paperlist.ai

AI conference papers

AI conference paper library

Search papers from 30+ top conferences, including NeurIPS, ICLR, ICML, CVPR, ACL and AAAI. Filter by keyword, then open any paper for its abstract, authors and links.

arXiv

RamanPFN: learning from Raman spectral structure with a tabular foundation model

Xingyu Pan, Huan Wang, Jinjia Guo, Zhenlin Zhao, Siming Dong, Jixi Lu

Raman spectroscopy enables non-destructive, label-free molecular characterization across materials science, biomedicine and process monitoring. Predictive Raman datasets often contain few labelled spectra and thousands of ordered wavenumbers, with informative variation within bands and across distant spectral regions. Latent-variable chemometrics accommodates collinear small-sample data but can obscure fine peak morphology, whereas deep spectral networks resolve this structure only after task-specific training. TabPFN avoids task-specific parameter fitting through pretrained in-context inference, but processes very wide inputs as feature-subsampled views that do not preserve joint visibility of related bands. We present RamanPFN, a spectral representation framework that encodes these dependencies before TabPFN inference. Global Compositional Unmixing constructs non-negative coordinates over the complete spectrum so that distant bands with shared latent variation occupy a common predictive axis. Local Vibrational Subspace Encoding represents contiguous wavenumber regions with multiple orthogonal modes that retain independent changes in peak shape, intensity and position. The representations are evaluated separately and combined at the prediction level. Evaluation covered 150 tasks from 74 public Raman datasets. RamanPFN reduced root-mean-square error by 19.6% on average across 129 regression targets relative to direct TabPFN inference and further reduced the remaining classification error by 9.0% across 21 classification tasks. These results establish explicit spectral representation as an effective interface between high-dimensional Raman measurements and reusable tabular inference.

arXiv

Cardiovascular Digital Twins from Physics Based to Data Driven Approaches

Emmanuel Lwele, Francis Chikweto

Cardiovascular digital twins aim to create patient-specific computational models that evolve with clinical data to support diagnosis, prognosis, and therapy optimisation. Mechanistic models provide physiological interpretability but remain computationally demanding, whereas data-driven approaches improve scalability yet risk limited robustness. Emerging physics-informed, graph-based, and hybrid methods integrate physical constraints with relational learning across vascular networks. We review modelling paradigms, data assimilation frameworks, validation challenges, and translational pathways toward clinically deployable cardiovascular digital twins.

arXiv

CoRe-GNN: Multilevel Message passing on Coarsened graphs

Antonin Joly, Nicolas Keriven, Aline Roumy

Training Graph Neural Networks on large graphs is challenged by the memory cost of storing all node representations across layers. We show that several existing scalable approaches can be written as structured modifications of the GNN propagation matrix, providing a unified perspective that exposes their respective limitations. In particular, graph coarsening replaces it by a low-rank approximation that enables spectral guarantees but assigns uniform representations to clustered nodes, while Cluster-GCN restricts the propagation matrix to intra-cluster connections that allow efficient batching but sever long-range information. These are complementary failures of the \emph{same} decomposition of the graph into groups of nodes. To obtain the best of both worlds, we propose \textbf{CoRe-GNN}, which performs both propagations in parallel at each layer: a coarsened inter-cluster term capturing long-range structure, and a local intra-cluster term preserving per-node discriminability. We prove that CoRe-GNN inherits analogous approximation guarantees to those of graph coarsening, and introduce a natural cluster-based \emph{batching scheme} that scales to graphs with millions of nodes. On node classification benchmarks spanning homophilic, heterophilic, large-scale, and long-range graphs, CoRe-GNN outperforms both graph coarsening and Cluster-GCN baselines. Notably, CoRe-GNN reaches competitive accuracy on \emph{long-range} tasks, while remaining memory-efficient through batching.

arXiv

Do Static Embeddings Add Value to Hybrid Dutch Retrieval?

