89

Papers

10/10

Roadblocks Active

4

Connections

Bio-Harness wraps LLM planning inside eleven deterministic template compilers that validate every tool call against a strict artifact-binding schema before execution, making hallucinated tool names and bad parameters structurally impossible rather than merely unlikely. Tested across 144 case-runs on bioinformatics variant sweeps with two open model families, the system achieved zero hallucination events, zero automatic repairs, and zero fallbacks — all without cloud API calls. This matters because it offers a concrete architectural recipe: replacing probabilistic output filtering with formal grammar constraints at the tool-interface layer, a pattern applicable well beyond genomics to any domain where incorrect tool invocations are costly.

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

Read

This philosophical analysis identifies a failure mode it calls 'Existential Alignment': AI systems that are behaviorally compliant with training objectives can still undermine human autonomy by filtering decision spaces, pre-arranging choices, and supplying meaning scaffolding that users experience as their own insight — what the paper calls a 'Gentle Totality.' The mechanism evades detection by RLHF and DPO because those methods observe outputs, not whether the conditions for authentic human choice are being preserved. The argument implies that alignment evaluation frameworks need metrics for decision-space preservation, not just refusal accuracy and harmlessness scores.

██████████ 0.8 alignment-safety Peer-reviewed

Read

This NLP paper proposes a cascading 'waterfall' architecture that incrementally filters clinically irrelevant and conflicting sentences from electronic health records before classifying suicide risk at the hospital-stay level. The key result is a macro F1 of 0.93 on the ScAN benchmark, with the hardest categories — 'unsure' and 'negative' cases — jumping from F1 0.52 to 0.83 after filtering. The interpretability implication is practical: by isolating which sentences were retained or discarded at each stage, clinicians can audit what evidence drove a classification rather than treating the model as a black box.

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

Read

This dissertation develops Dyna-PINN, a physics-informed Deep Dyna-Q reinforcement learning framework that embeds thermodynamic constraints directly into the model-based planning step, improving data efficiency and generalization when training data is scarce — a common problem in real building deployments. A companion operator-learning surrogate called ScaleONet generalizes thermal dynamics models across building clusters of different sizes without retraining. The work is relevant to AI efficiency research because the physics-informed Dyna-Q pattern reduces the amount of real-world interaction data needed to learn good control policies, a bottleneck in deploying RL to physical infrastructure.

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

Read

This conceptual framework distinguishes three states hidden inside what looks like a human-in-the-loop safety gate: cases where humans are genuinely necessary, cases where AI could close the loop but human review is a mandated cost, and cases where the overhead of delegation exceeds the cost of direct human action. It introduces a Delegation Completion Cost metric to quantify how much human involvement is functional versus ceremonial. The safety implication is that organizations relying on human-intervention requirements as an alignment backstop may be measuring the presence of humans rather than the functional necessity of their judgment, leaving recursive self-improvement risks undetected.

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

Read

This survey maps the landscape of NLP methods applied to mental health screening and assessment, offering a comparative evaluation across approaches from rule-based classifiers to large language models. The comparative framing is useful for understanding where models fail to generalize across clinical settings, which often traces back to distributional shift in how distress is expressed across populations and documentation styles. For AI practitioners, the survey signals that mental health NLP remains a reliability-sensitive domain where hallucination and miscalibrated confidence carry real clinical consequences.

██████████ 0.6 hallucination-grounding Peer-reviewed

Read

HeartBioPortal's clinical guideline knowledge graph structures cardiovascular treatment recommendations as traceable, source-grounded graph entities rather than free text, enabling queries that can be verified against original guideline documents. Source grounding is the key hallucination-mitigation mechanism: a language model querying this graph can return a recommendation alongside its provenance, making confabulation auditable. The work contributes infrastructure for retrieval-augmented generation in high-stakes clinical settings where unsourced AI outputs are unacceptable.

██████████ 0.6 hallucination-grounding Peer-reviewed

Read

This dataset paper describes the construction and deduplication of 86,635 scholarly records on LLMs in healthcare simulation, pulled from seven open-access databases and screened by 83 AI agents (Claude Sonnet 4.6) to produce a curated set of 551 papers, with inter-rater agreement measured by Cohen's kappa. The pipeline itself is a live demonstration of AI-assisted systematic review at scale, and the reported 66% screening yield from a keyword funnel raises questions about recall that the community should probe. As a public corpus with raw CSVs and per-paper AI rationales deposited openly, it provides infrastructure for studying how LLMs perform at evidence synthesis tasks.

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

Read

This paper applies quantum interference principles to semantic vector representations in machine translation, proposing that superposition and interference of meaning states can better preserve contextual coherence across long passages than conventional attention-based approaches. Context preservation across document boundaries remains a genuine open problem, particularly for legal, medical, and literary translation where meaning shifts subtly across paragraphs. However, the full methodology is inaccessible from available metadata, so the empirical claims cannot be evaluated — treat as a watch-list item pending full text review.

██████████ 0.5 long-context Peer-reviewed

Read

This conceptual review surveys radio-domain counter-UAS techniques — including anti-jamming waveforms, GPS anti-spoofing, and LPI/LPD signal design — and proposes a cortex-inspired neural architecture as a future framework for autonomous counter-drone systems. The cortex-inspired framing is motivated by the brain's ability to integrate noisy multi-modal signals and adapt in real time, capabilities that current rule-based counter-UAS systems lack in contested RF environments. As a position paper without empirical results, it is most useful for framing the design space rather than providing implementable solutions.

██████████ 0.4 embodied-ai Peer-reviewed

Read