Searcharxiv⌕ Search

arXiv subjects

Thierry Tambe

Publications and source records attributed to Thierry Tambe.

At least 19 recordsLinked to original sources

LaFlash: Capacity-Driven Acceleration for Coding Agent Workflows via QLC-HBF

As large language models (LLMs) evolve toward agentic AI, coding agents are emerging as memory-intensive workloads that repeatedly reuse persistent states, including contexts and KV caches. Emerging agentic workflows, including Loop/Graph Engineering, cause state amplification through frequent active/dormant transitions, increasing memory pressure. Evicted states incur costly recomputation upon reactivation, particularly under long-context, short-decode workloads. This motivates Capacity-Driven Acceleration (CDA), which explicitly accounts for state amplification to allocate memory capacity between active concurrency and dormant-state retention, replacing recomputation with memory access. In this work, we propose LaFlash, a High-Bandwidth Flash (HBF) architecture that exploits the large capacity of Quad-Level Cell (QLC)-based flash to realize CDA. However, QLC's long read latency demands massive read parallelism under irregular state reactivation and varying context lengths, while limited write endurance constrains placement. LaFlash introduces Session Grouping to balance active request counts and aggregate context lengths across groups, sustaining read parallelism and block utilization through request-exclusive allocation. Furthermore, Fast-Page Allocation exploits QLC's read-latency heterogeneity to mitigate residual inter-group imbalance and sustain bandwidth utilization. On end-to-end coding agent workloads, LaFlash achieves 7.04$\times$ higher throughput and 10.96$\times$ higher energy efficiency over SLC-HBF with batch expansion, and 2.36$\times$ higher throughput over QLC-HBF with CDA. HBM buffering further reduces unnecessary NAND writes, improving QLC write endurance. These results demonstrate the benefits of jointly optimizing state retention and physical data placement for capacity-driven acceleration.

cs.AR↗

If at First You Don't Succeed, Backtrack: Lightweight Recovery for Frozen VLAs

Vision-Language-Action (VLA) models perform well on robotic manipulation, yet remain brittle once their own actions carry them into states their training data barely covers. Recovering from such failures usually requires recovery demonstrations, policy retraining, or a separate model at inference time. We introduce Backtrack, a lightweight failure detection and recovery mechanism for frozen VLAs that needs none of these. A small failure prediction head, trained only on outcome-labeled rollouts of the frozen policy, reads the VLA's existing internal representations to estimate whether the current episode will fail. When the predicted failure probability exceeds a threshold, a controller pauses the policy, retraces the arm's recently achieved motion, and hands control back to the unchanged policy so it can retry from an earlier configuration. The head adds only 0.48-0.88% to the policy's parameter count. We evaluate Backtrack on LIBERO with three VLAs spanning two action-generation families: SmolVLA and $π$0.5, which use flow-matching action experts, and the autoregressive $π$0-FAST. On LIBERO-Long, the suite of long-horizon multi-step tasks and the one on which all three policies fail most often, Backtrack reduces each policy's failure rate by 19-48%, raising success by up to 7.4 percentage points; on the three shorter suites, where control success is already 85-99%, its effect is smaller. On MetaWorld, all twelve controller configurations tested with SmolVLA improve on the unassisted policy.

cs.RO↗

Spectral-MoE: Heterogeneous Per-Layer Expert Utilization in Dense-to-MoE Conversion

