A better visual tokenizer can make language-model pipelines produce higher-quality images/videos with fewer inference steps and offer a new compressed token format that speeds downstream generation and saves bandwidth.

ReWOO cuts API token usage and hosting cost by separating planning from tool calls, so multi-step tool-using pipelines can run cheaper and scale with smaller models.

If you run LLMs over tables in batches, reordering rows and fields can cut inference time and API bills materially by increasing prompt-cache reuse; it is a low-cost software change that often outperforms adding hardware.

LLMs let you do pairwise ER without labeled training data and with lower engineering effort; using short prompts can cut API costs substantially but you must combine LLMs with blocking to control scale.

MorphPiece yields better language modeling and embedding quality without changing model architecture, which can improve search, classification, and prediction pipelines but increases token counts and compute.

LazyLLM cuts the time-to-first-token on long prompts and lowers total token computation without model retraining, reducing latency and inference costs for long-context applications.

OThink-R1 cuts costly reasoning tokens at inference while keeping accuracy, lowering latency and per-request compute cost for products that use step-by-step reasoning.

You can run batch jobs (benchmarks, analytics) on the same expensive GPUs used for live LLM inference without degrading customer-facing latency. That turns idle capacity into usable throughput and reduces waste from overprovisioning.

ZipVL cuts compute and memory for large vision-language generation. That lowers cloud GPU costs for long images/videos, reduces time-to-first-token for interactive apps, and increases decoding throughput so more requests fit a GPU.

Compare reasoning methods by token cost, not just accuracy; cheaper self-consistency often gives better accuracy-per-dollar and reduces deployment cost.

You can cut embedding-costs and latency by ~20–34% using post-training token pruning without retraining per task, keeping few-shot accuracy competitive—useful when serving many small intents in production.

You can cut KV-cache memory and often improve long-context accuracy by storing more tokens at lower precision. This reduces GPU memory cost for long inputs and enables longer effective context without model changes.

Co-Saving can cut token bills and developer compute costs by reusing prior multi-agent transitions, while keeping or improving code quality on similar tasks, so teams can scale automated software generation under a fixed budget.

CORL enables predictable cost-vs-accuracy trade-offs from one deployed system. You can run a single LLM controller that adapts to customer budget tiers, saving inference spend at scale while keeping acceptable accuracy.

Adaptive reasoning reduces wasted compute on easy cases and directs budget to hard cases, lowering inference cost and improving reliability where it matters. Training-free solutions deliver quick wins; training-based solutions scale control into the model for repeated production use.

FutureWeaver helps you spend inference budget where it matters across cooperating agents, raising task success per dollar. It automates reusable multi-agent patterns and avoids leaving budget unused—useful for cost-sensitive production agents that combine search, browsing, and reasoning.

GVote can cut GPU memory used by KV-caches about in half without manual tuning. That frees headroom for larger batch sizes, longer contexts, or lower-cost GPUs and reduces engineering time spent tuning budgets per workload.

You can boost reasoning accuracy and cut inference memory by changing only the decoder strategy, making long-form reasoning cheaper and more reliable without retraining.

DOP cuts prefilling compute and real GPU latency substantially while keeping task accuracy near-original, enabling cheaper inference and higher throughput for multimodal deployments.

RadioLLM lets you reuse LLM priors for multiple radio tasks, improving classification and denoising while cutting prompt overhead and latency in many benchmark scenarios.

Cutting token usage by ~78% can substantially lower cloud inference bills and speed debugging by making model logic explicit.

ExLLM turns LLMs into a sample-efficient, no-training optimizer that cuts API cost and runtime and generalizes across chemistry, engineering and code tasks, lowering the barrier to rapid design under limited evaluation budgets.

AdaTP cuts inference FLOPs by up to ~73% on evaluated Video LLMs without losing task accuracy, lowering compute cost for production video understanding.

Replacing an English-focused tokenizer with a language-specific Unigram vocab can improve non-English accuracy and cut fine-tuning and inference costs, lowering time-to-market and cloud bills for localized LLM products.