JET (Justification Evaluation in Transformer) leverages pretrained language and vision‑language models to choose among a limited set of answers without extra training. It directly evaluates candidate likelihoods, reuses computation across candidates, and runs experiments on desktop CPUs and consumer GPUs to measure decision accuracy and execution cost. Results show high accuracy on the MMLU test set, significant speedups from prefix reuse and cache management, and a 30.8% reduction in process time through input preparation optimizations, all while maintaining unchanged outputs.
arXiv:2607. 09172v1 Announce Type: cross Abstract: Large Language Models are reshaping how software is developed and maintained.
The paper demonstrates that the number of candidates generated during test-time scaling of large language models does not fully capture the system cost. By comparing different generation schedules (e.g., one batched call versus multiple serial calls) while keeping the total candidate count fixed, the authors show that serial calls consume significantly more GPU energy and latency. The study suggests that reporting candidate count alone is insufficient; evaluations should also include generation schedule and GPU-level metrics.
arXiv:2609.36222v1 Announce Type: new Abstract: Large language models are increasingly expensive to serve. In large-scale serving systems, autoregressive decoding is often bottlenecked by transferrin...
arXiv:2608. 15383v1 Announce Type: new Abstract: Sparse mixture-of-experts (MoE) language models reduce arithmetic by activating only a small subset of experts per token, yet deployment still requires storing and moving the full expert bank.
The paper presents a Pareto atlas of LLM inference optimizations, mapping cost, quality, and latency trade‑offs for Qwen2.5‑7B‑Instruct on L4, A100, and H100 GPUs. Using 54 measured configurations and a calibrated simulator, it identifies 18 of 36 setups on the Pareto frontier, showing that combined methods outperform single ones. Quality tests reveal that AWQ 4bit and FP8 weights offer significant latency reductions while largely preserving accuracy, but naive FP8 KV caching fails to answer any questions correctly.
arXiv:2607. 25583v1 Announce Type: new Abstract: Parameter-efficient fine-tuning (PEFT) and low-bit quantization are now standard tools for adapting language models under tight compute budgets, yet their interaction is most often studied on billion-parameter models where the design space is expensive to explore.
arXiv:2609.01343v1 Announce Type: new Abstract: Looped Transformers increase effective depth by iterating a shared block of layers, but most evaluations compare at fixed model size, conflating archit...
The paper studies how the design of Mixture-of-Experts (MoE) routers affects inference speed when combined with Speculative Decoding (SD). It shows that routers promoting high expert coactivation reduce memory transfer costs and improve runtime. By integrating a global load‑balancing loss, shared experts, a consistency loss, and an autoregressive expert selection mechanism, the authors achieve a 21% throughput gain over baseline MoEs while preserving accuracy.
Parameter-efficient fine-tuning (PEFT) and low-bit quantization are now standard tools for adapting language models under tight compute budgets, yet their interaction is most often studied on billion-parameter models where the design space is expensive to explore. We ask a complementary question: on a specific, fully reproducible 60M-parameter encoder-decoder model (T5-small) and a single-table text-to-SQL benchmark (WikiSQL), how much task accuracy does each efficiency knob actually cost?
arXiv:2607. 24434v1 Announce Type: cross Abstract: Large Mixture-of-Experts (MoE) language models are attractive for end-device deployment because only a small subset of experts is active per token, but their routed expert weights often exceed accelerator memory.
The Transformer Accelerator (TFA) is a synthesizable, parameterizable INT8 memory‑to‑memory engine designed for transformer inference and machine translation. It features a one‑time‑multiplexed datapath that handles prompt processing and autoregressive generation, and implements key operations such as matrix multiplication, softmax, RMSNorm, and elementwise functions through eight 512‑bit macro‑op descriptors. In extensive verification, TFA achieved zero mismatches across 25 tests and 34 constrained‑random runs, matched floating‑point references on multiple translation tasks, and delivered a 20× speedup over a 22‑thread CPU while projecting significant energy reductions in larger designs.
PerfReasoning is a new benchmark that tests large language models (LLMs) on their ability to reason about hardware performance and generate analytical performance‑model code. The benchmark presents workloads, architectures, and mapping specifications, asking models to compare mappings and predict off‑chip traffic and buffer requirements. While the best closed‑source models achieve over 90% accuracy on reasoning‑based Q&A and the top open‑weight model scores 82.4%, constructing full performance models remains difficult, with most models scoring below 15% and significant variability across runs. Task‑specific reinforcement learning can improve a 4B model’s mapping‑reasoning accuracy by 15.7 points, but feedback‑free self‑revision prompting is not reliably effective.