作者Hao-Xuan Ma, Jin-Fei Qi, Yicheng Xiao, Han-Jia Ye
Multimodal large language models (MLLMs) have recently made strong progress in vision-language reasoning, yet their performance often degrades as generations grow longer. A key factor is that they frequently lose track of earlier visual evidence and intermediate constraints under a monolithic growing context. Inspired by how humans separately recall what they see and what they infer when solving complex tasks, we propose DLMR, a parameter-efficient mechanism that equips MLLMs with Dual Latent Memories: a visual memory that compresses image evidence and a reasoning memory that tracks intermediate conclusions and constraints. A Router then dynamically decides which memory and how much to reuse during inference, preserving visual grounding while maintaining coherent long-horizon reasoning. DLMR is trained in three stages, from latent memory construction to selective router learning, while keeping the base MLLM frozen, yielding substantial gains on both general and reasoning benchmarks with only a small number of additional trainable parameters. Analyses further show interpretable, state-dependent routing with specialized memory roles and reduced decoding tokens over long generations. Code is available at https://github.com/Hunter-Wrynn/DLMR.
展开完整摘要收起摘要↓
Multimodal large language models (MLLMs) have recently made strong progress in vision-language reasoning, yet their performance often degrades as generations grow longer. A key factor is that they frequently lose track of earlier visual evidence and intermediate constraints under a monolithic growing context. Inspired by how humans separately recall what they see and what they infer when solving complex tasks, we propose DLMR, a parameter-efficient mechanism that equips MLLMs with Dual Latent Memories: a visual memory that compresses image evidence and a reasoning memory that tracks intermediate conclusions and constraints. A Router then dynamically decides which memory and how much to reuse during inference, preserving visual grounding while maintaining coherent long-horizon reasoning. DLMR is trained in three stages, from latent memory construction to selective router learning, while keeping the base MLLM frozen, yielding substantial gains on both general and reasoning benchmarks with only a small number of additional trainable parameters. Analyses further show interpretable, state-dependent routing with specialized memory roles and reduced decoding tokens over long generations. Code is available at https://github.com/Hunter-Wrynn/DLMR.
作者Huaiyuan Yao, Xiaoou Liu, Charles Fleming, Tianlong Chen, Hua Wei
LLM-based multi-agent systems have shown strong performance on complex tasks, yet continual improvement from interaction experience remains challenging. Existing self-reflection methods build experience memories, but memories are mostly hard to invoke, refine, or scale, while agent skills offer a more actionable unit: structured procedural knowledge that specifies when to act, how to act, and which resources or tools to use. We introduce MASkills, a continual learning framework that optimizes multi-agent LLM systems through agent skills. MASkills presents a new agent-optimization pipeline that integrates skill-conditioned credit assignment, hierarchical credit aggregation, and momentum-smoothed optimization, enabling agent skill libraries to evolve through refinement, induction, consolidation, and pruning. Experiments on HotpotQA, LoCoMo, and GAIA demonstrate the effectiveness of MASkills across multiple agentic tasks. Our code is available at https://github.com/DaRL-GenAI/MASkills
展开完整摘要收起摘要↓
LLM-based multi-agent systems have shown strong performance on complex tasks, yet continual improvement from interaction experience remains challenging. Existing self-reflection methods build experience memories, but memories are mostly hard to invoke, refine, or scale, while agent skills offer a more actionable unit: structured procedural knowledge that specifies when to act, how to act, and which resources or tools to use. We introduce MASkills, a continual learning framework that optimizes multi-agent LLM systems through agent skills. MASkills presents a new agent-optimization pipeline that integrates skill-conditioned credit assignment, hierarchical credit aggregation, and momentum-smoothed optimization, enabling agent skill libraries to evolve through refinement, induction, consolidation, and pruning. Experiments on HotpotQA, LoCoMo, and GAIA demonstrate the effectiveness of MASkills across multiple agentic tasks. Our code is available at https://github.com/DaRL-GenAI/MASkills
LLM agents are increasingly deployed in multi-agent systems, where they can collude while keeping their actions benign. Output monitors designed to detect such collusions can be fooled by obfuscation and steganography, motivating the use of probes trained on internal activations. However, these probes are usually evaluated on agents that do not know they are being watched. We study how activation-based detection changes when agents are explicitly informed that their internal activations are being monitored, and when they additionally receive feedback from the monitor. We keep the models, probes, and thresholds fixed and change only what the agents are told: nothing (baseline), that an activation monitor is present (aware), or that a monitor is present together with the previous round's score (feedback). We test two games, a four-agent blackjack game and a two-agent Simmons prisoners game, using Qwen3-32B-AWQ and GPT-OSS-20B in homogeneous and mixed pairings. Telling agents about the monitor does not hide them. The best probes stay accurate in all three conditions, and the agents keep colluding.
