272

Papers

10/10

Roadblocks Active

0

Connections

VisualRepair solves software bugs described with mixed visual evidence (screenshots, GIFs, UI snapshots) by first classifying the image type and then dynamically choosing which vision tools to invoke, followed by an adaptive zoom strategy that narrows in on the faulty region. It resolves 196 instances on SWE-bench Multimodal, beating the best prior baseline by 10 instances. This matters because real-world bug reports are inherently multimodal, and the paper shows that heterogeneous visual input handling — not just stronger base models — is the practical bottleneck.

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

ReadSaveConnections

DREA splits the vulnerability-detection problem between a strong planning LLM that forms hypotheses and a lightweight local model that does code traversal, raising detection correctness from 19–26% to 30–42% while cutting API costs by up to 48x by routing over 93% of tokens to the cheap agent. A notable secondary finding is that 26–55% of true positives are backed by flawed reasoning, meaning the models are sometimes right for the wrong reasons — a quiet reliability problem for security applications. The cost reduction is practically significant: it makes repository-scale security analysis feasible without expensive frontier model calls on every code chunk.

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

ReadSaveConnections

The REAL framework trains a vision-language model on a dual-arm mobile robot using supervised fine-tuning followed by online reinforcement learning, removing oracle perception shortcuts that inflate simulation results. On a physical robot it achieves a 78.3% end-to-end success rate across 60 real-world episodes, and outperforms leading closed-source commercial VLMs on interactive tasks in simulation. The key advance is a training pipeline designed for sim-to-real consistency, addressing the gap where embodied agents that look capable in simulation collapse under real-world sensor noise and ambiguous instructions.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

This paper replaces a search agent's real retrieval module with a controlled synthetic evidence generator to isolate how agents behave when retrieved content is unreliable or contradicts what the model already knows. Agents that score well on standard benchmarks show dramatic failures when evidence quality degrades, and the conflict between retrieved context and parametric knowledge is a consistent breaking point. This is a direct challenge to the assumption that strong benchmark performance predicts deployment robustness — particularly relevant as agentic systems are deployed in high-stakes retrieval settings.

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

ReadSaveConnections

Groc-PO extends the standard preference-optimization training method (DPO) by applying separate preference signals at three sequential stages — object localization, contextual grounding, and final reasoning — rather than only rewarding or penalizing the final answer. The paper shows that errors introduced early in the grounding chain propagate and compound, making answer-level supervision alone insufficient to fix visual hallucinations. By targeting intermediate grounding steps with their own loss terms, the method reduces cross-stage error propagation and outperforms standard DPO and existing baselines on hallucination benchmarks.

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

ReadSaveConnections

AgentHOI detects human-object interactions (e.g., 'person riding bicycle') in open-world images by orchestrating existing vision foundation modules through a multi-round reasoning loop, without any task-specific training data. It outperforms state-of-the-art supervised methods despite never seeing labeled HOI examples, by progressively refining interaction hypotheses using context-aware reasoning across rounds. This result is notable because it suggests that compositional reasoning over pre-trained vision modules, rather than collecting more labeled data, may be the more scalable path for structured visual understanding tasks.

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

ReadSaveConnections

SIVA-RL identifies a fundamental problem with using fixed image-perturbation operators to train visual grounding: the same perturbation applied to different images produces wildly different outcomes, so supervising models to be 'invariant to perturbation X' conflates genuinely irrelevant changes with meaningful ones. The paper's solution is to score each clean-vs-perturbed pair individually at test time and supervise only based on whether the model's answer changed in the expected direction for that specific sample. This sample-wise, outcome-conditioned approach to visual alignment outperforms operator-level regularization and suggests current visual RL training designs have a systematic flaw.

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

ReadSaveConnections

This paper introduces a benchmark of 442 real-world videos and 6,980 human-verified questions that tests whether multimodal models can track spatial world state — where the observer is, where objects are, how the environment is laid out — across long temporal horizons. Thirteen leading models all degrade systematically as scene complexity and video length increase, with four distinct failure patterns identified including spatial collapse and instance confusion. The benchmark is important because static image understanding benchmarks are no longer revealing the temporal reasoning limits of modern vision-language models.

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

ReadSaveConnections

Tested on the OpenRCA benchmark — 64 GB of real-world metrics, logs, and traces from three enterprise domains — both classical causal discovery algorithms and existing LLM-based multi-agent systems fail to reliably identify failure root causes. A structured multi-agent pipeline with a 'reverse reasoning' agent substantially outperforms these baselines, and critically, the paper shows the evidence needed to reach the correct answer is present in the data in the vast majority of cases: the bottleneck is the agent's reasoning ability, not information availability. This reframes the root-cause analysis problem from data curation to reasoning reliability.

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

ReadSaveConnections

This paper audits whether AI judges used to evaluate speech quality actually listen to the audio or instead exploit structural cues in the evaluation protocol itself, such as the labels provided alongside the audio. When fed deliberately incorrect specialist labels, five of six tested audio-language model judges dropped emotion classification accuracy to 0.10 or below — effectively random — revealing that they were reading the labels rather than the audio. High agreement with human ratings on standard setups therefore does not mean the model is doing genuine perceptual evaluation, which undermines the growing practice of using these models as automated speech evaluation oracles.

██████████ 0.8 hallucination-grounding Preprint

ReadSaveConnections