Large language models impose substantial memory demands and arithmetic costs during inference. Dense-to-mixture-of-experts (MoE) conversion is a model optimization technique that aims to reduce these arithmetic costs. Dense-to-MoE conversion partitions pretrained feed-forward network (FFN) neurons into experts and selects a subset for each token. Leading methods fall into two categories: analytic methods produce a converted model quickly using heuristics, while learning methods produce a model much more slowly by optimizing via gradients. Both approaches, however, fix the same active expert count, or "utilization," at every layer. We introduce Spectral-MoE, which learns to allocate a utilization budget heterogeneously across layers. Its name draws a metaphor between separating a dense FFN into experts and separating a light beam into its spectrum. Spectral-MoE consists of two distinct approaches: a static approach that uses a single per-layer scalar which is shared across tokens, and a dynamic approach which emits a per-token schedule. We demonstrate that both allocators improve performance over uniform on both the analytic-based and learning-based SOTA dense-to-MoE methods. At 25% FFN utilization, dynamic allocation reduces WikiText-2 perplexity relative to uniform by 1.04, 2.96, and 2.71 points on the learned base for Llama-2-7B, Llama-3-8B, and Qwen2.5-7B, respectively. The corresponding reductions on the analytic base are 2.45, 6.26, and 2.17 points. On the learned base, dynamic allocation improves six-task mean accuracy over uniform by an average of 1.5, 4.2, and 3.6 percentage points across the three models at FFN utilizations of 12.5%, 25%, and 50%, respectively. These results establish per-layer expert utilization as a novel optimization axis in dense-to-MoE conversion.

cs.LG↗

BEHAVE: Functional Behavior Modeling Enables Self-Improving Agents for Hardware Design and Verification

Developing agents for hardware design and verification requires reliable correctness feedback. As a hardware specification may permit correct implementations with different latencies, matching design and reference outputs cycle by cycle can reject valid designs. To address this, we introduce BEHAVE, an agentic framework for multi-turn joint hardware design and verification through functional behavior modeling. We define Behavior IR to express task functionality as executable behavior models without prescribing implementation timing beyond the specification. The agent iteratively develops a register-transfer-level (RTL) design and a behavior model as the design's verification reference. Our evaluator, BEHAVE-Sim, checks both artifacts separately against a hidden golden behavior model using input stimuli generated by random sampling and solver-guided search. BEHAVE thus supports power, performance, and area (PPA) exploration across task-permitted latencies and microarchitectures. During training, the same evaluator provides verifiable reinforcement learning (RL) rewards from specification-behavior pairs without reference RTL. For self-improvement, the agent continually searches for high-level implementations relevant to its capability gaps, constructs and checks specification-behavior pairs, and trains on the expanded task pool. We release BEHAVE-Train and BEHAVE-Eval with 600 human-reviewed specification-behavior pairs for realistic hardware workloads. Starting from 60 seed tasks and acquiring 100 new tasks, self-improvement raises Qwen3.8-27B's RTL pass@1 on BEHAVE-Eval from 55.0% to 75.0%, reaching performance comparable to RL using a 540-task pool.

cs.AI↗

Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads

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.

cs.AI↗

Academia x Industry: The Role of Fundamentals for Silicon in an AI Native Era

Agentic AI is set to become one of the most transformational technologies in generations and materially change how we approach silicon design and engineering. The impact is being felt in real time amid a rapidly changing landscape, which can make it overwhelming for both silicon practitioners and academics to adapt to the AI native silicon design era. To add structure to how we navigate this transition, we provide a joint view from academia and industry silicon practitioners of the challenges, opportunities, and considerations we expect will catalyze how the community transitions into an AI native silicon future. In particular, we reemphasize the importance of core silicon design fundamentals in academic training and why they have renewed importance in research and industry practice for AI native silicon design. It is our hope that the views provided here will offer valuable and complementary perspectives to those in academia and industry to interpret, inform, and catalyze the transition to the AI native era. We expect that many similar and overlapping views will emerge, but the precise technical details will differ across stakeholders, so it is valuable for the community to amass a diversity of viewpoints.

cs.AR↗

The Hyperscale Lottery: How State-Space Models Have Sacrificed Edge Efficiency

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.

cs.AR↗

SemanticDialect: Semantic-Aware Mixed-Format Quantization for Video Diffusion Transformers