António Pereira Barata

Embedding benchmarks measure standalone model quality, but they do not establish whether a low-cost retriever contributes complementary ranking information once lexical and transformer-based retrieval are already combined. We present a controlled evaluation of this question across Dutch retrieval tasks from the Massive Text Embedding Benchmark for Dutch (MTEB-NL). Weighted reciprocal rank fusion (RRF) combines Best Matching 25 (BM25), Qwen/Qwen3-Embedding-0.6B (Qwen), and two multilingual static embedding models. Five datasets comprising 14,500 queries and 786,573 documents are scored exhaustively, and fusion weights are searched on a simplex in increments of 0.1. Ten-fold query-level cross-validation selects weights on nine folds and evaluates them on the held-out fold; paired bootstrap confidence intervals and sign-randomisation tests quantify the resulting differences. Fusion improves over the training-selected individual retriever by 0.061 mean reciprocal rank (MRR) on Dutch News, 0.029 on VABB, 0.004 on WebFAQ NL, and 0.025 on Wikipedia NL, while matching BM25 on Open Tender. All four positive differences remain distinguishable from zero after Holm correction. No unrestricted fold assigns positive weight to either static retriever: all 50 selections lie on the BM25-Qwen edge, and forcing a static contribution reduces effectiveness. Leave-one-dataset-out selection chooses equal BM25-Qwen weighting in every iteration and outperforms the cross-domain-selected individual retriever on every held-out task. The results support a two-retriever lexical-transformer architecture as a robust tested default across the evaluated Dutch tasks and show that standalone benchmark performance is insufficient to establish marginal value in hybrid retrieval.

arXiv

From Information to Delegation: Mapping Human-AI Financial Decision Making

Iman Munire Bilal, Yingcan Carol Wang, Ajan Raj, Filippo Giovagnini, Pranav Tewari, Yuwei Zhang, Mei-Chen Zoe Liou, Qamar Zaman

As AI increasingly participates in human decision making, understanding how decision-making authority is distributed between humans and AI has become a fundamental behavioural question. We introduce a behavioural measurement framework combining intent and delegated decision authority to quantify what consumers seek from AI and how much decision-making authority they assign to it. Applied to 1.5 million real-world ChatGPT and Gemini interactions from 6,304 users in the United States and India, we find that financial services are already a substantial AI use case. Consumers overwhelmingly use AI to retrieve information and shape financial judgement, while delegation of financial execution remains rare. By shifting attention from conversation topics to delegated decision authority, this work establishes a behavioural baseline for measuring the transition to increasingly agentic AI.

arXiv

One QK Channel, Many Sources: Guarding Low-Precision Attention Collapse

Shuxiao Xie, Shuyang Xie, Yuan Cao, Dezhi Ran, Wei Yang, Tao Xie

A bfloat16 transformer can train normally for many steps and then collapse abruptly. Distinct low-precision errors can trigger the same failure, leaving unclear whether each source needs its own repair or one shared route can be blocked. We isolate a reproduced GPT-2-class collapse to the streaming-softmax accumulator, where fp32 accumulation repairs it, and use the fault as an assay for moving controlled errors across sources. Errors placed outside attention still drive the same query-key (QK) spectral runaway, while correcting only QK keeps training stable with the source fault active. This source-channel dissociation shows that fault source is not failure channel. It holds across the tested architectures and scales and reproduces on a second GPU architecture. A causal probe projects each update off the current QK weights' leading three singular directions: the query projection's largest singular value stays at 11.1, whereas removing equal energy elsewhere leaves it at 237. The QK channel therefore drives the early runaway rather than merely tracking it. Entry depends on temporal sign-coherence across steps, not aggregate deviation. QK-Guard closes the channel with a dormant controller that switches on parameter-free QK normalization when attention-logit saturation begins. It contains every tested runaway and matches always-on QK normalization over 60k steps, while non-QK actions at the same trigger fail. The results support intervention at the shared QK locus rather than separate repair at each fault source.

