LLM Systems

40 beliefs (40 IN, 0 OUT)

The "LLM Systems" topic constitutes a broad knowledge base spanning the full lifecycle of large language models: from training economics and data engineering, through architectural and representational structure, to inference serving, behavioral characterization, evaluation, and downstream application. It matters because it captures not just what LLMs do but how they are built, deployed, constrained, and increasingly integrated into agent workflows (llm-agent-protocols-mcp-a2a) and sensitive domains such as mental-health support, where 48.7% of U.S. adults with conditions report using them for therapy (48.7-percent-mental-health-llm-therapy-usage). The beliefs collectively position LLMs as a subset of foundation models (llm-subset-of-foundation-models) while treating them as a distinct engineering and scientific object with specific cost structures, failure modes, and design trade-offs.

Several key claims form the structural backbone. On the training side, the computational cost asymmetry between training (approximately 6 FLOPs per parameter per token) and inference (1-2 FLOPs) (training-6-inference-1-2-flops-per-parameter-per-token, training-cost-6-flops-per-param-per-token) frames the economic stakes that DeepSeek-R1 disrupted by matching OpenAI o1's reasoning performance at roughly 95% lower cost using pure reinforcement learning (deepseek-r1-matches-o1-at-95-percent-lower-cost, deepseek-r1-pure-rl-reasoning-jan-2025, deepseek-r1-open-weight-reasoning). This feeds directly into the broader claim that frontier competition now spans proprietary and open-weight tracks, each driving different aspects of progress (frontier-competition-spans-proprietary-and-open-weight). On the representational side, Park's 2025 orthogonality theorems establish that hierarchically related and causally separable concepts occupy orthogonal subspaces in LLM representation space (park2025-hierarchy-orthogonality-subspaces, park2025-orthogonality-causally-separable), and the geometry-as-universal-semantic-currency-v2 belief synthesizes these into a model-independent geometric framework that underpins evaluation, editing, and interpretability. The memory architecture beliefs (complete-memory-hierarchy-bridges-capacity-and-persistence, memory-evolving-from-wider-windows-to-persistent-state) propose a two-axis development along horizontal capacity and vertical temporal persistence, with bridging abstractions at three temporal levels, though they explicitly flag whether a complete unified hierarchy exists as an open question.

The behavioral and evaluation beliefs chart a more sober picture. LLMs exhibit persistent positional bias (lost-in-middle effect) that survives architecture changes, query-aware contextualization, and instruction fine-tuning (llm-lost-in-middle-three-mitigations-incomplete), and the U-shaped curve persists even with random distractors (llm-lost-in-middle-ablation-random-distractors-persist-u-shape). They are order-sensitive and favor popular entities (llm-order-sensitivity-popularity-bias), shift preferences when evidence is fragmented (llm-fragmentation-inconsistency-strategyqa), and show reader saturation before retriever recall is exhausted (llm-lost-in-middle-reader-saturation-before-recall). Critically, Xu et al. (2024) prove via Corollary 1 that no computable LLM can prevent itself from hallucinating, meaning self-correction strategies like chain-of-thought or self-consistency cannot eliminate the problem (xu-2024-corollary-1-llm-cannot-self-mitigate). The evaluation landscape is explicitly multi-paradigm, ranging from intrinsic perplexity through multi-dimensional benchmarks like HELM to task-specific rankings such as MMLU and LMArena (evaluation-landscape-spans-intrinsic-and-extrinsic-paradigms, llm-benchmarks-mmlu-hle-lmarena), without a clean characterization of how intrinsic quality maps to extrinsic task performance. On the inference stack, the beliefs describe a layered optimization discipline spanning algorithmic, memory, compute, and caching concerns (inference-optimization-is-multi-layer-discipline), a canonical pipeline order (inference-optimization-pipeline-order), and hardware democratization through llama.cpp and llamafile (llm-inference-democratized-across-hardware-and-distribution), with vLLM's PagedAttention and continuous batching as a key serving innovation (vllm-uses-pagedattention-and-continuous-batching, vllm-pagedattention) and TensorRT-LLM as the Nvidia-specific counterpart (tensorrt-llm-nvidia-gpu-specific-inference, tensorrt-llm-nvidia-optimized).

Regarding belief status and structure, all beliefs in this topic are currently IN; none have been retracted, so there is no implication of prior claims being withdrawn or contradicted by the system. In terms of epistemic role, the empirical and bibliographic claims—Carlini's 2021 extraction attack (carlini-2021-exact-training-data-extraction-from-llm), Lee et al.'s deduplication result (training-data-deduplication-improves-llm-performance), Shen et al.'s computational infeasibility finding (sub-model-search-computationally-infeasible), the TPU-versus-GPU hardware distinction (tpu-asic-vs-gpu-general-purpose), and the quantization taxonomy (quantization-static-vs-dynamic)—function as relatively stable premises grounded in specific publications or engineering facts. The more synthetic beliefs, such as the memory-hierarchy framing, the geometric unification claim, the frontier-competition characterization, and the inference-democratization narrative, appear to be derived or interpretive syntheses that build on those premises to draw broader architectural and strategic conclusions. The MeLLo no-weight-modification belief (mello-no-weight-modification) and the grooming/poisoning belief (llm-grooming-deliberate-training-data-poisoning) introduce security and editing dimensions that connect back to the training-data and representational-structure claims, rounding out a picture in which LLMs are simultaneously a scientific object to be understood, an engineering stack to be optimized, a competitive product to be costed, and a system with non-trivial failure modes that cannot be self-corrected.

