87

Papers

10/10

Roadblocks Active

5

Connections

SCAIR introduces a framework where an LLM agent extracts entities and relations for knowledge graphs while being guided by a formal schema at each step, then self-corrects by looping back through validation checks. Grounding the agent in an explicit schema prevents it from inventing facts that violate the graph's structure, which is the core mechanism behind its hallucination reduction. For AI, this matters because enterprise knowledge graph construction is a high-value task where unchecked hallucination renders outputs unusable, and schema-conditioned iteration offers a principled, auditable path to reliability.

██████████ 0.9 reasoning-reliability Peer-reviewed

Read

This paper presents Terra AI, a system where an LLM orchestrates remote sensing tools and ML models (algal bloom classifier, peat moisture estimator) exposed through standardized Model Context Protocol interfaces. The key finding is that encoding explicit workflow rules directly in the MCP server prompts — rather than leaving the LLM to infer them — drives tool-ordering F1 from 0.79 to 0.99. This matters because it empirically isolates prompt engineering as the dominant reliability lever in multi-tool agentic pipelines, not model capability, which has immediate practical implications for anyone building LLM-orchestrated workflows. Note: evaluation covers only 20 test cases with no significance testing, so effect sizes should be treated cautiously.

██████████ 0.8 agent-tool-use Peer-reviewed

Read

SPARKLE trains a lightweight proxy model via reinforcement learning to decide when and how to retrieve information, treating the underlying LLM and retriever as fixed environment components rather than fine-tuning them. By using knowledge graph reasoning chains extracted from the LLM's intermediate thoughts to structure retrieval decisions, it achieves 9.17% average improvement on in-domain QA and 2.85% on out-of-domain QA over adaptive RAG baselines. The plug-and-play design means the retrieval policy can be swapped onto different LLM–retriever combinations without retraining everything, which addresses a real deployment bottleneck; code is publicly available on GitHub.

██████████ 0.8 hallucination-grounding Peer-reviewed

Read

This work uses vision-language models (VLMs) to detect AI-generated images and deepfake videos by combining visual retrieval with LLM-based reasoning, enabling the detector to explain its decisions comparatively rather than returning a black-box score. For video, temporal aggregation plus perceptual quality analysis improves robustness under realistic degradation like compression and resizing. The significance is that as generative models proliferate, detectors must generalize to unseen generators — and grounding detection in semantic descriptions rather than low-level artifacts appears to improve that generalization.

██████████ 0.8 multimodal-understanding Peer-reviewed

Read

This paper applies quantization to dense retrieval systems for long documents while exploiting document structure (sections, headings, hierarchy) to preserve retrieval accuracy that naive quantization would degrade. The mechanism is that structural signals guide which vector representations to compress aggressively versus preserve, cutting computational overhead while maintaining answer quality on QA benchmarks. For AI, this directly addresses the engineering reality that long-document RAG systems are expensive to run at scale, and structure-aware compression offers a principled way to reduce that cost without simply accepting accuracy loss.

██████████ 0.8 long-context Peer-reviewed

Read

This paper argues that ML adoption is stalled partly by opacity, and proposes combining participatory system dynamics modeling — where domain experts collaboratively build explicit causal diagrams — with ML to fill in empirical gaps those diagrams cannot capture. The mechanism is trust decomposition: humans verify the causal structure (interpretable), then check that ML respects those constraints, applied here to elderly care workforce management. For AI interpretability research, this is relevant because it demonstrates a practical hybrid architecture where ML's learned components are bounded by human-verifiable causal structure, making oversight feasible in complex real-world domains.

██████████ 0.8 interpretability Peer-reviewed

Read

NeuroLangSeg segments brain structures in 3D MRI scans by combining a visual encoder with a biomedical text encoder, guided by anatomical protocol descriptions encoded via an LLM acting as a training discriminator. This language-grounding mechanism enforces that segmentations respect morphological and topological constraints from clinical anatomy guidelines, yielding +4.1 Dice score in same-site settings and +14.5 Normalized Surface Distance improvement in cross-site generalization. The cross-site gains are the most practically significant result, since medical AI systems routinely fail when applied at hospitals using different scanner protocols than the training data.

██████████ 0.7 multimodal-understanding Peer-reviewed

Read

This paper applies a small-data massive-training artificial neural network (SMTANN) to detect scaphoid fractures — a commonly missed wrist injury — in plain X-rays under constrained training data conditions. The approach is relevant because scaphoid fracture datasets are inherently small due to the injury's rarity, making this a test case for whether neural networks can reach clinically useful performance without large labeled datasets. The result matters for data-scarce medical imaging contexts broadly, though detailed quantitative results and dataset sizes are not fully visible from available metadata.

██████████ 0.7 data-quality-curation Peer-reviewed

Read

This work develops a method for automatically synthesizing input preconditions for deep neural networks — formal statements about what inputs a network can handle reliably — backed by statistical guarantees rather than exhaustive formal verification. The practical implication is that a deployed model can be equipped with a gatekeeper: inputs failing the precondition are flagged before inference, preventing silent failures in safety-critical settings. A software artifact (StatPre.zip) is publicly available on Zenodo under Apache License 2.0, though the paper body was not directly accessible for deeper methodological assessment.

██████████ 0.7 alignment-safety Peer-reviewed

Read

This paper builds a Gamma Convolutional Neural Network (GCN) that encodes hydrological prior knowledge — specifically the Gamma distribution's shape and scale parameters, which represent physically meaningful rainfall-runoff time lags — directly into the CNN's kernel design. At just 5% of available training data, the GCN achieves a Nash-Sutcliffe Efficiency of 0.86 versus 0.67 for a traditional hydrological model and 0.73 for a standard MLP. For AI interpretability, the key contribution is demonstrating that injecting domain-meaningful mathematical structure into neural architectures both improves data efficiency and produces parameters that domain experts can audit against physical expectations.

██████████ 0.7 interpretability Peer-reviewed

Read