Science & Mathematics

Multi-agent LLM systems can formalize natural-language argumentation end-to-end, turning plain text into structured graphs by classifying components, connecting them with relations, and enriching the resulting graphs into formal models. Quality assurance for those graphs is almost always another LLMs: a reviewer, judge, or critic agent. On the other side, in adjacent neuro-symbolic fields, such generated output is often verified symbolically by external checks and solvers. In this work, we argue that in argument graph verification, the allocation should be similar, and based on what each instrument can guarantee. We propose layered verification: starting from deterministic checks, moving to solver-based and LLM reviews, and utilizing debate only under uncertainty, leaving human judgment for what no other instrument can decide. We ground our principle with a study on human-annotated graphs with injected structural defects. An LLM reviewer detects them at near-ceiling recall, including cycles in the support relation, but it objects to almost one in a ten correct graphs, even though it should not, happens to be inconsistent in verdict on identical inputs and is significantly slower than the equivalent exact check. It shows that while LLM reviewer does not lack the ability to detect structural defects, it is worth having a deterministic verification gates.

Quantum and physics-inspired solvers provide specialized sampling interfaces for Ising and quadratic unconstrained binary optimization models, but constrained optimization requires broader coordination and certification. This survey uses classical decomposition to analyze how such solvers can be integrated into hybrid quantum-classical algorithms. We distinguish formal decomposition, reformulation and reduction, and computational orchestration. For each interface, we ask what the transformation preserves, which assumptions make the outer method converge, and which certificates the implemented inner routine supplies. The central questions are whether that information meets the classical algorithm's requirements and whether hybridization improves solution quality or total cost enough to justify its coordination overhead. Benders decomposition, column generation, graph reduction, and copositive methods illustrate this distinction by separating heuristic proposals from classical certification. We also distinguish component effects within a fixed wrapper, reduction--solver interactions, and comparisons between complete quantum-enabled and classical systems. The analyzed studies provide conditional component benefits and favorable co-design evidence, but do not establish transferable end-to-end superiority over strong classical comparators. The resulting framework identifies opportunities for constraint programming and operations research in structure-preserving design, safeguarded hybrid algorithms, and controlled benchmarking.

Dual-polarization free-space quantum communication receivers produce two coupled complex heterodyne streams. We develop robust quaternion receiver methods for such links under fading, polarization mixing, calibration mismatch, impulsive contamination, and other non-Gaussian disturbances. The four polarization and quadrature components are encoded as quaternion-valued samples. We formulate an identifiable effective channel model, Quaternion-Valued Projection Statistics (QVPS), and Adaptive QVPS (AQVPS), which augments samplewise projection outlyingness with robust local second-order descriptors. Robust complex and quaternion channel estimators use radial Huber weighting, while QVPS and AQVPS are evaluated separately as disturbance diagnostics. Neural experiments compare matched real-, complex-, quaternion-, and widely linear complex-valued receivers. Robust estimation reduces sensitivity to non-Gaussian contamination. Local second-order descriptors improve detection of the tested localized disturbances relative to QVPS, and a descriptor-only ablation slightly outperforms full AQVPS. Complex LMMSE gives the lowest evaluation MSE in the clean Gaussian SNR sweep, whereas the Quaternion NN, among the compared receivers with their nominal activations, gives the lowest mean evaluation MSE in the tested contaminated sweeps. Activation-matched experiments indicate that much of the nominal Quaternion NN robustness is attributable to bounded activation rather than quaternion parameterization alone. The methods are evaluated at the receiver processing level and do not constitute a composable CV-QKD security proof.

In joint-embedding predictive architectures (JEPA), an encoder and a predictor are trained together to produce a representation that minimizes the error in predicting the representation of the next observation. Because the representations of both the current and the next observation are learned, it is unclear which features they will include. Earlier characterizations describe the features JEPA selects in terms of their individual predictability, raising the concern that unpredictable features, like a traffic light, will not be encoded into the representation. However, the prediction error is also affected by the relationships between features, e.g., the light's color helps predict whether the traffic will stop. We propose the predictive closure characterization, which accounts for these relationships. Based on a reformulation of the JEPA training objective, we show that selection depends on properties of feature sets. This view agrees with earlier works when features evolve independently, but makes new predictions when they do not, for instance that a feature like the traffic light can still be retained. We test this characterization in controlled experiments, which show that unpredictable features can become better represented when they help predict other features, and that changing one feature's predictive relationships can affect how well others are represented. We also demonstrate implications for steering feature selection: targeting one feature with an auxiliary objective, or supplying it to the predictor as side information, can strengthen the representation of another feature whose future values it helps predict, so steering can be indirect.

