LLM & EEM Research

Belief Wiki is a structured knowledge base that tracks what a particular research community currently believes about large language models, natural language processing, and the emerging field of machine interpretability. Every one of its 1946 entries is an active, unretracted belief, organized by a Truth Maintenance System into 29 topical clusters. The domain is broad but tightly focused on the post-BERT, post-GPT-3 era of NLP: how transformers actually work internally, how to measure and translate their internal representations, how to train and align them, how to evaluate them against each other, and where the field's foundational assumptions hold or break. You will find mathematical definitions of attention and representation similarity metrics sitting alongside observations about specific model releases, policy decisions around open-source licensing, and mechanistic analyses of where factual knowledge lives inside a feed-forward network.

The major themes that emerge across the topics are threefold. First, there is a deep investment in interpretability and the geometry of representation space: sparse autoencoders, function vectors, superposition, causal tracing experiments like ROME, and measures like CKA that quantify how aligned two sets of vectors are. Second, the knowledge base tracks the practical engineering and training side with real specificity: how RLHF and reward modeling shape behavior, how LoRA and fine-tuning alter parameters, how retrieval-augmented generation couples a retriever to a generator, and how in-context learning operates without weight updates. Third, there is a persistent thread of alignment, safety, and failure-mode analysis, covering hallucination, confirmation bias in parametric memory, context faithfulness, and the trajectories of frontier labs toward multimodal agentic capabilities.

In terms of structure, the 29 topics range from large clusters like "Other" (292 beliefs, a catch-all for cross-cutting observations that resist a single label), "Retrieval Augmentation" (123), and "Knowledge Editing" (121), down to narrower groups like "Belief Revision & TMS" (19), which reflects the meta-level of the system managing itself. The distribution suggests a community that works at the intersection of theory and empirical result, holding both formal definitions and specific experimental findings as first-class beliefs. The sample entries make this concrete: a belief about the sparsity pattern of query and key weight matrices in LLaMA-7B sits next to one about the Pareto-based evolutionary search in the GEPA prompt optimizer, and both are treated with the same epistemic status.

Notably, the network contains zero retracted (OUT) beliefs. In a mature TMS, OUT beliefs would mark the points where earlier understanding was contradicted by new evidence, giving a visible fossil record of how the field's consensus shifted. Their complete absence here most likely indicates that this knowledge base is in a relatively early phase of accumulation, where the community has been building up its inventory of what it knows rather than yet encountering the contradictions that force it to discard or revise prior claims. As the base grows and newer experimental results conflict with older assumptions, we would expect OUT beliefs to appear and to become one of the most informative parts of the wiki, showing exactly where understanding has moved.

The value of this knowledge base for a newcomer lies in its synthesis. The LLM research literature is scattered across thousands of papers, blog posts, model cards, and conference talks, and the relationships between a finding about attention sparsity, a benchmark result on MTEB, and a policy change in Llama's commercial license are rarely made explicit in any single source. Belief Wiki treats all of these as nodes in a shared network, tagged with enough specificity to be verifiable and organized enough to be navigable. For someone trying to build a mental model of how modern language models work, where their weaknesses lie, and how the field is evolving in 2025 and 2026, it offers a curated, internally consistent snapshot that would otherwise take months of reading to assemble.