Thierry Tambe

Yasmine Omri, Ziyu Gan, Zachary Broveak et al.

LLM agents are increasingly deployed on long-horizon tasks requiring sustained reasoning over extended interaction histories. Realizing this at scale requires agents to persistently store, retrieve, and update their own memory across sessions. A rich ecosystem of agent memory systems has emerged spanning flat retrieval, LLM-mediated extraction, consolidating fact stores, and agentic control flows. Yet, their system-level behavior remains uncharacterized. We present the first systems characterization of agent memory. First, we introduce a system-oriented taxonomy classifying agent memory systems along four axes. Second, we build a phase-aware profiling harness attributing cost to construction, retrieval, and generation. Third, we characterize ten representative systems across two benchmark suites, uncovering how design choices shift cost across the write and read paths. Finally, we derive 10 system recommendations covering construction scheduling, capability floors, amortization via query volume, freshness-latency tradeoffs, and fleet-scale management.

Yuzong Chen, Chao Fang, Xilai Dai et al.

The substantial memory bandwidth and computational demands of large language models (LLMs) present critical challenges for efficient inference. To tackle this, the literature has explored heterogeneous systems that combine neural processing units (NPUs) with DRAM-based processing-in-memory (PIM) for LLM acceleration. However, existing high-precision (e.g., FP16) PIM compute units incur significant area and power overhead in DRAM technology, limiting the effective computation throughput. In this paper, we introduce P3-LLM, a novel NPU-PIM integrated accelerator for LLM inference using hybrid numerical formats. Our approach is threefold: First, we propose a flexible mixed-precision quantization scheme, which leverages hybrid numerical formats to quantize different LLM operands with high compression efficiency and minimal accuracy loss. Second, we architect an efficient PIM accelerator for P3-LLM, featuring enhanced compute units to support hybrid numerical formats. Our careful choice of numerical formats allows to co-design low-precision PIM compute units that significantly boost the computation throughput under iso-area constraints. Third, we optimize the low-precision dataflow of different LLM modules by applying operator fusion to minimize the overhead of runtime dequantization. Evaluation on a diverse set of representative LLMs and tasks demonstrates that P3-LLM achieves state-of-the-art accuracy in terms of both KV-cache quantization and weight-activation quantization. Combining the proposed quantization scheme with PIM architecture co-design, P3-LLM yields an average of $4.9\times$, $2.0\times$, and $3.4\times$ speedups over the state-of-the-art LLM accelerators HBM-PIM, Ecco, and Pimba, respectively. Our quantization code is available at https://github.com/yc2367/P3-LLM.git

Hojoon Kim, Yuheng Wu, Thierry Tambe

Embodied AI agents increasingly rely on large language models (LLMs) for planning, yet per-step LLM calls impose severe latency and cost. In this paper, we show that embodied tasks exhibit strong plan locality, where the next plan is largely predictable from the current one. Building on this, we introduce AgenticCache, a planning framework that reuses cached plans to avoid per-step LLM calls. In AgenticCache, each agent queries a runtime cache of frequent plan transitions, while a background Cache Updater asynchronously calls the LLM to validate and refine cached entries. Across four multi-agent embodied benchmarks, AgenticCache improves task success rate by 22% on average across 12 configurations (4 benchmarks x 3 models), reduces simulation latency by 65%, and lowers token usage by 50%. Cache-based plan reuse thus offers a practical path to low-latency, low-cost embodied agents. Code is available at https://github.com/hojoonleokim/MLSys26_AgenticCache.

Coleman Hooper, Thierry Tambe, Gu-Yeon Wei

This work analyzes how attention-based Bidirectional Long Short-Term Memory (BLSTM) models adapt to noise-augmented speech. We identify crucial components for noise adaptation in BLSTM models by freezing model components during fine-tuning. We first freeze larger model subnetworks and then pursue a fine-grained freezing approach in the encoder after identifying its importance for noise adaptation. The first encoder layer is shown to be crucial for noise adaptation, and the weights are shown to be more important than the other layers. Appreciable accuracy benefits are identified when fine-tuning on a target noisy environment from a model pretrained with noisy speech relative to fine-tuning from a model pretrained with only clean speech when tested on the target noisy environment. For this analysis, we produce our own dataset augmentation tool and it is open-sourced to encourage future efforts in exploring noise adaptation in ASR.

