Marcus Armstrong

Navid Ayoobi, Marcus I Armstrong, Arjun Mukherjee

Large language models (LLMs) reason over discrete token ID sequences, yet modern subword tokenizers routinely produce non-unique encodings: multiple token ID sequences can detokenize to identical surface strings. This representational mismatch creates an unmeasured fragility wherein reasoning processes can fail. LLMs may treat two internal representations as distinct "words" even when they are semantically identical at the text level. In this work, we show that tokenization can betray LLM reasoning through one-to-many token ID mappings. We introduce a tokenization-consistency probe that requires models to replace designated target words in context while leaving all other content unchanged. The task is intentionally simple at the surface level, enabling us to attribute failures to tokenizer-detokenizer artifacts rather than to knowledge gaps or parameter limitations. Through analysis of over 11000 replacement trials across state-of-the-art open-source LLMs, we find a non-trivial rate of outputs exhibit phantom edits: cases where models operate under the illusion of correct reasoning, a phenomenon arising from tokenizer-induced representational defects. We further analyze these cases and provide a taxonomy of eight systematic tokenizer artifacts, including whitespace-boundary shifts and intra-word resegmentation. These findings indicate that part of apparent reasoning deficiency originates in the tokenizer layer, motivating tokenizer-level remedies before incurring the cost of training ever-larger models on ever-larger corpora.

Marcus Armstrong, ZiWei Qiu, Huy Q. Vo et al.

Large Language Models (LLMs) possess a theoretical capability to model information density far beyond the limits of classical statistical methods (e.g., Lempel-Ziv). However, utilizing this capability for lossless compression involves navigating severe system constraints, including non-deterministic hardware and prohibitive computational costs. In this work, we present an exploratory study into the feasibility of LLM-based archival systems. We introduce \textbf{Hybrid-LLM}, a proof-of-concept architecture designed to investigate the "entropic capacity" of foundation models in a storage context. \textbf{We identify a critical barrier to deployment:} the "GPU Butterfly Effect," where microscopic hardware non-determinism precludes data recovery. We resolve this via a novel logit quantization protocol, enabling the rigorous measurement of neural compression rates on real-world data. Our experiments reveal a distinct divergence between "retrieval-based" density (0.39 BPC on memorized literature) and "predictive" density (0.75 BPC on unseen news). While current inference latency ($\approx 2600\times$ slower than Zstd) limits immediate deployment to ultra-cold storage, our findings demonstrate that LLMs successfully capture semantic redundancy inaccessible to classical algorithms, establishing a baseline for future research into semantic file systems.

Marcus Armstrong, Navid Ayoobi, Arjun Mukherjee

We present a feedforward graph architecture in which heterogeneous frozen large language models serve as computational nodes, communicating through a shared continuous latent space via learned linear projections. Building on recent work demonstrating geometric compatibility between independently trained LLM latent spaces~\cite{armstrong2026thinking}, we extend this finding from static two-model steering to end-to-end trainable multi-node graphs, where projection matrices are optimized jointly via backpropagation through residual stream injection hooks. Three small frozen models (Llama-3.2-1B, Qwen2.5-1.5B, Gemma-2-2B) encode the input into a shared latent space whose aggregate signal is injected into two larger frozen models (Phi-3-mini, Mistral-7B), whose representations feed a lightweight cross-attention output node. With only 17.6M trainable parameters against approximately 12B frozen, the architecture achieves 87.3\% on ARC-Challenge, 82.8\% on OpenBookQA, and 67.2\% on MMLU, outperforming the best single constituent model by 11.4, 6.2, and 1.2 percentage points respectively, and outperforming parameter-matched learned classifiers on frozen single models by 9.1, 5.2, and 6.7 points. Gradient flow through multiple frozen model boundaries is empirically verified to be tractable, and the output node develops selective routing behavior across layer-2 nodes without explicit supervision.