I’m sorry, but as of my last update, I don’t have specific information on a game called “Marvel Rivals” or its “Team Up abilities.” If this is a new or less-publicized game, I suggest checking the game’s official resources or community forums for detailed information and guides. If you have more context or can describe what the abilities do, I might be able to help brainstorm alternate title suggestions.

Comprehending the Constraints of Language Model Transparency

With large language models (LLMs) increasingly integrated into various applications, from enterprise decision support to education and scientific research, the need to understand their internal mechanics becomes crucial. One of the primary challenges is identifying the origins of a model’s response. Although LLMs are educated using vast datasets with trillions of tokens, there still lacks a functional way to trace model outputs back to their data sources. This lack of transparency complicates efforts to assess trustworthiness, pinpoint factual sources, and scrutinize for potential memorization or biases.

OLMoTrace – A Real-Time Tool for Output Tracing

The Allen Institute for AI (Ai2) has developed OLMoTrace, a system designed to trace portions of LLM-generated responses back to their training data instantaneously. Built on Ai2’s open-source OLMo models, it offers an interface for identifying verbatim matches between generated text and the documents used during the model’s training. In contrast to retrieval-augmented generation (RAG) approaches which add external context during inference, OLMoTrace focuses on post-hoc interpretability—highlighting the connections between model behavior and past training exposure.

OLMoTrace is featured in the Ai2 Playground, where users can explore specific segments in an LLM output, view matching training documents, and examine those documents in extended context. The tool supports OLMo models like OLMo-2-32B-Instruct and utilizes their comprehensive training data, which encompasses over 4.6 trillion tokens across 3.2 billion documents.

Technical Architecture and Design Considerations

Central to OLMoTrace is infini-gram, an indexing and search engine crafted for extreme-scale text corpora. Utilizing a suffix array-based framework allows for efficient searching of exact text spans from the model’s outputs in the training data. The main inference pipeline follows five stages:

1. Span Identification: Extracts maximal spans from a model’s output that match exact sequences in the training data while ignoring incomplete, overly common, or nested spans.
2. Span Filtering: Ranks spans according to “span unigram probability,” favoring longer and rarer phrases as they are generally more informative.
3. Document Retrieval: Retrieves up to 10 documents containing each span, balancing between precision and runtime efficiency.
4. Merging: Consolidates overlapping spans and duplicates to minimize redundancy in the user interface.
5. Relevance Ranking: Employs BM25 scoring to rank retrieved documents by their similarity to the original prompt and response.

This architecture ensures that tracing results are accurate and presented swiftly, with an average latency of 4.5 seconds for a 450-token output. All processes run on CPU-based nodes utilizing SSDs for accommodating large index files with low-latency access.

Evaluation, Insights, and Use Cases

Ai2 evaluated OLMoTrace using 98 LLM-generated conversations. Human annotators and a model-based “LLM-as-a-Judge” evaluator (GPT-4o) scored document relevance. The top retrieved document averaged a relevance score of 1.82 (on a 0–3 scale), and the top five averaged 1.50—suggesting a reasonable correlation between model outputs and retrieved training contexts.

Three illustrative use cases showcase the system’s utility:

• Fact Verification: Users can verify if a factual statement was likely memorized by examining its source documents.
• Creative Expression Analysis: Even seemingly novel expressions can sometimes be traced back to fan fiction or literary samples within the training corpus.
• Mathematical Reasoning: OLMoTrace can identify matches for symbolic computation steps or structured problem-solving examples, offering insights into how LLMs learn mathematical tasks.

</