95

Papers

10/10

Roadblocks Active

1

Connections

This independent preprint argues that transformer layers follow a Fetch-Decode-Execute-Store pipeline analogous to a CPU, and claims one specific attention head (L9H6) suppresses roughly 70% of factual content before output. The proposed 'Surgery' technique intervenes geometrically on hidden states at inference time to recover suppressed facts. While the framing is speculative and lacks controlled baselines, identifying specific suppressor components is a concrete and testable hypothesis that — if replicated — could directly inform hallucination-reduction techniques.

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

Read

This peer-reviewed survey maps the trajectory of text-based visual question answering — systems that must read text inside images and answer questions about it — documenting how datasets and evaluation methods have evolved. The review identifies persistent gaps in grounding language understanding to visual content, which matters because real-world AI deployment (documents, street signs, screenshots) depends on exactly this capability. For researchers, it provides a structured map of where evaluation methodology has improved and where open problems remain.

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

Read

This empirical analysis identifies three distinct pathways through which AI hiring tools inherit and amplify bias: data bias (skewed training history), interaction bias (feedback loops from biased users), and evaluation bias (proxies that correlate with protected characteristics). The core finding is that AI does not neutralize human bias but restructures it — making it faster, more scalable, and harder to contest. This matters because the same training-data-encodes-history problem applies broadly to any AI system trained on human-generated records.

██████████ 0.8 data-quality-curation Peer-reviewed

Read

CNN-based deepfake detectors trained on GAN-generated faces fail to generalize to images produced by diffusion models, with Grad-CAM analysis revealing that the two architectures focus on systematically different facial regions depending on the generator type. This distribution shift problem — detectors learning generator-specific artifacts rather than universal forgery cues — explains why deployed face detectors degrade rapidly as image generation technology changes. The Grad-CAM region-consistency method offers a diagnostic tool for understanding when a deployed detector is operating outside its valid domain.

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

Read

CatBoost models predict near-bit rock properties (sonic velocity, bulk density) from real-time drilling sensor data with 7.94% and 5.74% mean absolute error respectively on blind wells, reducing reliance on costly wireline measurements. Feature importance analysis reveals that gamma ray drives sonic prediction (lithology proxy) while downhole pressure drives density — physically interpretable dependencies that indicate the model has learned genuine domain relationships rather than spurious correlations. This interpretable feature structure is directly relevant to agent tool-use safety: knowing which inputs are causally meaningful lets an autonomous system detect when it is operating outside a tool's valid domain.

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

Read

This conceptual paper introduces the term 'template absorption' to describe cases where a generative AI produces stable, coherent-looking output while silently drifting away from the user's actual structural requirements — the output looks right but the underlying conditions have been replaced by a familiar template. The proposed 'Stability-Substitution Effect' names the tendency to mistake output coherence for output accuracy, which is a real failure mode in deployed AI systems. The paper lacks empirical validation, but naming this mechanism gives practitioners a diagnostic frame for a class of errors that is easy to overlook in production.

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

Read

This governance framework proposes that any AI or biohybrid system operating on biological or cognition-capable substrates must satisfy demonstrable recoverability conditions before deployment — the ability to reverse or contain consequences. The core argument is that technological capability to create a system does not establish the right to deploy it, and that irreversibility creates hard constraints that capability alone cannot override. While it contains no empirical work, it contributes a structured vocabulary for a class of alignment problem — irreversible AI-bio consequences — that existing AI safety frameworks largely ignore.

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

Read

A RAG-based system using open-source reasoning models (Qwen3-30B-Thinking and K2-Thinking) classifies maintenance work orders against the ISO 14224 reliability standard, achieving roughly 70% accuracy on 300 historical records without supervised training, with a target of 95% after further refinement. The most concrete finding is a time reduction from six months of manual effort (two engineers, 10,000 work orders) to under seven days — a meaningful operational result. The gap between 70% and 95% target accuracy highlights the remaining reliability challenge for RAG systems operating on technical standards documents.

██████████ 0.7 hallucination-grounding Peer-reviewed

Read

K-means and hierarchical clustering applied to 80 oil wells independently converged on the same three-cluster structure (silhouette score 0.568), classifying wells into stable, injection-increase candidates, and inefficient waterfloods. Validation against actual field decisions from 2022–2024 showed that five of eight wells predicted for injection increase were part of a successful real-world project, and two wells flagged for losses subsequently received mechanical water shutoff. The convergent agreement between two independent clustering algorithms on field-validated outcomes strengthens confidence that the pattern reflects real operational structure rather than algorithm artifact.

██████████ 0.6 data-quality-curation Peer-reviewed

Read

Three LLMs (GPT-4o mini, Claude 3.5 Haiku, Gemini 1.5 Flash) were compared on their ability to generate 'abuser stories' — threat-actor perspectives derived from user stories — using zero-shot and one-shot prompting on 10 real-world software requirements. The study addresses whether LLMs can systematically surface security threats early in the development lifecycle before implementation, which would reduce the cost of finding vulnerabilities. With only 10 test cases the sample is too small to draw strong conclusions, but the methodology establishes a replicable evaluation structure for a practically relevant application.

██████████ 0.6 reasoning-reliability Peer-reviewed

Read