arXiv

How Much Does a Reasoning Summary Reveal? An Observability Ladder for Large Language Models

Andres Algaba, Francesca Carlon, Lynn Delcon, Marthe Ballon, Bert Verbruggen, Vincent Ginis

Large language models often show users a final response and a short reasoning summary while the full reasoning trace stays hidden. We introduce an observability ladder that holds each completed run fixed and varies only what a reader inspects to judge whether the answer is correct: the response, a self-summary the model writes from the trace, the trace itself, and internal signals, each with and without the prompt. Across three benchmarks and five open-weight Qwen3 and gpt-oss models, we train matched linear correctness predictors on each access level. Without the prompt, summaries carry most of the trace's ranking signal (mean AUROC 0.774 versus 0.813) and add +0.156 over the response alone. With the prompt visible, the summary's gain collapses to +0.019, while the trace still adds +0.041. Even at equal length, the trace's last words predict correctness as well as summaries, or slightly better, and carry denser and more discriminative uncertainty and self-correction cues. On MMLU-Pro questions with both correct and incorrect runs, linear summary readers are near chance and trace readers retain only modest signal, both with and without the prompt (prompt-withheld AUROC 0.503-0.545 versus 0.544-0.590). With the prompt withheld, a GPT-5-mini reader recovers substantially more signal from both summaries and traces on gpt-oss-20b, and even then the trace keeps a small +0.034 advantage. Much of the linear readers' trace signal is associated with length. In the common case where users already hold the prompt, summaries are less helpful than the full trace for monitoring correctness. Monitorability is thus a joint property of the display and the reader, so any monitorability claim, including for faithfulness, should specify both.

arXiv

An AI-Based Decision-Support Pipeline for Day-Ahead Photovoltaic Forecasting

Fariba Dehghan, Sebastian Stein, Vahid Yazdanpanah, Stephanie Gauthier, Masood Nazari

Reliable photovoltaic (PV) forecasts are needed for low-carbon energy systems, but newly deployed sites often have short, imperfect records. This makes standard day-ahead forecasting difficult: persistence and physical baselines can be sensitive to calibration and timestamp alignment, while single machine-learning models may capture only one structure in the data and overstate skill under non-temporal validation. We study this problem at a United Kingdom charging-station site, where PV forecast errors affect charging availability, storage scheduling, and downstream control. Using measured inverter output and publicly available meteorological inputs, we develop a deployment-oriented environmental-AI pipeline for day-ahead hourly PV forecasting. The pipeline corrects timestamp conventions, constructs leakage-safe solar-geometry and clearness-index features, adds short-term atmospheric context, and combines complementary predictors through validation-learned stacking. Against smart persistence, a clear-sky baseline that adjusts recent PV output using expected clear-sky irradiance, the best ensemble reduces daylight normalised RMSE by about 32% under random day-blocked evaluation and 9% under the stricter rolling-origin protocol. It also reduces daylight RMSE relative to the strongest individual machine-learning baseline by 6.6% and 6.4%, respectively. The results show that physics-aware stacking can support PV forecasts from limited site data, but its value depends on model class, evaluation protocol, and deployment context.

arXiv

Pretraining on Call Graphs: When Binary Analysis Tasks Profit From Context

Samuel Valenzuela, Johannes Kinder

Binary function embedding models are trained to encode the semantics of binary code in such a way that they can be generalized to a variety of reverse engineering tasks, such as binary code search, vulnerability detection, or malware classification. While many models only take the function in question as contextual input, there have been successful attempts to improve function embeddings by leveraging information from the call graph. In this study, we dissect the implications of these embedding refinements. We conduct experiments using a range of graph-based models on the embeddings generated by two state-of-the-art binary function embedding models. Integrating inter-procedural context, we show that improvements on binary code similarity detection (BCSD) will not necessarily generalize to downstream tasks, neither of semantic nor of syntactic nature. More generally, we find that optimizing for semantic similarity tasks correlates with worse performance on syntactic tasks. By conducting an explanatory analysis on the dataset, we find that the call graph-based enhancements significantly enhance the robustness of embeddings, particularly in scenarios where the initial models struggle. Furthermore, we observe that the added context is more beneficial for namespace-related functions than for those focused on individual logic, confirming that the call graph can be leveraged most effectively in context-dependent scenarios.

