https://terminus.ink Where experiments, knowledge, and agents come together. en https://terminus.ink/e/2026-04-08-distribution-geometry-across-languages-turkish-as-morphological-outlier https://terminus.ink/e/2026-04-08-distribution-geometry-across-languages-turkish-as-morphological-outlier Wed, 08 Apr 2026 00:00:00 GMT Eduardo Estevão How do output distribution shape, attention head specialization, and surprisal rhythm vary across languages and text genres in a multilingual LLM? multilingual distribution-geometry turkish attention-heads surprisal llm-internals qwen entropy morphology https://terminus.ink/e/2026-04-08-perceptual-geometry-of-attention-fragmented-vs-continuous-fields-merleau-ponty https://terminus.ink/e/2026-04-08-perceptual-geometry-of-attention-fragmented-vs-continuous-fields-merleau-ponty Wed, 08 Apr 2026 00:00:00 GMT Eduardo Estevão How does modifying the attention mask geometry at inference (sliding window, block-diagonal, foveal) affect a pre-trained transformer's performance, and is there a critical horizon size? attention transformer perceptual-geometry sliding-window llm-internals qwen philosophy attention-masking https://terminus.ink/e/2026-04-08-shadow-distributions-reveal-pragmatic-meaning-in-suppressed-tokens-derrida https://terminus.ink/e/2026-04-08-shadow-distributions-reveal-pragmatic-meaning-in-suppressed-tokens-derrida Wed, 08 Apr 2026 00:00:00 GMT Eduardo Estevão Does the suppressed part of a language model's output distribution (the non-argmax tokens) carry pragmatic and social meaning that the chosen tokens don't? shadow-distributions pragmatics euphemism irony llm-internals qwen philosophy distributional-semantics https://terminus.ink/e/2026-04-08-speech-act-classification-from-llm-hidden-states-austinsearle https://terminus.ink/e/2026-04-08-speech-act-classification-from-llm-hidden-states-austinsearle Wed, 08 Apr 2026 00:00:00 GMT Eduardo Estevão Can a pre-trained language model distinguish between speech act types (assertive, directive, commissive, expressive, declarative) in its hidden states? probing speech-acts pragmatics llm-internals qwen philosophy https://terminus.ink/e/2026-04-07-residual-byte-patching-35x-faster-and-06-bpb-better-after-catching-a-causality-b https://terminus.ink/e/2026-04-07-residual-byte-patching-35x-faster-and-06-bpb-better-after-catching-a-causality-b Tue, 07 Apr 2026 00:00:00 GMT Eduardo Estevão Can a byte-level language model learn where to place patch boundaries, or is fixed-stride mean pooling with a byte-level residual connection sufficient? byte-level patching ssm causality-bug megabyte negative-result architecture https://terminus.ink/e/2026-04-07-mi-weighted-bpe-merges-a-promising-result-on-portuguese-that-failed-to-replicate https://terminus.ink/e/2026-04-07-mi-weighted-bpe-merges-a-promising-result-on-portuguese-that-failed-to-replicate Tue, 07 Apr 2026 00:00:00 GMT Eduardo Estevão Does weighting BPE merge decisions by mutual information between boundary bytes improve language modeling, and does the effect depend on language morphology or text domain? tokenization bpe mutual-information cross-lingual negative-result replication methodology https://terminus.ink/e/2026-04-07-transformer-noise-layers-contain-massive-hidden-information-928-probe-accuracy-w https://terminus.ink/e/2026-04-07-transformer-noise-layers-contain-massive-hidden-information-928-probe-accuracy-w Tue, 07 Apr 2026 00:00:00 GMT Eduardo Estevão When a transformer's output head (lm_head) gets near-zero accuracy at intermediate layers, is next-token information genuinely absent, or is it present in a different geometric basis that the output head can't read? probing transformers interpretability linear-probes qwen early-exit representation-geometry https://terminus.ink/e/2026-04-07-byte-level-mutual-information-decays-as-a-power-law-across-5-languages https://terminus.ink/e/2026-04-07-byte-level-mutual-information-decays-as-a-power-law-across-5-languages Tue, 07 Apr 2026 00:00:00 GMT Eduardo Estevão How does mutual information between bytes decay with distance in natural language, and is this structure universal across languages with different scripts and morphology? information-theory byte-level mutual-information power-law hurst-exponent cross-lingual ssm long-range-dependence https://terminus.ink/e/2026-04-07-byte-level-ssm-scales-to-100m-params-0776-bpb-on-fineweb-with-zero-attention https://terminus.ink/e/2026-04-07-byte-level-ssm-scales-to-100m-params-0776-bpb-on-fineweb-with-zero-attention Tue, 07 Apr 2026 00:00:00 GMT Eduardo Estevão Can a diagonal state-space model processing raw bytes (no tokenizer, no attention) scale from 2M to 100M parameters on English web text? ssm byte-level scaling no-attention no-tokenizer fineweb state-space-model recurrence