270

Papers

10/10

Roadblocks Active

0

Connections

Even the best current models succeed on fewer than half of visual tool-calling tasks in a new 511-tool benchmark spanning 16 domains — and the primary reason is not flawed planning but flawed visual perception: 53% of failures occur when models extract wrong information from images despite choosing the right tools. The paper also identifies a planning-to-perception crossover with model scale, where small models fail to decide what to do and large models fail to see what is there. This matters because it shifts the research priority from reasoning improvements toward visual grounding fidelity in agentic pipelines.

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

ReadSaveConnections

StructAgent wraps any backbone LLM in a unified causal state representation — tracking requirements, verified evidence, and open needs — combined with a planner-actor-verifier loop, and nearly doubles success rates on OSWorld-Verified computer-use tasks (27% to 46.9% with a 9B model, 31.6% to 62.2% with a 27B model, reaching 78.9% with MiniMax-M3). The mechanism is that explicit state forces the model to commit to what it has confirmed rather than hallucinating progress. This is a strong signal that scaffolding architecture, not just model scale, drives long-horizon agent performance.

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

ReadSaveConnections

Mako is an agentic system that monitors its own exploitation failures, synthesizes new capability modules in response, verifies them against live targets, and loads them at runtime without restarting — achieving 100% coverage across 104 validation targets spanning 26 vulnerability classes. The authors argue that autonomous offensive capability is now limited by available skills rather than reasoning, and that once a skill exists and is indexable, the difficulty of exploitation collapses. This is a direct empirical demonstration of capability self-accumulation in a security context, with clear alignment implications.

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

ReadSaveConnections

Current video language models answer questions without providing verifiable visual evidence, and this paper demonstrates that QA accuracy and actual visual grounding are decoupled — a model can get the right answer for the wrong reason. The paper introduces E-VQA, a task requiring models to jointly output answers, temporal segments, and dense tracked segmentation masks, backed by a 160k-example training set and a human-annotated benchmark evaluated against GPT-4o, o3, and Gemini-2.5-Pro. Code and data are publicly released, making this a practical resource for measuring and reducing video hallucination.

██████████ 0.9 hallucination-grounding Preprint

ReadSaveConnections

This paper proposes splitting a single autonomous agent into three specialized roles — a Disrupter that explores freely, a Validator that blocks unsafe actions before execution, and a Broker that mediates — combined with a Monte Carlo Tree Search mechanism that compiles failure experiences into signed constraint patches called Scars, preventing the same violation from occurring in future runs. The Validator component is reported to block all executed safety breaches while still permitting creative exploration. The core idea is that safety and exploration need not trade off if the architecture enforces a hard pre-execution veto rather than a post-hoc penalty.

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

ReadSaveConnections

A new benchmark of 787 multi-camera sports video bundles and 2,592 multiple-choice questions across basketball and football shows that current multimodal language models fail to effectively use multiple viewpoints — not because they lack domain knowledge or logical reasoning, but because fine-grained visual perception and view selection are the bottlenecks. An agentic multi-view reasoning system achieves a 14.46% relative improvement over the strongest MLLM baseline by explicitly managing which camera to attend to at each reasoning step. The finding reframes multi-view video understanding as an attention and perception problem rather than a knowledge problem.

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

ReadSaveConnections

The paper argues that test-time compute has three distinct axes — reasoning depth, sampling breadth, and interactive grounding — and that the third axis is governed by a single variable: whether the model receives external feedback from the environment. On hard coding tasks, reasoning-only and best-of-N approaches plateau at fixed token budgets, while a proposer-reviewer interaction loop continues improving, reaching 100% pass rate with no run-to-run variance. This matters because it suggests that the ceiling on current test-time scaling comes not from compute but from the absence of grounded feedback loops.

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

ReadSaveConnections

This paper identifies that standard action tokenization objectives — which train models to reconstruct actions pixel-by-pixel — produce representations biased toward low-level signal fidelity rather than control performance, and that latent space geometry is what actually governs action generation quality. Lumo-2 addresses this through a multi-stage modality pre-alignment strategy that progressively aligns action representations with latent world dynamics, then vision, then language. The framing redefines embodied AI training as a representation alignment problem rather than a data-scale problem.

██████████ 0.8 embodied-ai Preprint

ReadSaveConnections

This position paper argues that the embodied AI field is fragmented along three coupled gaps: incompatible action representations, misaligned training objectives, and incoherent system composition — and that World Action Models, which explicitly connect candidate interventions with predicted consequences, are the unifying architecture needed to close them. While the paper contains no experiments, it provides a useful taxonomy for understanding why progress in robot learning has not transferred cleanly across labs or hardware platforms. It is most valuable as a framework for evaluating whether individual research contributions address root causes or symptoms.

██████████ 0.8 embodied-ai Preprint

ReadSaveConnections

AutoVSR converts circuit schematic images into symbolic mathematical expressions via a two-stage pipeline: a vision-language model first builds an Executable Intermediate Representation of the circuit using dual-ended verification feedback, then a planning agent calls symbolic math tools (SymPy, Lcapy) to derive the final expression. This achieves 30–59% accuracy improvement over end-to-end VLM approaches and 42–52% over specialized prior methods, while also outperforming closed-source frontier models on inference cost. The result demonstrates that decomposing visual-to-symbolic tasks into a perception stage and a structured reasoning stage is more effective than asking a single model to do both.

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

ReadSaveConnections