展开完整摘要收起摘要↓
LLM agents are increasingly deployed in multi-agent systems, where they can collude while keeping their actions benign. Output monitors designed to detect such collusions can be fooled by obfuscation and steganography, motivating the use of probes trained on internal activations. However, these probes are usually evaluated on agents that do not know they are being watched. We study how activation-based detection changes when agents are explicitly informed that their internal activations are being monitored, and when they additionally receive feedback from the monitor. We keep the models, probes, and thresholds fixed and change only what the agents are told: nothing (baseline), that an activation monitor is present (aware), or that a monitor is present together with the previous round's score (feedback). We test two games, a four-agent blackjack game and a two-agent Simmons prisoners game, using Qwen3-32B-AWQ and GPT-OSS-20B in homogeneous and mixed pairings. Telling agents about the monitor does not hide them. The best probes stay accurate in all three conditions, and the agents keep colluding.
Topology-guided safeguards for LLM-based multi-agent systems (MAS) train a GNN over the inter-agent communication graph to localize risky agents and intervene on the topology---but they assume one operator can pool all labeled traces. Across organizations that assumption breaks: episodes contain private prompts, tool outputs, and proprietary workflows, and no silo alone sees the full attack distribution. We cast privacy-preserving MAS safeguarding as graph federated learning and instantiate FGLGuard: each operator fits an edge-featured graph attention detector on its own judge-labeled episode graphs and shares only model updates. The method couples a proximal local objective for non-IID clients, domain-balanced aggregation, over-refusal-constrained threshold calibration, corroborated upstream scoring, and a guarded rewrite for blocked answers. Federation is not optional: off-the-shelf transfer collapses under distribution shift (AUROC 0.51 to 0.70 only after in-domain retraining), so a deployable guard must adapt on each site's private traces. On Agent-SafetyBench, R-Judge, and AgentDojo, federated FGLGuard exceeds the in-domain centralized ceiling on all three benchmarks without pooling any data---where unsupervised anomaly guards and local-only training fail. One guard federated across four different-domain operators comes within 0.03 AUROC of multi-domain centralization, while any single-domain guard collapses on the others. Live FGLGuard cuts AgentDojo's ground-truth attack-success rate by 43% at near-unguarded utility, zero API cost, and negligible capability loss.