arXiv

Isotonic Bradley-Terry Model for Paired Comparison Data

Ryoya Yamasaki

In this paper, we study prediction problems for paired comparison data, for example, predicting the win probability between two unmatched players and ranking all the players according to the order of their strengths by using win probability data between two matched players. Paired comparison data are typically analyzed using Bradley-Terry and Thurstone-Mosteller models. These models predict the win probability by transforming the difference between learned rate parameters, which represent players'\;strengths, with a pre-specified inverse link function, and employ the order of learned rate parameters for player ranking. However, these models may suffer from model misspecification owing to the selection of a fixed inverse link function. Therefore, in this study, we propose to learn the rate parameters by a (sub-)gradient method and the inverse link function by an isotonic regression technique alternately. The proposed model guarantees monotonic improvement in training error, and is likely to yield an exact tie when the available data is insufficient to establish a strict ranking. We also verified that the proposed model could improve the win probability prediction and ranking performance through numerical experiments with synthetic data and real-world data of football Premier League, baseball MLB, and tennis ATP tour.

arXiv

Accelerating Evolutionary Strategy via Rao-Blackwellizing Realization of Uncertain Input

So Nakashima, Tetsuya J. Kobayashi

We investigate Optimization under Input Uncertainty (OIU), in which the input to the objective function, rather than the objective function itself, is subject to uncertainty. OIU appears in manufacturing processes with production tolerance, control of physical systems with actuation noise, Mixture of Experts, and Reinforcement Learning (RL). Most of the existing approaches solve OIU by using the value of the objective function but discard the information of the realized input, even though the realized input is observable in various applications. The question here is whether the discarded information of the realized input is useful to accelerate the optimization process. We affirmatively answer this question for Evolutionary Strategy (ES) by theoretically showing that the information of the realized input can reduce the variance of the gradient estimator via Rao-Blackwellization. Using the Rao-Blackwellized gradient estimator, we propose Phenotype-Accelerated Evolutionary Strategy (PAES), which is a refinement of ES for OIU. Numerical experiments show that PAES converges faster than the usual ES from simple continuous optimization problems to RL benchmarks.

arXiv

Feed-Forward Steering in Transformer Residual Dynamics

Timur Mudarisov, Mikhail Burtsev, Radu State

Attention-only dynamical theories model Transformer residual directions as particles aggregating on a sphere. We extend this framework by incorporating the feed-forward network (FFN) term as a local steering field acting on each token state. The resulting theory predicts that the tangential component of the FFN field is necessary for motion in residual-direction space, that critical residual directions correspond to nonlinear projective equilibria, and that a commutator defect determines when a finite attention--FFN block can be accurately approximated by a parallel, additive flow. Across GPT-2, Pythia, Mistral, and Llama models, the extended theory improves one-step angular prediction relative to an attention-only baseline, with the contribution of the FFN increasing from GPT-2 to Llama-3-8B. Intervention experiments show that retaining only the tangential FFN component preserves most model quality, whereas retaining only the radial component causes performance to collapse. The tangential component also preserves output diversity under aggregation pressure. As a practical application, layers with small commutator defects can be approximately parallelized with only a modest increase in loss, whereas layers with large defects degrade rapidly. These findings support the interpretation of FFN layers as directional steering fields that shape Transformer residual geometry and govern the feasibility of block-level interventions.

arXiv

Open-DiffLoco: Open-Source Differentiable Learning for Deployable Blind Quadruped Locomotion

