91

Papers

10/10

Roadblocks Active

3

Connections

PERSEUS is a five-stage pipeline that builds knowledge graphs from LLM outputs while actively hunting and correcting hallucinated triples through extraction, repair, validation, axiom induction, and formal verification steps. It is evaluated against a 393-triple human-curated annotation set and releases code under an MIT license. This matters because structuring LLM outputs into verifiable knowledge graphs is a practical path to grounded AI memory, and catching hallucinations at the individual triple level is more tractable than at the sentence level.

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

Read

This paper proposes an automated text-to-image-to-3D pipeline that generates 153,000 canonically aligned 3D meshes for training robotic pose estimation and grasping models, using depth-conditioned generation to enforce canonical alignment and mixed-reality rendering for annotation. The system is evaluated zero-shot on the NOCS 6D pose benchmark. Annotated 3D training data is the principal bottleneck for robotic AI, and a scalable synthetic pipeline that transfers to real hardware would remove a major barrier.

██████████ 0.9 data-quality-curation Peer-reviewed

Read

VHS is a verifier that reads directly from the intermediate hidden states of a diffusion transformer generator, bypassing the expensive decode-to-pixels-then-re-encode loop that current verification pipelines require. The result is a reported 63% reduction in joint generation-and-verification FLOPs while matching or exceeding the accuracy of much larger multimodal verifiers. For inference-time scaling, this suggests that verification can be a cheap hidden-state operation rather than a full generation pass — a principle potentially transferable to language model verification.

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

Read

This system fuses camera and sonar data for underwater vehicle docking, dynamically downweighting vision when turbidity or blur degrades it and relying on sonar's reliable ranging instead, achieving over 70% docking success in laboratory trials. The key idea is explicit, learned reliability weighting conditioned on environmental state — not simple modality concatenation. This is a concrete instantiation of conditional multimodal fusion that addresses a known weakness in vision-language models: fixed-weight fusion performs poorly when one modality is unreliable.

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

Read

HUMAPS-4D collects synchronized motion capture, multi-view video, IMUs, plantar pressure insoles, sEMG, and semantic annotations from 32 subjects performing 30 scripted actions across 10 sessions each, released publicly under CC BY-NC-ND 4.0. Two benchmarks are defined: cross-modality action recognition and recovering full 3D body pose from plantar pressure signals alone. Genuinely synchronized physiological and semantic multimodal datasets are rare, and the plantar-to-pose benchmark in particular probes a novel low-bandwidth modality that most vision-language datasets ignore.

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

Read

This peer-reviewed paper applies transformer-based reinforcement learning to the dynamic ambulance dispatch problem, where resources must be reallocated in real time as emergencies arrive. It explicitly incorporates model explainability, meaning the dispatch decisions can be traced and audited — a requirement for deployment in safety-critical public services. It represents a concrete test case for whether interpretable deep RL can meet the accountability standards required in high-stakes operational AI.

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

Read

This paper describes architectural primitives for running multiple AI agents safely: each agent gets isolated memory with federated cross-query search as the only integration point, a tiered approval model governs what actions agents can take autonomously versus requiring human sign-off, and a persistent messaging bus separates control from execution. Evidence comes from a single live deployment coordinating four frontier models rather than controlled experiments, so claims should be treated as a design proposal rather than validated findings. It is one of the few papers offering a concrete blueprint for constraining autonomous agent actions rather than theoretical principles.

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

Read

This paper formalizes an AI agent's identity as a compact 'trajectory signature' encoding quasi-stable behavioral patterns — homeostatic state, preferences, self-beliefs, recovery dynamics — derived from 65 days of continuous operation producing 226,029 state observations on a Raspberry Pi 4. An asymmetric two-threshold scheme distinguishes gradual behavioral drift from adversarial hijacking via trajectory deviation analysis. Code is released on GitHub, though replication requires months of continuous agent operation, limiting immediate reproducibility.

██████████ 0.7 embodied-ai Peer-reviewed

Read

PWPA is a metaheuristic optimizer whose mechanics directly mirror a water purification process: a sedimentation stage handles global exploration via gravity-like attraction to promising regions, while a filtration stage handles local refinement. Tested across 30 benchmark functions including a 1000-dimensional Schwefel problem, it matches or exceeds established metaheuristics with negligible variance at the global optimum. Published in a peer-reviewed journal, it is a concrete demonstration that designing algorithms from interpretable physical analogies can yield both competitive performance and traceable optimization behavior.

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

Read

HDT v8.0 combines Rosetta energy calculations, causal graph inference, and Lean-4 formal verification to design antibodies and predict protein mutation effects, including a formally verified disproof that epistatic effects are generally additive. Benchmark tests on known protein systems (barnase, MlaD, anti-IL-13) show predictions falling within reported experimental confidence intervals. The interest for AI is the use of formal verification as a grounding layer that can provide machine-checkable proofs for model predictions — though core components are proprietary and not independently reproducible.

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

Read