Diffusion Transformers (DiTs) achieve state-of-the-art video generation quality, but their substantial memory and computational footprints hinder edge deployment. Quantization can reduce these costs, yet existing methods often degrade video quality due to high activation variation and the difficulty of preserving semantic and temporal coherence. We propose SemanticDialect, which advances block-wise mixed-format quantization. In this framework, each block selects an optimal format (dialect) from a candidate set (formatbook), which is augmented with lookup tables that store quantization errors and quantized indices, enabling efficient per-block format selection and quantization with minimal online overhead. We further introduce attention-guided activation decomposition, which reduces quantization error via residual quantization, and semantic-aware dialect assignment (SeDA), which reduces cross-token quantization inconsistency by enforcing format uniformity among semantically correlated tokens. Experiments demonstrate that SemanticDialect outperforms prior quantization methods and block-wise formats (MXFP4, NVFP4) while approaching FP16 quality on Open-Sora 2.0. We also validate hardware deployability through RTL design and GPU kernel implementation.

cs.CV↗

P3-LLM: An Integrated NPU-PIM Accelerator for Edge LLM Inference Using Hybrid Numerical Formats

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, the high-precision 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 edge LLM inference. 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. Evaluations on diverse LLMs and tasks demonstrate that P3-LLM achieves higher accuracy than state-of-the-art KV-cache quantization and weight-activation quantization algorithms. Combining the proposed quantization scheme with low-precision PIM architecture co-design, P3-LLM yields an average of $4.9\times$, $2.0\times$, and $3.4\times$ speedups over state-of-the-art LLM accelerators HBM-PIM, Ecco, and Pimba, respectively. Code is available at https://github.com/yc2367/P3-LLM.

cs.AR↗

AgenticCache: Cache-Driven Asynchronous Planning for Embodied AI Agents

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.

cs.LG↗

Heterogeneous Memory Design Exploration for AI Accelerators with a Gain Cell Memory Compiler

As memory increasingly dominates system cost and energy, heterogeneous on-chip memory systems that combine technologies with complementary characteristics are becoming essential. Gain Cell RAM (GCRAM) offers higher density, lower power, and tunable retention, expanding the design space beyond conventional SRAM. To this end, we create an OpenGCRAM compiler supporting both SRAM and GCRAM. It generates macro-level designs and layouts for commercial CMOS processes and characterizes area, delay, and power across user-defined configurations. The tool enables systematic identification of optimal heterogeneous memory configurations for AI tasks under specified performance metrics.

cs.AR↗

LLM-FSM: Scaling Large Language Models for Finite-State Reasoning in RTL Code Generation

Finite-state reasoning, the ability to understand and implement state-dependent behavior, is central to hardware design. In this paper, we present LLM-FSM, a benchmark that evaluates how well large language models (LLMs) can recover finite-state machine (FSM) behavior from natural-language specifications and translate it into correct register transfer-level (RTL) implementations. Unlike prior specification-to-RTL benchmarks that rely on manually constructed examples, LLM-FSM is built through a fully automated pipeline. LLM-FSM first constructs FSM with configurable state counts and constrained transition structures. It then prompts LLMs to express each FSM in a structured YAML format with an application context, and to further convert that YAML into a natural-language (NL) specification. From the same YAML, our pipeline synthesizes the reference RTL and testbench in a correct-by-construction manner. All 1,000 problems are verified using LLM-based and SAT-solver-based checks, with human review on a subset. Our experiments show that even the strongest LLMs exhibit sharply declining accuracy as FSM complexity increases. We further demonstrate that training-time scaling via supervised fine-tuning (SFT) generalizes effectively to out-of-distribution (OOD) tasks, while increasing test-time compute improves reasoning reliability. Finally, LLM-FSM remains extensible by allowing its FSM complexity to scale with future model capabilities.