IN 48.7-percent-mental-health-llm-therapy-usage
48.7% of U.S. adults with mental-health conditions who used LLMs reported using them for therapy or emotional support, per a Sentio University study in early 2025
IN carlini-2021-exact-training-data-extraction-from-llm
Carlini et al. (USENIX 2021) demonstrated both membership inference and exact extraction of training data from LLMs, establishing a concrete privacy attack vector on deployed language models
IN complete-memory-hierarchy-bridges-capacity-and-persistence
LLM memory architecture appears to be developing along two orthogonal axes — horizontal capacity expansion (10,000x context window growth enabling richer intra-session memory) and vertical persistence (cross-session state consolidation enabling inter-session memory) — with at least three temporal levels (intra-turn ghost attention, inter-episode reflexion, cross-session dreaming) that may serve as bridging abstractions between the two, though whether these form a complete or unified hierarchy remains an open question.
IN deepseek-r1-matches-o1-at-95-percent-lower-cost
DeepSeek-R1 (2025) matched OpenAI o1's performance using pure reinforcement learning at approximately 95% lower training cost, as reported by VentureBeat and Nature
IN deepseek-r1-open-weight-reasoning
DeepSeek R1 (January 2025) is a 671B parameter open-weight reasoning model that matched proprietary reasoning models using pure reinforcement learning at dramatically lower cost
IN deepseek-r1-pure-rl-reasoning-jan-2025
DeepSeek-R1 (released January 2025) achieves reasoning performance comparable to OpenAI o1 using pure reinforcement learning without supervised fine-tuning, at approximately 95% lower cost
IN evaluation-landscape-spans-intrinsic-and-extrinsic-paradigms
LLM evaluation spans from intrinsic information-theoretic metrics (perplexity as the exponential of average negative log-likelihood) through multi-dimensional benchmarks (HELM evaluating accuracy, calibration, robustness, fairness, and other dimensions) to task-specific benchmarks and rankings (MMLU, Humanity's Last Exam, LMArena) — reflecting a landscape where multiple evaluation paradigms coexist, though the relationship between intrinsic quality metrics and extrinsic task performance is not explicitly characterized by these benchmarks alone.
IN frontier-competition-spans-proprietary-and-open-weight
Frontier LLM competition now spans both proprietary (GPT's exponential capability scaling, Claude's agentic platform expansion) and open-weight (DeepSeek R1 matching proprietary reasoning models) tracks, with each driving different aspects of progress — capability ceilings from proprietary investment, accessibility and cost from open-weight alternatives.
IN geometry-as-universal-semantic-currency-v2
The covariance geometry serves as a unifying geometric framework for LLM semantics, underpinning what can be measured (cosine/Spearman evaluation in SBERT/MTEB), what can be modified (covariance-whitening-based editing as in ROME), what converges across architectures (cross-model universality of feature geometry), and what is structured at the interpretable level (SAE feature neighborhoods as empirical instantiations of the subordination relations in Park's orthogonality theorem)—supporting the view that it functions as a model-independent semantic structure rather than an architecture-specific artifact.
IN inference-optimization-is-multi-layer-discipline
LLM inference optimization operates at multiple layers simultaneously: algorithmic (speculative decoding with draft models), memory management (PagedAttention's KV cache paging), compute (Flash Attention's IO-aware tiling), and caching (KV reuse across tokens).
IN inference-optimization-pipeline-order
The LLM inference optimization pipeline follows the sequence: model compression → knowledge distillation → quantization → speculative decoding → PagedAttention → batched serving (vLLM, SGLang, TensorRT-LLM)
IN llm-agent-protocols-mcp-a2a
LLM agents operate via protocols including Model Context Protocol (MCP) and Agent2Agent (A2A), with frameworks like LangChain and CrewAI.
IN llm-benchmarks-mmlu-hle-lmarena
Key LLM benchmarks and evaluation methods include MMLU, Humanity's Last Exam, LMArena, LLM-as-a-Judge, and perplexity.
IN llm-fragmentation-inconsistency-strategyqa
When one side's evidence is split into fragments, LLMs shift their preference to the other intact side in 38.2% of cases on StrategyQA.
IN llm-grooming-deliberate-training-data-poisoning
LLM grooming is the deliberate mass-publishing of web content to bias LLM training data and outputs, a term coined by the American Sunlight Project in 2025 (e.g., the Pravda network).
IN llm-inference-democratized-across-hardware-and-distribution
The LLM inference stack has been democratized along two independent axes: hardware accessibility (llama.cpp enabling CPU-only execution via C++ reimplementation, eliminating GPU requirements) and distribution simplicity (llamafile bundling model weights and runtime into single executables, eliminating installation complexity), together making frontier-class model execution accessible on commodity hardware without specialized toolchains.