Wonsuk Jang, Thierry Tambe

Diffusion Transformers (DiT) achieve strong video generation quality, but their memory and compute costs hinder edge deployment. Quantization can reduce these costs, yet existing methods often degrade video quality under high activation variation and the need to preserve semantic/temporal coherence. We propose SemanticDialect, which advances recent block-wise mixed-format quantization-selecting a per-block optimal format (a dialect) from multiple candidates (a formatbook)-by scaling the formatbook with lookup tables for quantization error and quantized values, enabling efficient per-block format selection and quantization at low online cost. We also introduce activation decomposition that reduces quantization error by re-quantizing and adding back residual errors, with attention-guided salient token selection. We further propose semantic-aware dialect assignment (SeDA) to improve quantized value consistency by sharing a sub-formatbook among semantically correlated tokens. Experiments on video DiT (VDiT) models show that SemanticDialect outperforms prior VDiT quantization methods and fine-grained block-wise format baselines, while approaching FP16 quality on Open-Sora 2.0.

Yasmine Omri, Parth Shroff, Thierry Tambe

The exponential growth of Large Multimodal Models (LMMs) has driven advancements in cross-modal reasoning but at significant computational costs. In this work, we focus on visual language models. We highlight the redundancy and inefficiency in current vision encoders, and seek to construct an adaptive compression method for multimodal data. In this work, we characterize a panoply of visual token selection and merging approaches through both benchmarking and qualitative analysis. In particular, we demonstrate that simple cluster-level token aggregation outperforms prior state-of-the-art works in token selection and merging, including merging at the vision encoder level and attention-based approaches. We underline the redundancy in current vision encoders, and shed light on several puzzling trends regarding principles of visual token selection through cross-modal attention visualizations. This work is a first effort towards more effective encoding and processing of high-dimensional data, and paves the way for more scalable and sustainable multimodal systems.

Wonsuk Jang, Thierry Tambe

The rapidly increasing size of large language models (LLMs) presents significant challenges in memory usage and computational costs. Quantizing both weights and activations can address these issues, with hardware-supported fine-grained scaling emerging as a promising solution to mitigate outliers. However, existing methods struggle to capture nuanced block data distributions. We propose BlockDialect, a block-wise fine-grained mixed format technique that assigns a per-block optimal number format from a formatbook for better data representation. Additionally, we introduce DialectFP4, a formatbook of FP4 variants (akin to dialects) that adapt to diverse data distributions. To leverage this efficiently, we propose a two-stage approach for online DialectFP4 activation quantization. Importantly, DialectFP4 ensures energy efficiency by selecting representable values as scaled integers compatible with low-precision integer arithmetic. BlockDialect achieves 10.78% (7.48%) accuracy gain on the LLaMA3-8B (LLaMA2-7B) model compared to MXFP4 format with lower bit usage per data, while being only 5.45% (2.69%) below full precision even when quantizing full-path matrix multiplication. Focusing on how to represent over how to scale, our work presents a promising path for energy-efficient LLM inference.

Robin Geens, Jonas De Schouwer, Marian Verhelst et al.

The Hardware Lottery posits that research directions are dictated by available silicon compute platforms. We identify a derivative phenomenon, the Hyperscale Lottery, where model architectures are optimized for cloud throughput at the expense of algorithmic efficiency. While State-Space Models (SSMs) such as Mamba were lauded for their linear complexity, ideal for edge intelligence, their evolution from Mamba-1 to Mamba-3 reveals a systematic divergence from edge-native efficiency. We demonstrate that Mamba-3's architectural changes, designed to saturate hyperscale GPUs, impose a significant edge penalty: a 28% latency increase at 880M parameters, worsening to 48% for 15M-parameter models. We argue for decoupling cloud-scale saturation strategies from core architectural design to preserve the viability of single-user, real-time edge intelligence.