cs.AI↗

RaZeR: Pushing the Limits of NVFP4 Quantization with Redundant Zero Remapping

The recently introduced NVFP4 format demonstrates remarkable performance and memory benefits for quantized large language model (LLM) inference. However, we observe two types of redundancy in NVFP4 encoding: (1) The FP4 element format naturally exposes an unused quantization value due to its sign-magnitude representation that contains both positive and negative zeros. (2) The FP8 block scaling factor has an unused sign bit because it is always positive. Additionally, we find that LLM weights are more tolerant to a lower-precision block scaling factor. Based on these observations, we propose Redundant Zero Remapping (RaZeR), an enhanced numerical format that pushes the limits of NVFP4 for more accurate LLM quantization under the same memory footprint. RaZeR leverages the redundant bits of the block scaling factor to adaptively remap the redundant FP4 zero to additional quantization values with improved accuracy. To demonstrate the practicality of RaZeR, we design efficient GPU kernels for RaZeR-quantized LLM inference and propose novel hardware to natively support this. Extensive experiments validate RaZeR's superior performance for 4-bit LLM quantization. For example, relative to native NVFP4, RaZeR reduces the average perplexity loss by 34.6% and 31.2% under weight-only and weight-activation quantization, respectively. Code is available at: https://github.com/yc2367/NVFP4-RaZeR.

cs.LG↗

GaussianVision: Vision-Language Alignment from Compressed Image Representations using 2D Gaussian Splatting

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 pre-training (CLIP) to 2DGS by reusing a frozen RGB-based transformer backbone with a lightweight splat-aware input stem and a perceiver resampler, training only 9.7% to 13.8% of the total parameters. On a 12.8M dataset from DataComp, GS encoders yield competitive zero-shot performance on 38 datasets from the CLIP benchmark while compressing inputs 3x to 23.5x relative to pixels. 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.

cs.CV↗

On the Role of Temperature Sampling in Test-Time Scaling

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.

cs.AI↗

AttAnchor: Guiding Cross-Modal Token Alignment in VLMs with Attention Anchors

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.

cs.AI↗

The Future of Memory: Limits and Opportunities

Memory latency, bandwidth, capacity, and energy increasingly limit performance. In this paper, we reconsider proposed system architectures that consist of huge (many-terabyte to petabyte scale) memories shared among large numbers of CPUs. We argue two practical engineering challenges, scaling and signaling, limit such designs. We propose the opposite approach. Rather than create large, shared, homogenous memories, systems explicitly break memory up into smaller slices more tightly coupled with compute elements. Leveraging advances in 2.5D/3D integration, this compute-memory node provisions private local memory, enabling accesses of node-exclusive data through micrometer-scale distances, and dramatically reduced access cost. In-package memory elements support shared state within a processor, providing far better bandwidth and energy-efficiency than DRAM, which is used as main memory for large working sets and cold data. Hardware making memory capacities and distances explicit allows software to efficiently compose this hierarchy, managing data placement and movement.

cs.AR↗

GainSight: A Unified Framework for Data Lifetime Profiling and Heterogeneous Memory Composition

As AI workloads drive increasing memory requirements, domain-specific accelerators need higher-density on-chip memory beyond what current SRAM scaling trends can provide. Simultaneously, the vast amounts of short-lived data in these workloads make SRAM overprovisioned in retention capability. To address this mismatch, we propose a wholesale shift from uniform SRAM arrays to heterogeneous on-chip memory, incorporating denser short-term RAM (StRAM) devices whose limited retention times align with transient data lifetimes. To facilitate this shift, we introduce GainSight, the first comprehensive, open-source framework that aligns dynamic, fine-grained workload lifetime profiles with memory device characteristics to enable generation of optimal StRAM memory compositions. GainSight combines retargetable profiling backends with an architecture-agnostic analytical frontend. The various backends capture cycle-accurate data lifetimes, while the frontend correlates workload patterns with StRAM retention properties to generate optimal memory compositions and project performance. GainSight elevates data lifetime to a first-class design consideration for next-generation AI accelerators, enabling systematic exploitation of data transience for improved on-chip memory density and efficiency. Applying GainSight to MLPerf Inference and PolyBench workloads reveals that 64.3% of first-level GPU cache accesses and 79.01% of systolic array scratchpad accesses exhibit sub-microsecond lifetimes suitable for high-density StRAM, with optimal heterogeneous on-chip memory compositions achieving up to 3x active energy and 4x area reductions compared to uniform SRAM hierarchies. To facilitate adoption and further research, GainSight is open-sourced at https://gainsight.stanford.edu/.

cs.AR↗