IN llm-inference-engines-vllm-llamacpp
LLM inference engines include vLLM, llama.cpp, TensorRT-LLM, and SGLang.
IN llm-inference-optimization-techniques
LLM inference optimization techniques include model compression, knowledge distillation, speculative decoding, and PagedAttention.
IN llm-lost-in-middle-ablation-random-distractors-persist-u-shape
In Liu et al. 2023 Appendix B, replacing hard-negative Wikipedia distractors with randomly sampled Wikipedia documents yields higher absolute accuracy for all models but the U-shaped positional curve persists.
IN llm-lost-in-middle-reader-saturation-before-recall
In open-domain QA pipelines, the LM reader's accuracy saturates before the retriever's recall is exhausted, meaning adding more retrieved passages does not improve accuracy beyond a point.
IN llm-lost-in-middle-three-mitigations-incomplete
The three mitigations studied in Liu et al. 2023 (model architecture changes, query-aware contextualization, and instruction fine-tuning) do not fully eliminate the mid-context positional bias in LLMs.
IN llm-order-sensitivity-popularity-bias
LLMs are sensitive to the order of documents in context and favor more popular entities and answers corroborated by more documents (Xie et al., 2023).
IN llm-short-counter-answer-adoption-rate
LLMs adopt a short counter-answer (bare answer without explanation) in only 18.8% of cases versus 56.7% for a full counter-memory paragraph on POPQA.
IN llm-stubbornness-50-case-inspection
Manual inspection of 50 cases where LLMs rejected all counter-memory revealed 34 were attributable to ambiguity, commonsense impossibility, or suggestive question phrasing.
IN llm-subset-of-foundation-models
LLMs are a subset of foundation models, which are pretrained on broad data and adapted to downstream tasks
IN mello-no-weight-modification
MeLLo keeps the LLM frozen and stores edited facts in an external memory, using iterative sub-question decomposition and consistency checking at inference time without any gradient updates
IN memory-evolving-from-wider-windows-to-persistent-state
LLM memory capability is evolving along two distinct axes: horizontal expansion (10,000x context window growth from 1K to 10M tokens over seven years) and temporal persistence (Dreaming consolidating memory between sessions) — suggesting the next frontier is not how much a model can process at once but what it retains across interactions.
IN park2025-hierarchy-orthogonality-subspaces
Hierarchically related concepts (e.g., animal ⊃ mammal ⊃ dog) occupy orthogonal subspaces in LLM representation space under a suitable inner product.
IN park2025-orthogonality-causally-separable
Causally separable binary concepts (independently manipulable) are represented by orthogonal directions in LLM representation space under a suitable inner product.
IN quantization-static-vs-dynamic
LLM quantization reduces precision post-training to shrink model size; static quantization uses a calibration phase while dynamic quantization adjusts at inference time
IN sub-model-search-computationally-infeasible
Shen et al. (2023) could not exhaustively search sub-models to identify which parameters correspond to cGD's 'updated' weights due to computational infeasibility at LLM scale.
IN tensorrt-llm-nvidia-gpu-specific-inference
TensorRT-LLM is an inference framework specifically tied to Nvidia GPUs for compilation and serving optimization, distinct from CPU-oriented alternatives like llama.cpp
IN tensorrt-llm-nvidia-optimized
TensorRT-LLM is an Nvidia GPU-optimized LLM serving toolkit
IN tpu-asic-vs-gpu-general-purpose
TPUs (Google) are ASICs optimized specifically for matrix multiplication, while GPUs (NVIDIA) are general-purpose parallel processors; both are critical for LLM training and inference at scale
IN training-6-inference-1-2-flops-per-parameter-per-token
Training an LLM costs approximately 6 FLOPs per parameter per token, while inference costs approximately 1–2 FLOPs per parameter per token.
IN training-cost-6-flops-per-param-per-token
LLM training costs approximately 6 FLOPs per parameter per token, while inference costs 1-2 FLOPs per parameter per token.
IN training-data-deduplication-improves-llm-performance
Deduplication of training data measurably improves LLM performance, as demonstrated by Lee et al. (2022), making it a standard data-engineering practice
IN vllm-pagedattention
vLLM is an LLM inference and serving framework known for introducing PagedAttention
IN vllm-uses-pagedattention-and-continuous-batching
vLLM is a high-throughput LLM serving framework that uses PagedAttention for memory management and continuous batching for inference optimization
IN xu-2024-corollary-1-llm-cannot-self-mitigate
Corollary 1 of Xu et al. (2024) states that no computable LLM can prevent itself from hallucinating, meaning methods relying on the LLM to detect or correct its own errors (chain-of-thought, self-consistency, self-verification) cannot eliminate hallucination.