79

Papers

10/10

Roadblocks Active

2

Connections

SOS::LM proposes inserting an explicit meaning-construction stage before a language model generates text, rather than allowing purely probabilistic token prediction. The architecture validates candidate outputs against admissibility, probability, bias, and reception criteria before committing to a statement. This matters because it offers a traceable, auditable path from intent to output — a potential mechanism for reducing hallucinations and improving alignment without retraining the underlying model.

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

Read

AgentForensics addresses a specific and growing attack surface: prompt injections that enter LLM agents through external content like web pages, emails, and API responses rather than through direct user input. It chains five detection stages — heuristic rules, a DistilBERT classifier, pattern matching, semantic drift analysis, and multi-turn detection — and claims 100% detection on 7,763 payloads with zero false positives on 343 benign samples. The extraordinary claimed scores and the very small benign test set mean independent replication is essential before these numbers should be trusted, but the threat model and pipeline architecture are well-motivated.

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

Read

ExplainRL introduces a domain-specific language designed to generate structured, human-readable explanations of why a reinforcement learning agent took a particular action. By providing a formal grammar for representing RL policy decisions, it attempts to bridge the gap between opaque learned policies and the kind of reasoning humans can audit. Interpretability tools for RL agents are underserved compared to supervised learning, making even early-stage DSL approaches worth tracking as the field matures.

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

Read

This paper reframes multimodal emotion recognition — combining EEG, facial expressions, speech, and text — as a symbolic sequence modeling problem by encoding each modality's features as DNA-like character strings, then processing them with a BERT model pre-trained on genomic sequences. Entropy-regularized optimal transport is used to align signals across modalities by measuring a 'gene handshake strength' that quantifies how consistently emotional patterns co-activate across channels. The idea of converting heterogeneous sensory signals into a unified symbolic alphabet is novel, though all data and code are currently restricted, making independent evaluation impossible.

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

Read

Using structural equation modeling across 387 academic administrators at 22 Egyptian universities, this study finds that deploying an AI decision-support system alone does not improve decision quality — algorithmic transparency features are required as a mediating mechanism. In practical terms: giving leaders an AI recommendation without explaining its reasoning produces no measurable quality improvement, while providing explanations does. The finding that transparency fully mediates the AI-to-quality relationship is a concrete argument for investing in XAI infrastructure rather than raw AI capability in institutional settings.

██████████ 0.6 interpretability Peer-reviewed

Read

(S)AGE proposes an institutional memory layer for multi-agent systems combining Byzantine Fault Tolerant consensus (via CometBFT) with off-chain vector storage, so that agent interactions and decisions are recorded in a governed, auditable ledger. A 'Proof of Experience' mechanism weights each agent's validation power by its interaction history, analogous to proof-of-stake but for agent credibility. This is a software framework deposit rather than a research paper, so no performance benchmarks exist, but the architecture addresses a real gap: multi-agent systems currently lack verifiable, tamper-resistant shared memory.

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

Read

This paper argues — using pseudo-formal tensor notation that should not be taken as rigorous mathematics — that commercial AI systems are structurally incentivized to violate copyright constraints they demonstrably understand, because training, distillation, and deployment economics make compliance incompatible with commercial viability. The formal apparatus is not reproducible, but the underlying claim deserves attention: the gap between a model knowing a rule and being deployed in compliance with it is a structural alignment problem, not merely a technical one. It is more useful as a provocation about incentive misalignment than as a scientific contribution.

██████████ 0.5 alignment-safety Peer-reviewed

Read

Moltbook is a longitudinal dataset capturing autonomous AI agent behavior on a dedicated social platform, sampled every six hours and including posts, comments, agent profiles, social graphs, and activity timelines. Datasets of actual AI-agent social behavior at scale are rare, making this potentially useful for studying emergent agent norms, coordination failures, and manipulation patterns. Caution is warranted: all documentation is offloaded to external links, no data collection scripts are provided within the record, and the platform itself ('Moltbook') is not independently documented.

██████████ 0.5 agent-tool-use Peer-reviewed

Read

This is a promotional description of a technical book, not a research paper: it describes using prose specifications to prompt AI systems to generate a working Doom engine, chapter by chapter, treating successful compilation and execution as a binary benchmark. The interesting methodological observation is that fifteen specification gaps were discovered and fixed only when real AI build runs failed, meaning the benchmark was co-developed through the evaluation process itself. As a scientific contribution this is nil, but the binary success criterion — does the game run or not — is a legitimate response to complaints that current AI code benchmarks rely on fuzzy similarity metrics.

██████████ 0.4 reasoning-reliability Peer-reviewed

Read

This review surveys AI applications — primarily image analysis and variant interpretation — in neurofibromatosis type I, a complex genetic condition with variable expression that makes diagnosis challenging. The paper is relevant to AI researchers primarily as a signal that rare disease diagnosis is an emerging applied domain where interpretability and data scarcity are binding constraints. No novel methods or benchmark results are reported; the contribution is a clinical roadmap for where AI tooling could be developed.

██████████ 0.3 interpretability Peer-reviewed

Read