Yuheng Wu, Azalia Mirhoseini, Thierry Tambe

Large language models (LLMs) can improve reasoning at inference time through test-time scaling (TTS), where multiple reasoning traces are generated and the best one is selected. Prior work shows that increasing the number of samples K steadily improves accuracy. In this paper, we demonstrate that this trend does not hold indefinitely: at large K, further scaling yields no gains, and certain hard questions remain unsolved regardless of the number of traces. Interestingly, we find that different sampling temperatures solve different subsets of problems, implying that single-temperature scaling explores only part of a model's potential. We therefore propose scaling along the temperature dimension, which enlarges the reasoning boundary of LLMs. Averaged over Qwen3 (0.6B, 1.7B, 4B, 8B) and five representative reasoning benchmarks (AIME 2024/2025, MATH500, LiveCodeBench, Hi-ToM), temperature scaling yields an additional 7.3 points over single-temperature TTS. Temperature scaling also enables base models to reach performance comparable to reinforcement learning (RL)-trained counterparts, without additional post-training. We further provide a comprehensive analysis of this phenomenon and design a multi-temperature voting method that reduces the overhead of temperature scaling. Overall, our findings suggest that TTS is more powerful than previously thought, and that temperature scaling offers a simple and effective way to unlock the latent potential of base models.

Junyang Zhang, Tianyi Zhu, Thierry Tambe

A fundamental reason for the dominance of attention over RNNs and LSTMs in LLMs is its ability to capture long-range dependencies by modeling direct interactions between all tokens, overcoming the sequential limitations of recurrent architectures. Similarly, a key reason why today's vision language models (VLMs) hallucinate and underperform pure language models is that they rely on direct concatenation of image and text tokens with a modality-blinded positional encoding, which conveniently adopts the pretrained LLM backbone but forces unnecessary long-distance attention between semantically related tokens across modalities. This underscores the urgent need for mechanisms that efficiently enhance token locality and cross-modal alignment. In response, we propose Attention Anchor, a parameter-free framework that efficiently groups semantically similar tokens across modalities, improving cross-modal locality. By inserting text tokens near relevant visual patches, we create semantic signposts that reveal true content-based cross-modal attention scores, guiding the model to focus on the correct image regions for tasks such as VQA, MMBench and POPE. This improves answer accuracy and reduces hallucinations without disrupting the prompt's semantic flow. AttAnchor achieves improvements across 13 out of 15 different metrics and benchmarks, including up to 32% gains on reasoning tasks and up to 15% improvements on hallucination benchmarks. AttAnchor enables TinyLLaVA 1B to outperform much larger models like LLaVA 7B and QwenVL 3B on POPE with only 0.1% inference time overhead. To the best of our knowledge, this work is among the first to investigate mixed-modal token grouping, where text and image tokens are clustered jointly into shared groups rather than being grouped within a single modality or merely aligned post-hoc with additional alignment losses.

Yasmine Omri, Connor Ding, Tsachy Weissman et al.

Modern vision language pipelines are driven by RGB vision encoders trained on massive image text corpora. While these pipelines have enabled impressive zero shot capabilities and strong transfer across tasks, they still inherit two structural inefficiencies from the pixel domain: (i) transmitting dense RGB images from edge devices to the cloud is energy intensive and costly, and (ii) patch based tokenization explodes sequence length, stressing attention budgets and context limits. We explore 2D Gaussian Splatting (2DGS) as an alternative visual substrate for alignment: a compact, spatially adaptive representation that parameterizes images by a set of colored anisotropic Gaussians. We develop a scalable 2DGS pipeline with structured initialization, luminance aware pruning, and batched CUDA kernels, achieving over 90x faster fitting and about 97% GPU utilization compared to prior implementations. We further adapt contrastive language image pretraining (CLIP) to 2DGS by reusing a frozen RGB-based transformer backbone with a lightweight splat aware input stem and a perceiver resampler, training only about 7% of the total parameters. On large DataComp subsets, GS encoders yield meaningful zero shot ImageNet-1K performance while compressing inputs 3 to 20x relative to pixels. While accuracy currently trails RGB encoders, our results establish 2DGS as a viable multimodal substrate, pinpoint architectural bottlenecks, and open a path toward representations that are both semantically powerful and transmission efficient for edge cloud learning.

Sai Qian Zhang, Thierry Tambe, Nestor Cuevas et al.