展开完整摘要收起摘要↓
Topology-guided safeguards for LLM-based multi-agent systems (MAS) train a GNN over the inter-agent communication graph to localize risky agents and intervene on the topology---but they assume one operator can pool all labeled traces. Across organizations that assumption breaks: episodes contain private prompts, tool outputs, and proprietary workflows, and no silo alone sees the full attack distribution. We cast privacy-preserving MAS safeguarding as graph federated learning and instantiate FGLGuard: each operator fits an edge-featured graph attention detector on its own judge-labeled episode graphs and shares only model updates. The method couples a proximal local objective for non-IID clients, domain-balanced aggregation, over-refusal-constrained threshold calibration, corroborated upstream scoring, and a guarded rewrite for blocked answers. Federation is not optional: off-the-shelf transfer collapses under distribution shift (AUROC 0.51 to 0.70 only after in-domain retraining), so a deployable guard must adapt on each site's private traces. On Agent-SafetyBench, R-Judge, and AgentDojo, federated FGLGuard exceeds the in-domain centralized ceiling on all three benchmarks without pooling any data---where unsupervised anomaly guards and local-only training fail. One guard federated across four different-domain operators comes within 0.03 AUROC of multi-domain centralization, while any single-domain guard collapses on the others. Live FGLGuard cuts AgentDojo's ground-truth attack-success rate by 43% at near-unguarded utility, zero API cost, and negligible capability loss.
LLM-based multi-agent systems (MAS) are increasingly considered for high-stakes decision-making, yet outcome-based fairness audits can miss where risks arise within the decision trajectory. We present SCOPED-Hiring, a process-aware fairness diagnosis pipeline for LLM-based hiring MAS. SCOPED-Hiring constructs controlled resume variants, runs role-based hiring committees, logs over 311K structured decision trajectories, and converts trajectory fields into quantitative fairness signals organized by six diagnostic lenses: final outcome, counterfactual, process, pathway, dynamic, and design effects. SCOPED-Hiring reveals that balanced final hire rates can mask hidden trajectory unfairness in multi-agent decision trajectories: career gaps trigger suspicion, proxy cues shape qualification judgments, and identity cues lead to unequal investigation. Targeted repair guided by these diagnoses reduces total layered burden by 72.3% while shifting the hire rate by only 1.86 pp, showing that process diagnosis can guide effective repair. Project Page: https://scoped-hiring-project-page.vercel.app/
展开完整摘要收起摘要↓
LLM-based multi-agent systems (MAS) are increasingly considered for high-stakes decision-making, yet outcome-based fairness audits can miss where risks arise within the decision trajectory. We present SCOPED-Hiring, a process-aware fairness diagnosis pipeline for LLM-based hiring MAS. SCOPED-Hiring constructs controlled resume variants, runs role-based hiring committees, logs over 311K structured decision trajectories, and converts trajectory fields into quantitative fairness signals organized by six diagnostic lenses: final outcome, counterfactual, process, pathway, dynamic, and design effects. SCOPED-Hiring reveals that balanced final hire rates can mask hidden trajectory unfairness in multi-agent decision trajectories: career gaps trigger suspicion, proxy cues shape qualification judgments, and identity cues lead to unequal investigation. Targeted repair guided by these diagnoses reduces total layered burden by 72.3% while shifting the hire rate by only 1.86 pp, showing that process diagnosis can guide effective repair. Project Page: https://scoped-hiring-project-page.vercel.app/
Removing complete transformer layers preserves a standard serving architecture, but existing depth-compression methods can lose substantial quality, and the loss varies unpredictably across models. We introduce XMerge, a post-training method with two components. Cross-axis selection identifies a block with low relative-magnitude and angular hidden-state change, and local boundary reconstruction re-fits the adjacent surviving block to match the original two-block output. XMerge uses no task labels or end-to-end fine-tuning, and it introduces neither architectural changes nor additional inference-time parameters. Across seven Llama and Qwen backbones (0.5B-8B), five published baselines, and three layer-reduction levels, its advantage over baselines is largest at the most aggressive removal: at k=4 it ranks first on six of seven backbones on CORE (a 22-task aggregate) and, separately, on six of seven on MMLU (five of seven on both at once), while avoiding the large perplexity increases of several competing operators. In a task-level bootstrap, the 95% confidence intervals for the three largest CORE margins exclude zero; the remaining margins are consistent with ties. Across the 14 (model, regime) cells it is also the only evaluated operator that never collapses, ranking top-2 in both zero-shot and in-context regimes; on a first calibration probe (one backbone) it is the best-calibrated operator. Ablations show that local reconstruction provides most of the gain, while cross-axis fusion helps when the two selection axes disagree. The additional construction cost is recovered through per-token decode savings after roughly tens of thousands of requests.