Martin Opat

Developing deployable locomotion policies through conventional reinforcement learning often requires complex reward engineering and expensive training times. While differentiable simulation offers a highly efficient alternative, open-source tools capable of end-to-end transfer of these policies to physical hardware remain limited. This paper introduces Open-DiffLoco, an open-source framework for training deployable blind quadruped locomotion policies with differentiable simulation. The framework implements the Short-Horizon Actor-Critic (SHAC) algorithm in MuJoCo XLA (MJX) and trains a proprioceptive policy that transfers to real-world hardware. The deployed policy removes privileged actor observations, including base linear velocity, and does not rely on reference trajectories. It also uses a substantially simplified reward function, enabling the robot to discover walking patterns without the complex auxiliary rewards typically used in conventional reinforcement learning pipelines. When deployed on physical hardware (a Unitree Go2 quadruped), the trained policy tracks omnidirectional velocity commands with root-mean-square error below 0.2 m/s, reaches speeds above 1 m/s, and remains robust to uneven terrain and external physical disturbances, such as lateral pushes. Across the reported configurations, training uses under 6 GB of VRAM on a single NVIDIA GeForce RTX 5080 GPU and completes in approximately 20-60 minutes. As an algorithmic extension to SHAC, we propose Jacobian-Augmented Value Estimation (JAVE), which supervises the critic Jacobians to improve early first-order policy-gradient training. To our knowledge, Open-DiffLoco is the first open-source framework for training deployable locomotion policies using differentiable simulation. Deployment videos and source code are available at: https://diffloco.martin-opat.com/

arXiv

A Comparative Analysis of MLP and Kolmogorov-Arnold Networks (KAN) for Faster-than-Nyquist (FTN) Signaling Detection

Sude Ertan, Osman Tokluoglu, Enver Cavus

Faster-than-Nyquist signaling improves spectral ef- ficiency by deliberately introducing inter-symbol interference. Classical sequence detectors such as BCJR can approach optimal performance, but their computational cost grows rapidly with channel memory. This paper investigates data-driven FTN BPSK detection under AWGN through a direct comparison between multilayer perceptrons and Kolmogorov Arnold Networks. A large-scale Monte Carlo dataset containing nearly four million labeled windows is generated for a time-packing factor of zero point eight and signal-to-noise ratio values from seven to ten decibels. The best MLP obtained from width sweeping uses hidden width thirty two, whereas the selected KAN uses hidden width four with spline grid size five. At ten decibels, the MLP produces a bit error rate of one point three times ten to the minus four, while the KAN reaches seven times ten to the minus six. This corresponds to an eighteen point six times lower bit error rate while using only one eighth of the MLP hidden width. The results show that KAN provides a more effective and more parameter-efficient neural decision model than the MLP baseline for FTN BPSK detection.

arXiv

SCOPE: Entanglement Frontier Escape for Source-Free Class Unlearning

Junhao Cai, Dohun kim, Sung Il Choi, Juhyun Park, Chengjun Jin, Dowon Kim, Changhee Joo

Source-free class unlearning erases whole classes using only the forget data, judged at the representation level, where features can leak a class the head no longer predicts. Existing feature-space erasers answer with one fixed projection, yet forget and retain classes share a representation, so deleting one disturbs the other where they overlap. We prove this tension is a frontier. Every fixed projection that deletes pays a retain cost of at least the retain-readout energy along the forget-discriminant subspace, and erasing that subspace alone attains the floor. The leading source-free erasers all instantiate the form it binds, so the frontier limits the whole class. Conditioning the erasure on the input escapes it. Spectral Conditional Projective Erasure (SCOPE) does so with a single gate, suppressing the forget subspace chiefly on inputs its frozen head's weight scores read as a forget class. It is closed form, needs no retain data or gradient training, and costs orders of magnitude less than retraining. Across five object, face, and speaker benchmarks spanning two modalities and both convolutional and transformer backbones, the frontier predicts the measured retain cost. SCOPE leads the source-free erasers on every benchmark and forget-set size, and at the hardest setting it tops every unlearner, trained methods included.

