Philipp Nazari, Neehal Tumma, Daniela Rus, T. Konstantin Rusch
Linear recurrences are becoming an increasingly common building block in large language models, replacing the growing KV cache of softmax attention with a fixed-size recurrent state that enables efficient decoding. However, the quantization stack has largely been developed around softmax Transformers, whose numerical properties and inference state differ from those of recurrent models. Our goal is to understand these differences and study the design space of quantization methods for recurrent models. In our analysis, we differentiate between two targets. First, extra-state quantization concerns weights and activations outside the recurrent state, whose distributions are shaped indirectly by the sequence mixer during training. In controlled 1.3B parameter models, we find that decayed recurrences develop weaker residual outlier channels than softmax attention and are therefore easier to quantize without incoherence processing. Second, intra-state quantization concerns the recurrent state itself during decode, where repeated updates create a different numerical problem from KV cache quantization. We instantiate our findings as LoReS, a state quantization pipeline combining value-axis scale grouping to match output channel magnitude, decay-based mixed-precision to protect slowly decaying heads or channels, and stochastic rounding to limit error accumulation under repeated requantization. Across Qwen-3.5 9B, Nemotron-H 8B, and Kimi-Linear 48B-A3B, LoReS supports robust 8-bit state quantization and lower effective precision on shorter decode horizons. We implement LoReS in vLLM and demonstrate end-to-end decode speedups.
Neehal Tumma, Philipp Nazari, Rohan Cherukuri, Daniela Rus, T. Konstantin Rusch
Linear recurrences are emerging as practical alternatives to softmax attention in large language models, appearing in both standalone and hybrid architectures. However, despite their rapid emergence, low-precision inference remains much less developed for these models than for softmax attention. During prefill, chunkwise-parallel algorithms cast much of the recurrence as dense matrix multiplications, suggesting a natural path to low-precision execution. Simply quantizing these operations in their native form, however, can be fragile. In this work, we identify functional invariances that preserve the full-precision computation while making the operands more amenable to quantization. Building on this observation, we introduce LinQuant, a framework for low-precision recurrent prefill based on seven such invariances, including rotations, diagonal rescalings, and alternative placements of recurrent decay and write factors. We derive when these transformations remain valid across common recurrence families spanning linear, gated, and delta-rule updates, and show that they provide complementary degrees of freedom for improving low-precision accuracy. Finally, we translate this framework to production by augmenting SSDi8, an existing low-precision prefill pipeline for Mamba-2, with the applicable invariances. LinQuant improves perplexity at matched latency on Mamba-2 1.3B, 2.7B, and 8B and Nemotron-H 8B, demonstrating the practical utility of invariance-based quantization for recurrent models.
Noel Loo, Neehal Tumma, T. Konstantin Rusch, Mathias Lechner, Alexander Amini, Daniela Rus
Test-time training (TTT) layers replace the linear fast-weight memory of modern recurrent architectures with a small neural network trained online by gradient descent. This nonlinear memory is more expressive than the linear memories of Gated DeltaNet and Mamba-2, but is difficult to parallelize: existing implementations such as TTT-MLP rely on sequential, block-wise weight updates that run several times slower than chunkwise-parallel linear recurrences. Moreover, their block-wise execution only approximates the token-by-token online recurrence, whereas linear recurrences admit an exact chunkwise-parallel evaluation. We introduce ParaTTT (PTTT), an algorithm for parallelizing arbitrarily deep nonlinear memories by iteratively refining estimates of the inner-loop gradient trajectory. Starting from a guess of the hidden-layer gradients, each refinement consists of three chunkwise-parallel passes: a forward pass that evaluates the resulting feature trajectory, an exact gated-delta-rule pass that computes the readout trajectory and its feature gradients, and a backward pass that propagates these gradients into a new estimate. The exact online gradient descent recurrence is a fixed point of the iteration, while decoding requires no refinement iterations. We prove geometric convergence for a two-layer variant under sufficient contraction conditions. In practice, two iterations suffice. With two refinement steps at 4K context, PTTT evaluates the inner-loop trajectory $32\times$ faster than TTT-MLP with the block approximation at matched error, while the full 340M model trains $4.9\times$ faster. Models trained at 340M and 1.3B parameters match or exceed Gated DeltaNet and Mamba-2 and remain competitive with same-size Transformers on language-modeling perplexity and zero-shot accuracy, with better recall than the linear recurrences.
