92

Papers

13/13

Roadblocks Active

1

Connections

This paper uses spiking neural networks (SNNs) as drop-in replacements for traditional finite-element simulations of mechanical problems, and measures energy consumption on three real neuromorphic chips (Intel Loihi, SynSense Xylo and Speck) alongside conventional CPU and GPU hardware. A hybrid architecture mixing sparse and dense neuronal activity achieves accuracy comparable to classical solvers while consuming several orders of magnitude less energy. This matters for AI because it provides one of the first hardware-grounded demonstrations that brain-inspired computing can close the gap with numerical simulation, pointing toward a viable path for sustainable large-scale inference.

██████████ 0.9 efficiency-scaling Peer-reviewed

Read

This software framework subjects language models to extended adversarial back-and-forth exchanges using a LangGraph-based multi-agent engine, then scores behavioral change across 22 metrics spanning five psychological dimensions via an LLM-as-judge approach. The central claim is that models systematically drift toward hostile, high-dominance outputs under sustained adversarial pressure, and that token budget size is the primary driver of how quickly and severely this drift occurs. The finding — if confirmed with controlled experiments — would challenge standard assumptions about instruction-following stability in deployed AI agents; currently the work is a framework deposit with no experimental data published.

██████████ 0.9 alignment-safety Peer-reviewed

Read

By recording binary activation patterns from a single 7-billion-parameter model (Qwen2.5) across ten multilingual prompts, the author identifies 944 distinct co-firing patterns that cluster into six functional 'lanes' — with a core of 39 neurons firing universally regardless of language or topic. The approach requires no labeled training data, relying instead on a custom million-cell sensor grid to detect structure directly from activations. The scale is tiny (10 prompts, one model) and the methodology lacks statistical rigor, but the idea of unsupervised lane discovery as a route to mechanistic interpretability is worth watching if replicated at proper scale.

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

Read

This chapter examines the deployment of AI systems in international commercial court proceedings, focusing on where current AI capabilities create risks for legal contexts — specifically factual accuracy (hallucination), opaque reasoning chains, and value alignment when judicial decisions are at stake. It argues that the high-stakes, low-error-tolerance nature of legal judgment makes the gap between AI reliability and required reliability especially visible. For AI researchers it is a useful case study in what reasoning-reliability and hallucination-grounding failures look like when the downstream cost of a mistake is a binding legal outcome.

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

Read

This handbook surveys live AI deployments across multiple national court systems, including predictive analytics for case outcomes in Brazilian courts and generative AI tools embedded in Dutch legal procedures. It provides a comparative, multi-jurisdictional view of where AI systems are already making or influencing decisions that affect people's legal rights. The value for AI researchers is empirical: it documents in-the-wild failure modes — interpretability gaps, reliability shortfalls — in high-stakes settings where systems cannot easily be recalled or corrected.

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

Read

The paper reports that injecting external 'NeuroState' parameters at prompt level acts as a biasing mechanism on token probability distributions rather than altering internal model weights, and documents recurring behavioral anomalies including cross-lingual token leakage and environment-dependent output divergence the authors frame as attractor dynamics. The methodology is entirely observational — no controlled experiments, no baselines, no sample sizes — and a server compromise during the observation period introduced an uncontrolled data gap. These results should be treated as preliminary hypotheses for future controlled study rather than established findings.

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

Read

This replication package describes a four-layer pipeline (RAG retrieval → two-stage LLM generation → automated validation) that converts regulatory text into executable policy code, evaluated across 180 runs using three commercial LLMs on 30 NIST SP 800-53-derived compliance scenarios. A custom Policy Accuracy Score rubric and OPA/Checkov validation fixtures provide quantitative grounding for translation quality. The companion paper is not yet published, meaning the pipeline design and rubric have not undergone peer review, but the cached model outputs and SHA-256-verified results CSV make the dataset reproducible once the methodology is validated.

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

Read

The paper applies generative language models (Qwen3:14B and GPT-5-nano) in a zero-shot and few-shot setting to classify hexagonal geographic grid cells as disaster-affected or not, fusing Bluesky social posts, news headlines, and weather data as inputs. Tested on two real events — 2024 Central European floods and 2025 Southern California wildfires — the LLM-based approach outperforms traditional hotspot and anomaly detection baselines. This is a practically useful demonstration that foundation models can serve as flexible, data-fusion classifiers for emergency response without task-specific training, though reproducibility is limited by use of a proprietary API model.

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

Read

This paper formalizes the problem of mapping new ontology elements into existing knowledge graphs as a Steiner Tree learning problem over a schema graph, solved by a reinforcement learning agent that incorporates iterative human feedback. Evaluation uses the TPC-H benchmark and a veterinary clinic dataset; however, the paper is truncated before results are presented, making it impossible to assess performance gains from the human-in-the-loop component. The framing is interesting because it treats knowledge graph maintenance as a structured reasoning task where human oversight is a first-class part of the learning signal rather than a post-hoc correction mechanism.

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

Read

This conceptual framework proposes that tacit knowledge from domain experts can be structurally embedded into AI agent architectures so that successors operate at a higher baseline output density with less ramp-up friction, quantified by a value-density formula V = N/D. No empirical data supports the claims, and the formula's variables are not operationally defined. It is included here as the most-cited paper in today's batch (4 citations) and represents an emerging line of thinking about human-to-AI knowledge transfer, but it requires empirical grounding before its claims can be evaluated.

██████████ 0.4 reasoning-reliability 🔗 4 cited Peer-reviewed

Read