On-device learning allows AI models to adapt to user data, thereby enhancing service quality on edge platforms. However, training AI on resource-limited devices poses significant challenges due to the demanding computing workload and the substantial memory consumption and data access required by deep neural networks (DNNs). To address these issues, we propose utilizing embedded dynamic random-access memory (eDRAM) as the primary storage medium for transient training data. In comparison to static random-access memory (SRAM), eDRAM provides higher storage density and lower leakage power, resulting in reduced access cost and power leakage. Nevertheless, to maintain the integrity of the stored data, periodic power-hungry refresh operations could potentially degrade system performance. To minimize the occurrence of expensive eDRAM refresh operations, it is beneficial to shorten the lifetime of stored data during the training process. To achieve this, we adopt the principles of algorithm and hardware co-design, introducing a family of reversible DNN architectures that effectively decrease data lifetime and storage costs throughout training. Additionally, we present a highly efficient on-device training engine named \textit{CAMEL}, which leverages eDRAM as the primary on-chip memory. This engine enables efficient on-device training with significantly reduced memory usage and off-chip DRAM traffic while maintaining superior training accuracy. We evaluate our CAMEL system on multiple DNNs with different datasets, demonstrating a $2.5\times$ speedup of the training process and $2.8\times$ training energy savings than the other baseline hardware platforms.

Srivatsan Krishnan, Thierry Tambe, Zishen Wan et al.

Aerial autonomous machines (Drones) has a plethora of promising applications and use cases. While the popularity of these autonomous machines continues to grow, there are many challenges, such as endurance and agility, that could hinder the practical deployment of these machines. The closed-loop control frequency must be high to achieve high agility. However, given the resource-constrained nature of the aerial robot, achieving high control loop frequency is hugely challenging and requires careful co-design of algorithm and onboard computer. Such an effort requires infrastructures that bridge various domains, namely robotics, machine learning, and system architecture design. To that end, we present AutoSoC, a framework for co-designing algorithms as well as hardware accelerator systems for end-to-end learning-based aerial autonomous machines. We demonstrate the efficacy of the framework by training an obstacle avoidance algorithm for aerial robots to navigate in a densely cluttered environment. For the best performing algorithm, our framework generates various accelerator design candidates with varying performance, area, and power consumption. The framework also runs the ASIC flow of place and route and generates a layout of the floor-planed accelerator, which can be used to tape-out the final hardware chip.

Thierry Tambe, Coleman Hooper, Lillian Pentecost et al.

Transformer-based language models such as BERT provide significant accuracy improvement for a multitude of natural language processing (NLP) tasks. However, their hefty computational and memory demands make them challenging to deploy to resource-constrained edge platforms with strict latency requirements. We present EdgeBERT, an in-depth algorithm-hardware co-design for latency-aware energy optimization for multi-task NLP. EdgeBERT employs entropy-based early exit predication in order to perform dynamic voltage-frequency scaling (DVFS), at a sentence granularity, for minimal energy consumption while adhering to a prescribed target latency. Computation and memory footprint overheads are further alleviated by employing a calibrated combination of adaptive attention span, selective network pruning, and floating-point quantization. Furthermore, in order to maximize the synergistic benefits of these algorithms in always-on and intermediate edge computing settings, we specialize a 12nm scalable hardware accelerator system, integrating a fast-switching low-dropout voltage regulator (LDO), an all-digital phase-locked loop (ADPLL), as well as, high-density embedded non-volatile memories (eNVMs) wherein the sparse floating-point bit encodings of the shared multi-task parameters are carefully stored. Altogether, latency-aware multi-task NLP inference acceleration on the EdgeBERT hardware system generates up to 7x, 2.5x, and 53x lower energy compared to the conventional inference without early stopping, the latency-unbounded early exit approach, and CUDA adaptations on an Nvidia Jetson Tegra X2 mobile GPU, respectively.

Thierry Tambe, En-Yu Yang, Zishen Wan et al.

Conventional hardware-friendly quantization methods, such as fixed-point or integer, tend to perform poorly at very low word sizes as their shrinking dynamic ranges cannot adequately capture the wide data distributions commonly seen in sequence transduction models. We present AdaptivFloat, a floating-point inspired number representation format for deep learning that dynamically maximizes and optimally clips its available dynamic range, at a layer granularity, in order to create faithful encoding of neural network parameters. AdaptivFloat consistently produces higher inference accuracies compared to block floating-point, uniform, IEEE-like float or posit encodings at very low precision ($\leq$ 8-bit) across a diverse set of state-of-the-art neural network topologies. And notably, AdaptivFloat is seen surpassing baseline FP32 performance by up to +0.3 in BLEU score and -0.75 in word error rate at weight bit widths that are $\leq$ 8-bit. Experimental results on a deep neural network (DNN) hardware accelerator, exploiting AdaptivFloat logic in its computational datapath, demonstrate per-operation energy and area that is 0.9$\times$ and 1.14$\times$, respectively, that of equivalent bit width integer-based accelerator variants.