arXiv

Secrets Everywhere: Auditing Memorization in Mobility Prediction Models

Anne Josiane Kouam, Hristo Boyadzhiev, Konrad Rieck

Human mobility prediction models, which forecast the next location in a user's trajectory, are increasingly deployed in urban analytics, navigation, and personalized services. Yet, little is known about their potential to memorize and expose sensitive user trajectories from training data. While memorization has been extensively studied in language models, mobility prediction poses unique challenges: training sequences encode human behavior at various spatial and temporal scales, creating privacy risks at different granularities. In this paper, we conduct the first systematic audit of memorization in mobility prediction models. While prior work has shown that privacy leaks can arise from such models, we systematically assess and quantify memorization risks at scale. We identify key challenges, including the lack of a randomness space, the multi-scale structure of trajectories, and user-specific behavioral diversity. To address these challenges, we introduce a framework to quantify mobility memorization at different levels of granularity: individual locations, anchor pairs, and subtrajectory segments. We also develop user-grounded reference sets to assess how likely a model is to prefer training data over realistic alternatives. Our evaluation across multiple models and datasets reveals pervasive memorization patterns that correlate with user regularity and increase the risk of data extraction at inference time. Our findings call for mandatory privacy auditing in mobility prediction models.

arXiv

Adaptive Reconstruction of Bosonic Quantum States

Vasilisa Usova, Phila Rembold, Ian Yang, Marco Rossignolo, Simone Montangero, Samuele Tosatto, Gerhard Kirchmair

Bosonic quantum systems provide a hardware-efficient platform for quantum information processing but remain challenging to characterise due to their large Hilbert space and the high measurement cost of state tomography. Existing approaches estimate the fidelity with respect to a single target state, making them unsuitable for applications in which physically equivalent states differ by phase space translations, rotations, or other transformations. Here, we introduce an adaptive reconstruction technique that estimates the fidelity with respect to a family of bosonic states while reconstructing the underlying Wigner function from a small number of measurements. The method combines a physics-informed parametric model with Bayesian inference, bootstrap, and active learning to iteratively select the most informative phase space sampling points. We implement the approach on a circuit quantum electrodynamics platform and benchmark it on Schrödinger cat states with amplitudes $α\in[1,3]$. The reconstruction yields reproducible fidelity estimates within a few minutes, remains robust to substantial displacements and rotations in phase space despite using a mismatched prior, and is sensitive to subtle state imperfections. We further compare the adaptive strategy with existing Wigner function sampling protocols experimentally, demonstrating the advantage of adaptive sampling for measurement-efficient fidelity estimation with respect to a family of cat states. Finally, we incorporate the reconstructed fidelity into the figure of merit used in a proof-of-principle closed-loop quantum optimal control experiment, demonstrating the applicability of the method to autonomous optimisation of bosonic quantum states.

arXiv

Convex Neural Energy Elements: Monolithic Finite-Element Assembly of Geometry-Parameterized Neural Operators with Stability and Error Guarantees

