97

Papers

10/10

Roadblocks Active

2

Connections

Current vision models build scene graphs by detecting every possible object and relationship in an image, producing cluttered outputs that hurt downstream reasoning tasks. This paper shows that training scene graph models with loss functions guided by human-written captions—which implicitly encode what is relevant—produces sparser, more human-like graphs. The key finding is that standard recall metrics miss these improvements, but graph edit distance against human annotations reveals them, suggesting the field has been measuring the wrong thing.

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

Read

When multiple AI agents operate simultaneously in enterprise systems without coordination, they collide on shared state, block each other, and produce outputs that cannot be traced to any single agent. This paper proposes a five-layer orchestration framework—analogized to traffic control—with two architectural models: a centralized 'traffic light' approach for predictability and human oversight, and a decentralized 'roundabout' for throughput and resilience. The work is purely conceptual with no benchmarks, but it names and structures a real operational problem that most multi-agent deployments currently solve ad hoc.

██████████ 0.7 agent-tool-use Peer-reviewed

Read

Deploying accurate object-detection models in real-time edge settings like X-ray scanners requires trading off accuracy against compute, a core AI efficiency challenge. GSA-YOLO addresses this by combining structured sparsity—removing entire weight groups rather than individual values, which maps cleanly to hardware—with adaptive knowledge distillation that transfers capabilities from a large model to the compressed one. The result maintains detection accuracy at substantially lower computational cost, demonstrating a deployment-ready compression pipeline for safety-critical vision tasks.

██████████ 0.7 efficiency-scaling Peer-reviewed

Read

Machine unlearning—removing specific knowledge from a trained model without retraining from scratch—is an emerging requirement for privacy compliance and alignment. This artifact visualizes the 'Who's Harry Potter?' unlearning method as edge surgery on a knowledge graph, making the mechanism intuitive: the model's associations between specific entities are cut while surrounding language structure is preserved. It is an educational tool rather than original research, but it usefully frames approximate unlearning in a way that highlights what current methods can and cannot selectively forget.

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

Read

This paper attempts to characterize transformer internals using thermodynamic and dynamical-systems analogies, reporting a claimed universal constant for attention's 'specific heat' (dU/dT ≈ −18) and a consistently negative Lyapunov exponent suggesting transformers are stable attractors. The framing is creative, but critical caution is warranted: experiments use only two small model sizes (0.5B and 1.5B), the physical analogies are metaphorical rather than derived from first principles, and the work comes from a single unaffiliated researcher with no peer review. Treat as a hypothesis-generating curiosity, not an established result.

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

Read

Most language models scale by adding parameters; this architecture proposes a second axis—iterating a shared transformer block multiple times within a single forward pass, with a learned controller deciding how many refinement steps to take. Internal verification signals and conditional repair updates are woven into this loop, aiming to let the model self-correct mid-generation. The idea is interesting, but the only experiment run was a 15M-parameter smoke test to confirm the code executes; no benchmarks against existing models exist, and no code is publicly released.

██████████ 0.7 reasoning-reliability Peer-reviewed

Read

This PRISMA-compliant systematic review surveys AI applications in forensic personal identification—fingerprints, facial recognition, dental records—and finds a median accuracy of 91.4% across studies, with deep learning modestly outperforming classical machine learning. The review is notable for a high-stakes deployment context where false identifications carry severe consequences, making data quality and model interpretability non-negotiable. It provides a useful calibration point for how mature AI performance actually is in a regulated real-world domain.

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

Read

Agentic AI systems that sense and decide often lack a principled mechanism for improving over time from past episodes. GESA proposes a five-component optimization loop borrowing simulated annealing's temperature schedule—starting with broad exploration of past experiences and gradually focusing on the most relevant ones for synthesis. The framework is purely conceptual, extracted post-hoc from a proprietary production pipeline, and the 11.6 kB software deposit suggests minimal implementation; it is worth tracking if an empirical evaluation surfaces.

██████████ 0.6 agent-tool-use Peer-reviewed

Read

As AI agents are asked to collaborate and share state across organizational boundaries, questions of trust, accountability, and tamper-evident memory become critical. This software deposit proposes using Byzantine fault-tolerant consensus (CometBFT) as an institutional memory layer for multi-agent systems, with a 'Proof of Experience' mechanism that weights agents' votes by their demonstrated reliability track record. The concept connects distributed systems trust guarantees to AI agent coordination, though no evaluation exists and the artifact is a software stub rather than a validated system.

██████████ 0.6 alignment-safety Peer-reviewed

Read

This deposit applies E8 root vector geometry—a 248-dimensional mathematical structure from theoretical physics—to analyze token routing in a Mixture-of-Experts language model architecture, reporting a compression ratio of 0.772 for cluster 145 of 240 total root vectors. The connection between E8 geometry and MoE routing efficiency is not derived from first principles and lacks experimental validation. It is an unusual framing that may be worth watching if the geometric compression claims are tested empirically, but currently reads as speculative pattern-matching.

██████████ 0.5 efficiency-scaling Peer-reviewed

Read