We study trail-weighted pruning of evidence for Bayesian-network prediction. A traversal enumerates the active trails from each evidence node to the outcome, scores each trail by the product of its edge strengths, and drops evidence whose score falls below a threshold. For collider-free networks (chains and forks) over binary variables, we show that the trail score equals the exact correlation between a trail's endpoints, that a chain and a fork with equal edge strengths have identical joint distributions, and that the pruned posterior is within a computable sum of per-trail terms of the exact posterior when each dropped trail lies in its own branch of the outcome node. We prove this bound, check it exhaustively on all small trees, and show that it fails when dropped trails share a branch. In simulations on random trees, thresholding trail scores gives 1.4 to 13 times lower error than random pruning of the same size. We also apply the score, as an unproven heuristic, to seven bnlearn networks with colliders and loops. Keeping the top half of the active evidence by trail score gives lower error than random pruning on all seven, but it is not clearly better than a mutual-information ranking, and scoring costs more than the inference it saves on six of the seven, so we show no speedup. Models trained on trail, graph and information features and tested on networks they never saw do not beat the trail score or mutual information, but trail features carry most of what they use and clearly outperform generic graph features. A bound for general networks is left to future work.

Reflective prompt optimization improves instructions using task feedback, but coupling candidate generation with full-validation-based selection can make search costly and sequential. We propose a two-stage framework that decouples reflective candidate-pool construction from validation allocation. During construction, sampled parents first answer training questions, and groups of their saved responses guide parallel reflection and revision. Each child is screened on its group's questions against the corresponding parent scores. Improving, nonduplicate prompts enter the pool and can become parents without full validation. After construction, an exploration policy selects candidates from the frozen pool under a fixed full-validation budget, and the prompt with the highest measured validation score is returned. This separation allows candidate-pool size and validation budget to be set independently. We evaluate random selection and a multilayer perceptron with similarity-based exploration (MLP + similarity) on five benchmarks using Qwen3.5-9B and Llama-3.1-8B-Instruct. With 100 retained candidates and 20 full validations, our framework improves over the initial solver by up to 11.22 percentage points. MLP + similarity improves over the initial solver in all ten comparisons, although neither selection policy consistently outperforms the other. These results support separating reflective search from budgeted validation while leaving room for stronger candidate-selection policies.

Span-level hallucination detection pinpoints unsupported content in LLM outputs, enabling actionable feedback and targeted mitigation. Existing evaluations for span-level hallucination detection methods are fragmented across tasks, annotation formats, and metrics, and most span-level datasets cover only a single NLG domain, making it difficult to assess whether a detection method actually generalizes across tasks and domains. In this paper, we introduce SHAM-Bench (Span-level Hallucination Benchmark Across Multiple Tasks), a unified benchmark for span-level hallucination detection spanning six NLG tasks, aggregated from existing human-annotated datasets totaling 6,461 examples under a common schema and evaluation protocol. Using SHAM-Bench, we evaluate existing hallucination detection methods and find that fine-tuned detectors often perform well in-domain but degrade under distribution shift, whereas prompt-based methods with reasoning are more robust across tasks. Our analysis further shows that subtle unsupported content remains difficult to detect and span-level evaluation is still challenged by ambiguous span boundaries and noisy labels. We release the dataset, evaluation code, and a public leaderboard to support future research on broadly generalizable hallucination detection.

Normalizing Flows (NFs) have shown strong results in generative modeling, recently reaching state-of-the-art image generation performance. However, their repetitive, sequential operations limit the computational and memory efficiency achievable with GPUs and TPUs. We propose M3D-FLOW, a novel monolithic 3D-stackable (M3D) DRAM-based processing-near-memory (PNM) architecture that eliminates memory bottlenecks using a Cu-Cu hybrid-bonded logic die integrated with an M3D DRAM die. The M3D DRAM provides dense bank integration and a tiered organization with variable access latencies that M3D-FLOW exploits for the efficient processing of different NF workloads. To maximize compute density while minimizing area, energy, and thermal overhead, M3D-FLOW employs hybrid binary--stochastic floating-point multipliers, performing mantissa multiplication in the stochastic domain and exponent operations in the binary domain. A pipelined execution strategy maps sequential NF stages across dynamically partitioned PE groups, and together with the stochastic hardware, enables M3D-FLOW to achieve at least $6.6\times$ higher performance and $3\times$ better energy efficiency than prior PNM accelerators, with only a 2.2% bits-per-dimension (BPD) increase over bfloat16 and 4.1% over FP32. An analysis of stochastic hardware variants further shows that shorter stochastic bitstreams reduce execution time by up to $3.3\times$ at the cost of higher approximation error, providing a concrete design guide for latency--accuracy trade-offs.

Is there a fundamental tradeoff between desiderata for safe AI systems, such as harmlessness and helpfulness? We answer by formalizing the interaction between AI systems and users as a game with three players: a defender and an attacker that are computationally or informationally bounded, and an AI that provides public advice and is less constrained. Our main result shows that there is a fundamental tradeoff in providing advice that helps the defender while advice remains public. We also consider whether the information that the AI shares can be manipulated to benefit only the defender and conclude with a discussion of how we can use these results in designing safer AI systems.

Neural image codecs face a fundamental trade-off between pixel fidelity and perceptual quality. Distortion-optimized codecs preserve pixel-wise fidelity but produce blurred reconstructions at low bitrates, whereas generative codecs leverage learned priors to improve perceptual quality but tend to exhibit saturating PSNR at higher bitrates and require separate models for different bitrate targets. We propose PerQ, which reframes perceptual reconstruction as inverse generative modeling with bounded support. Specifically, we define the quantization cell as the set of latents that yield the same transmitted symbols and derive a transport bound showing that reconstructions from this cell remain within a rate-determined radius of the codec output. This motivates performing generative reconstruction within the codec-induced quantization cell, rather than sampling from an unconstrained image prior. PerQ implements this principle by attaching a rate-conditioned flow-matching compensator to a frozen, multi-rate distortion-optimized backbone, enabling perceptual enhancement while preserving the backbone's coding performance. Experiments on Kodak and CLIC2020 show that PerQ achieves perceptual performance comparable to generative codecs such as MS-ILLM in LPIPS and FID, while maintaining competitive PSNR relative to distortion-optimized codecs such as ELIC. A single model checkpoint supports a wide bitrate range with fast decoding, and the compensator provides consistent gains across three different backbones. The implementation is available at https://jasminepp.github.io/PerQ/.

Sparse autoencoders (SAEs) seek to decompose the dense activations of large language models (LLMs) into sparse, potentially interpretable features using a learned overcomplete dictionary. In this work, we ask whether it is possible to enhance the semantic selectivity of the learned dictionary atoms by linearly \emph{shaping} the dense activations prior to learning an SAE. We explore rank-reducing transformations based on the recently introduced J-space corresponding to the averaged Jacobian from the current layer to the LLM output, projecting activations onto the ``live'' subspace spanned by eigenmodes of the Jacobian with larger singular values (which we expect to be more semantically relevant). We also introduce a new approach for training SAEs in which (a)~an affine shaping transformation is learned along with the dictionary atoms, (b)~the atom selection mechanism is based on the TopK projections along the one-dimensional subspace spanned by the atoms, and (c)~reconstruction is carried out using the same atoms. We call this approach a Matched Filter SAE (MF-SAE) based on analogies with communication theory. The promise of shaping is demonstrated via evaluations against dense activations and state-of-the-art SAEs on benchmarks including SAEBench and SAE Probes.

Migrating transformer inference to the edge requires sustainability considerations that encompass both operation and manufacturing, rather than operational energy alone. Silicon photonics enables high-bandwidth, low-energy matrix multiplication; however, existing microring-resonator-based Dot Product Unit (DPU) organizations have been tailored to convolutional computations and evaluated using analog signaling without accounting for embodied carbon footprints. We adapt three DPU organizations (AMW, MAW, and MWA) to transformer inference workloads and extend each organization to stochastic signaling, yielding six organizations in total. We then calculate their embodied and operational carbon per inference. Across five transformer models and four foundry grids, the photonic system achieves improvements of four, four, and six orders of magnitude in Energy-Delay, Carbon-Delay, and Carbon-Energy Product, respectively, relative to CPUs, while also being compared against H100 SXM5 GPU and TPU v5p baseline systems. Analog-to-stochastic conversion increases die area by 4.3-4.6x and embodied carbon by 4.4-4.8x, while reducing operational energy by approximately 91% and causing embodied carbon to equal or exceed operational carbon per inference. Importantly, the DPU organization and signaling scheme, rather than the foundry grid, are the key determinants of carbon impact. Consequently, the fastest organization (MWA) is not necessarily the most carbon-efficient, whereas MAW-Stochastic offers a favorable balance between sustainability and performance.

Large language model (LLM) agents now autonomously complete long-horizon tasks, such as repository-level coding, through hundreds of think-act steps, yet the experience each run produces is rarely reused on the next. Prior memory systems concentrate on retrieval, finding the most relevant past trajectory. Retrieval cannot observe the running agent, so it cannot decide whether the agent needs an experience, at which point, or for how long. We introduce REM (Reusable Experience Memory), a framework that manages the memory lifecycle from inside the agent's think-act loop through three explicit decisions. What to store: each completed trajectory is consolidated into a phase-structured document, two orders of magnitude smaller than the trajectory. When to recall: the document is injected at the start of the task; the costlier implementation record is withheld until an in-loop gate detects the agent struggling. What to forget: the record is evicted after a single exposure. On repository-level bug fixing (SWE-ContextBench Lite), REM retrieves the related trajectory more accurately than every state-of-the-art memory system (68.7% vs. 55.6% match@1) and improves, in every replicate, end-task resolution (27.6% vs. 24.9%) at 17% lower cost. Injecting the same experience as a summary or raw trajectory, even when oracle-retrieved, does not. Ablating any of the three decisions reduces resolution. Altogether, these results demonstrate that the bottleneck in experience reuse is not retrieval fidelity but how experience is stored and governed across its lifecycle, opening a control-theoretic perspective on agent memory.

Shipboard equipment must survive the transient loads of underwater explosions. The U.S. Navy certifies such equipment with the Dynamic Design Analysis Method (DDAM), a modal response analysis driven by Navy shock design spectra. DDAM is well established as a verification tool, yet it is rarely used as a design driver in topology optimization. The design spectrum depends on the modal mass, a non-smooth coupling that classical sensitivity formulations do not address. This paper presents a complete differentiable formulation of the DDAM criterion for gradient-based topology optimization. Eigenvalue sensitivities follow the Fox-Kapoor method, and eigenvector sensitivities follow Nelson's method. The chain rule covers a spectral displacement that depends on both the natural frequency and the modal participation factor. Automatic differentiation assembles the sensitivity, and the engineer can certify components inside the optimization loop rather than after design. Three certification problems on a representative shipboard equipment frame demonstrate the workflow. In each, the optimizer removes 43.2% to 46.1% of the design-domain material and holds the certifying NRL Sum within 0.41% of the Navy allowable. Two companion studies follow. A compliance-optimized frame of nearly the same mass has a 5.2 Hz fundamental, outside the range the certification can qualify. Its nominal response also exceeds the allowable by a factor of 2.75. A lower bound on the mean of the two lowest eigenvalues separates the repeated pair that the demonstrations drive onto the frequency floor. The response then stops oscillating across the acceptance boundary. It ends at most 0.5% above the allowable, at a cost of under two percentage points of volume fraction.

Block-diffusion speculative decoding accelerates large language model inference by drafting multiple tokens in parallel. However, one-step drafting in discrete diffusion is limited by token factorization error: each token is predicted from the same masked block, without conditioning on the tokens selected at other positions in that step. This can produce inconsistent drafts and reduce acceptance length. Iterative refinement mitigates this limitation by conditioning later predictions on previously committed tokens, though additional passes increase drafting cost. To address this trade-off between draft fidelity and inference latency, we introduce DStill, a two-stage framework that distills a target-conditioned drafter's own multi-step refinement into one-step drafting. We first train the drafter to complete partially revealed blocks, then use its refinement trajectories to train a student to predict the final completion in one pass. Using a curriculum that gradually transitions from partial to fully masked inputs, we successfully transfer the multi-step refinement capabilities into one step. Across reasoning, math, code benchmarks on three model families DStill improves acceptance length by up to 28% and speedup by up to 25% over one-step DFlash, achieving up to 6.45x speedup over auto-regressive decoding.