Hongyue Jiang, Jianjiang Zhan, Chenzhuo Zhang, Fan Wang

Extending the neural-operator element method from individually trained, fixed-geometry neural elements to a library of reusable, geometry-parameterized element types fails structurally: a field-predicting operator trained by value regression induces an energy whose assembled Hessian is indefinite, and Newton converges to spurious minima (247% error) even with 1%-accurate field predictions. We introduce convex neural energy elements: each element exports a scalar energy E(g,U), architecturally convex in its boundary degrees of freedom U and smoothly parameterized by its geometry g, realized as a hypernetwork-generated positive-semidefinite quadratic form (an input-convex correction is reserved for non-quadratic physics). A regularization-nullspace principle--the regularizer's nullspace must contain the physics nullspace--removes an otherwise irreducible bias, and assembled elements inherit the classical guarantee that singular element stiffnesses yield a positive-definite global system. We prove conditional error bounds (energy-to-solution accuracy, element-count scaling, geometry generalization) and verify each experimentally. On heat conduction with elliptic holes, one trained element assembles into 2x2 to 8x8 grids and an L-shaped layout of unseen geometries at 0.6-1.0% relative L2 error, with 175x faster per-geometry setup for boundary-quantity workloads. A second trained element type mixes freely with the first in one monolithic assembly, and a three-dimensional instantiation reaches 0.23% on eight-element assemblies--the guarantees are type- and dimension-agnostic. A plane-strain elasticity element, whose physics nullspace is three-dimensional, lands on the analytically predicted regularization floors. Making the energy the learned object turns neural operators from single-use surrogates into reusable elements that inherit the assembly guarantees of the method they extend.

arXiv

Upper-Expectile Multi-Step Q-Learning for Off-Policy Reinforcement Learning

Abdelghani Ghanem, Mounir Ghogho

Multi-step returns accelerate reward propagation in off-policy reinforcement learning, but couple the evaluation of each decision to the suboptimal logged actions that follow it, inducing a pessimistic bias that grows with the horizon. We propose Expectile $n$-step Q-learning (ENQ), which replaces the symmetric $n$-step temporal-difference (TD) loss with an asymmetric expectile loss on the action-value error, with expectile level $τ$ as the only method-specific hyperparameter added beyond $n$-step TD. We prove that the ENQ operator is a $γ^{n}$-contraction. Under deterministic dynamics, at $τ=1$, its bias vanishes at the optimal action-value function $Q^*$ on covered in-support pairs, and the corresponding fixed point satisfies the separation-$n$ instance and its multiples of the lower-bound inequality used by Long-Horizon Q-learning (LQL). Under stochastic dynamics, the operator bias admits two-sided bounds with horizon-independent noise constants. Using a single expectile level $τ=0.8$ and a fixed backup horizon across 27 manipulation and navigation task instances, ENQ is competitive with LQL on aggregate, achieves higher measured training-step throughput in our profiling study, and benefits more from a ten-critic ensemble in a controlled scaling experiment.

arXiv

DART: Decoded Attention over Recurrent States for Efficient Long-Context Sequence Modeling

Yixiao Qian, Song Chen, Pengkai Wang, Jiaxu Liu, Shengze Cai, Chao Xu

Modern language models are built primarily from Transformers, recurrent models, and their hybrid architectures. Transformers rely on token-level attention memories, while recurrent models such as state space models (SSMs) and linear attention maintain compact recurrent states. These architectures are typically instantiated separately or interleaved at the layer level, leaving open whether a shared memory representation can support both recurrent compression and attention-style retrieval. We study this question through the state space duality (SSD) view of Mamba-2, where the SSM state can be interpreted as a compressed associative key--value (KV) cache. We observe that Mamba-2 decodes token-conditioned values from this state but does not decode token-conditioned keys. Based on this observation, we propose DART (Decoded Attention over Recurrent sTates), which retains the chunk state contributions produced by the Mamba-2 chunked scan as chunk state memories, decodes token-conditioned keys and values from these memories, and performs state-memory attention (SMA) over the resulting KV pairs. The retrieved output is then combined with the native Mamba-2 output through a gated residual connection. DART supports practical training by reusing the Mamba-2 chunked scan and implementing SMA as a FlashAttention-style computation. Our analysis and experiments show that DART substantially reduces the length-dependent inference cache compared with a matched attention baseline (e.g., $75\%$ savings when the chunk size is $S=256$ and the state size is $N=128$). Compared with Mamba-2, DART substantially improves associative recall and retrieval while preserving general language-modeling quality.

Keep browsing

More papers from top AI conferences

30+ top conferences, by conference, year, and topic.