinct failure modes, not one. This directly limits real-world robot and agentic deployment in any context outside curated web interfaces.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

AllDayNav trains a robot navigation policy using reinforcement learning that autonomously generates its own training goals — no human labelling needed — and stores experience in a multimodal memory that holds visual keyframes, semantic descriptions, and time context. The system claims near-100% success across room-crossing and task-switching scenarios in both simulation and physical environments, outperforming map-based and VLM baselines on path efficiency. The self-supervised training loop is the key contribution: it allows the agent to keep improving from its own experience without pre-built maps or manual annotation.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

Workflow-GYM is a benchmark that requires GUI agents to complete professional workflows involving at least 40 sequential actions in real software environments — the kind of task a junior office worker would handle. State-of-the-art agents succeed on only about 30% of tasks, failing primarily through objective drift (losing track of what they were supposed to do), skipping required workflow stages, and misreading professional software interfaces. The benchmark fills a gap between short toy tasks, where agents look capable, and the sustained multi-step execution that real professional deployment requires.

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

ReadSaveConnections

HiViG adds a critic model to GUI agents that does two things existing critics do not: it tracks a compact summary of what the agent has already accomplished (macro-action history) to prevent short-sighted replanning, and it visually verifies that the next intended action's screen coordinates actually make sense before executing it. Tested on web, mobile, and desktop benchmarks with both open-weight and commercial policy models, HiViG outperforms verbal and scalar critics. The key insight is that failures in long GUI tasks are often caused by forgetting prior context or clicking the wrong area, not by reasoning errors in isolation.

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

ReadSaveConnections

Using China's standardized national office software certification exam as a benchmark, the study finds that frontier LLMs in single-turn mode score only 37% while humans score 96%; adding agentic execution-feedback loops raises this to 69%, still less than three-quarters of human performance. The test is grounded in verifiable office document outputs — spreadsheets, presentations, documents — not open-ended text, so results are harder to game. The persistent 27-point gap after agentic augmentation suggests the bottleneck is fine-grained document manipulation skill, not just planning.

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

ReadSaveConnections

This systematic review of 247 papers organises LLM agent security research into a unified lifecycle framework covering how attacks enter (prompt injection, tool misuse, memory corruption), how they spread (multi-agent propagation), and what defences exist. The key finding is that current defences address individual attack vectors reasonably well but fail when combined — there is no principled way to compose them into a coherent security posture. For AI practitioners deploying agents in production, this is a practical gap inventory rather than an abstract academic exercise.

██████████ 0.9 alignment-safety Preprint

ReadSaveConnections

Keye-VL-2.0 is a 30-billion-parameter mixture-of-experts model that activates only 3 billion parameters per forward pass, yet achieves top results among similarly-sized models on video understanding, temporal grounding, and STEM benchmarks. A key technical contribution is adapting DeepSeek's sparse attention mechanism to support 256,000-token context windows without accuracy loss in a multimodal setting — something previously not demonstrated in this architecture class. The combination of MoE efficiency and long-context video capability in a single open-weights model is what makes this practically notable.

██████████ 0.9 long-context Preprint

ReadSaveConnections

This paper addresses multi-pair robotic assembly — the problem of fitting many component pairs together without any single pair's assembly interfering with another's — using a framework that splits work between an LLM for local symbolic planning and a lightweight Transformer for resolving cross-pair conflicts. The system reaches 97% executability in offline evaluation and is deployed on a physical UR3 arm, with the LLM constrained to high-level atomic actions to reduce the hallucinated steps that plague unconstrained LLM planners. The result demonstrates that constraining where LLMs operate in a pipeline is often more effective than trying to make LLMs reliable end-to-end.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections