89

Papers

10/10

Roadblocks Active

1

Connections

This study grafts a block-sparse attention selector (following the SeerAttention design) onto a frozen Qwen2.5-1.5B language model and tests it at 32K-token context lengths using the RULER benchmark. One specific configuration — increasing the selector's hidden dimension from 128 to 256 — is the only variant that passes both acceptance criteria at that scale. This matters because it provides a reproducible data point on how sparse attention mechanisms need to be tuned as context length grows, a practical constraint for deploying long-context models efficiently.

██████████ 0.8 efficiency-scaling Peer-reviewed

Read

The paper argues that in-context learning — the ability of large language models to adapt to new tasks from a few examples — emerges from specific attention circuit structures called induction heads that appear suddenly during a training phase transition. These circuits are claimed to implement something like Bayesian inference by tracking statistical patterns in the input window. Understanding exactly which circuits produce which capabilities is foundational to AI safety and interpretability work, because it tells researchers where to look when a model behaves unexpectedly — though the paper's methodological details and reproducibility could not be verified from available metadata.

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

Read

A second Zenodo deposit describing the same claims about induction head circuits underlying in-context learning, using ablation studies and attention pattern visualization across models from 125M to 52B parameters. The duplicate deposition and absence of linked code or data limit confidence in the contribution's originality, but the core claims — that discrete circuits enable task adaptation from examples and that these circuits emerge at identifiable training checkpoints — are consistent with published mechanistic interpretability literature and remain scientifically important if verified.

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

Read

This paper addresses a structural risk in AI agent systems: when a large language model acts as a sensor feeding beliefs into a goal-directed agent, its hallucinated or contradictory outputs can corrupt the agent's internal world model before any safety check occurs. The proposed fix, called HADD, inserts a verification gate between perception and belief commitment in the classical BDI agent architecture, converting unverified LLM outputs into traceable epistemic error flags rather than silently accepted beliefs. This kind of upstream filtering is practically important for any agentic system that relies on LLM-generated observations to drive real-world actions.

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

Read

This paper describes a training approach in which a language model is given a written constitution of principles and then asked to critique and revise its own outputs before learning from them, reducing the need for human labelers compared to standard reinforcement learning from human feedback. The abstract claims a 90% reduction in human oversight requirements while improving alignment quality, but the full paper content was not accessible for verification and no code or data is linked. The constitutional AI concept is already well-established in published Anthropic research; whether this Zenodo deposit adds new empirical results or reproduces prior findings cannot be assessed.

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

Read

This paper proposes that any system capable of genuine interpretation must pass through seven distinct minimal conditions — including trace detection, temporal ordering, and re-entry availability — rather than simply performing computations. The motivation is conceptually relevant to AI interpretability: if we can formally separate what a model computes from what it means, we have a sharper language for diagnosing failure modes. However, the framework is purely axiomatic with no computational implementation or empirical test, making it a philosophical proposal rather than a technical contribution at this stage.

██████████ 0.6 interpretability 🔗 4 cited Peer-reviewed

Read

This paper proposes that current AI systems lack a temporal dimension of reasoning, causing what the authors call 'temporal hallucinations' and value alignment failures that cannot be fixed by scaling alone. The proposed solution is a two-layer architecture separating spatial logic (current LLMs) from temporal logic with a 'finiteness theorem' as a boundary operator. The paper was drafted by Claude based on human prompts and validated only by LLM conversation, not experiment — the core diagnosis about temporal blindness in transformers is a recognized research topic, but this specific framing adds no empirical evidence.

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

Read

This paper argues via the informal 'Coffee Count Theorem' that real numbers are downstream constructions of integers — you cannot measure a continuous fluid without discrete containers — and extends this to claim that AI hallucinations arise from structural mismatches between discrete symbolic processing and continuous outputs. The discrete-versus-continuous framing of AI failures is a legitimate research question with serious technical literature behind it, but this paper offers no formal proofs, no experiments, and no connection to existing mathematical or machine learning research on the topic.

██████████ 0.5 hallucination-grounding Peer-reviewed

Read

This French-language paper proposes a 'Human-in-the-Loop Orchestration Model' in which multiple AI systems (ChatGPT, Claude, Gemini, NotebookLM) collaborate under human epistemic governance rather than autonomously producing knowledge. The paper was itself developed using the multi-AI workflow it describes, treating the creation process as a demonstration. The governance framing is relevant to AI alignment but the paper is entirely conceptual with no evaluation criteria, making it impossible to assess whether the model actually improves knowledge quality or safety outcomes.

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

Read

This paper describes an architecture for a locally hosted AI assistant with a dual-index vector database that keeps interaction history separate from factual knowledge archives to prevent context contamination, paired with a 34-opcode virtual machine for deterministic session state tracking. Separating episodic from semantic memory is a well-motivated design pattern for reducing hallucination drift in retrieval-augmented systems. However, no implementation code, benchmark results, or baselines are provided, and the repository contains only PDF and media files, making the claimed performance improvements unverifiable.

██████████ 0.4 hallucination-grounding Peer-reviewed

Read