Tudor Braicu, Joseph DelPreto, Jared Boyer, Mónika Farsang, Maximilian Stölzle, Tim Niklas Seyde, Daniela Rus
Agile robots require spatiotemporal intelligence to interpret scenes, reason about interactions, and anticipate future events. Yet it remains unclear how well foundation vision models support these capabilities, particularly from a moving egocentric camera, which is a natural view for robots. To investigate this, we collect an annotated tennis dataset and train a suite of vision models on spatiotemporal tasks. Our dataset contains egocentric video synchronized with five allocentric videos, estimated global positions, labeled contact times, and annotated stroke types. We then evaluate seven prediction tasks from egocentric images: 2D ego player and 3D ball localization, future player and ball trajectories, forehand or backhand stroke anticipation, time to the next ego player ball contact, and 3D ball position at that contact. We compare image encoders, video encoders, and vision-language models with jointly trained prediction heads. The best performing models achieve mean current-frame localization errors of 0.14 m for the ego player and 0.54 m for the ball, future forecasting errors of 0.34 m for the ego player and 0.91 m for the ball over 1.1 s, and a mean contact-time prediction error of 0.14 s. Auxiliary depth information substantially improves 3D ball localization, suggesting that inferring metric depth remains a challenge. Longer image histories do not improve ball forecasting, indicating that representations exploit scene context rather than rely solely on inter-frame motion. Finally, we propose a conditional diffusion model to generate whole-body kinematic tennis motion from vision encoder predictions on egocentric video at real-time rates, adapting movements as predictions are refined. These results move towards a deeper understanding of how foundation vision models can support real-time dynamic robots using egocentric video.
MEM1: Learning to Synergize Memory and Reasoning for Efficient Long-Horizon Agents
Zijian Zhou, Ao Qu, Zhaoxuan Wu, Sunghwan Kim, Alok Prakash, Daniela Rus, Jinhua Zhao, Bryan Kian Hsiang Low 等 9 位作者
AI 速览：MEM1 是一个端到端强化学习框架，让语言智能体在长多轮任务中以恒定内存运行，通过紧凑共享状态协同记忆巩固与推理。
Modern language agents must operate over long-horizon, multi-turn interactions, where they retrieve external information, adapt to observations, and answer interdependent queries. Yet, most LLM systems rely on full-context prompting, appending all past turns regardless of their relevance. This leads to unbounded memory growth, increased computational costs, and degraded reasoning performance on out-of-distribution input lengths. We introduce MEM1, an end-to-end reinforcement learning framework that enables agents to operate with constant memory across long multi-turn tasks. At each turn, MEM1 updates a compact shared internal state that jointly supports memory consolidation and reasoning. This state integrates prior memory with new observations from the environment while strategically discarding irrelevant or redundant information. To support training in more realistic and compositional settings, we propose a simple yet effective and scalable approach to constructing multi-turn environments by composing existing datasets into arbitrarily complex task sequences. Experiments across three domains, including internal retrieval QA, open-domain web QA, and multi-turn web shopping, show that MEM1-7B improves performance by 3.5x while reducing memory usage by 3.7x compared to Qwen2.5-14B-Instruct on a 16-objective multi-hop QA task, and generalizes beyond the training horizon. Our results demonstrate the promise of reasoning-driven memory consolidation as a scalable alternative to existing solutions for training long-horizon interactive agents, where both efficiency and performance are optimized.