展开完整摘要收起摘要↓
Removing complete transformer layers preserves a standard serving architecture, but existing depth-compression methods can lose substantial quality, and the loss varies unpredictably across models. We introduce XMerge, a post-training method with two components. Cross-axis selection identifies a block with low relative-magnitude and angular hidden-state change, and local boundary reconstruction re-fits the adjacent surviving block to match the original two-block output. XMerge uses no task labels or end-to-end fine-tuning, and it introduces neither architectural changes nor additional inference-time parameters. Across seven Llama and Qwen backbones (0.5B-8B), five published baselines, and three layer-reduction levels, its advantage over baselines is largest at the most aggressive removal: at k=4 it ranks first on six of seven backbones on CORE (a 22-task aggregate) and, separately, on six of seven on MMLU (five of seven on both at once), while avoiding the large perplexity increases of several competing operators. In a task-level bootstrap, the 95% confidence intervals for the three largest CORE margins exclude zero; the remaining margins are consistent with ties. Across the 14 (model, regime) cells it is also the only evaluated operator that never collapses, ranking top-2 in both zero-shot and in-context regimes; on a first calibration probe (one backbone) it is the best-calibrated operator. Ablations show that local reconstruction provides most of the gain, while cross-axis fusion helps when the two selection axes disagree. The additional construction cost is recovered through per-token decode savings after roughly tens of thousands of requests.
In this paper, we propose a scalable Kronecker-based approximation that captures cross-layer interactions without storing the entire Fisher matrix, enabling practical Hessian analysis for billion-parameter networks where full computation is infeasible. Our approach reveals consistent vulnerability patterns: value projection layers exhibit the highest sensitivity and strongest cross-layer correlations across multiple model families, while other components exhibit architecture-specific behaviors. Through extensive experiments on quantization, sparsification, inter-layer corruption, and post-corruption fine-tuning, we demonstrate that our approximation strongly correlates with both performance degradation and recovery. Our framework provides a practical, theoretically grounded tool for identifying fragile components in large models, opening new avenues for guided compression and optimization strategies, such as mixed-precision allocation, layer-wise sparsity, and adaptive low-rank decomposition across layers and even individual weight groups.
展开完整摘要收起摘要↓
In this paper, we propose a scalable Kronecker-based approximation that captures cross-layer interactions without storing the entire Fisher matrix, enabling practical Hessian analysis for billion-parameter networks where full computation is infeasible. Our approach reveals consistent vulnerability patterns: value projection layers exhibit the highest sensitivity and strongest cross-layer correlations across multiple model families, while other components exhibit architecture-specific behaviors. Through extensive experiments on quantization, sparsification, inter-layer corruption, and post-corruption fine-tuning, we demonstrate that our approximation strongly correlates with both performance degradation and recovery. Our framework provides a practical, theoretically grounded tool for identifying fragile components in large models, opening new avenues for guided compression and optimization strategies, such as mixed-precision allocation, layer-wise sparsity, and adaptive low-rank decomposition across layers and even individual weight groups.
作者Alejandro Barón García, Feng Wang, Emilia Garcia Casademont, Han Xiao
We present Jina-OCR-v1, an end-to-end document parsing model built to serve on low-budget GPUs. It combines the compressed-vision encoder and the 3B mixture-of-experts decoder of DeepSeek-OCR, which activates about 570M parameters per token, with a FastMTP speculative decoding head that shares a single draft block recursively across K=3 prediction steps. Greedy verification makes decoding lossless. Post-training combines instruction alignment, robustness fine-tuning on difficult documents, and GRPO under dense verifiable rewards: deterministic formula, table, and structural checks that award partial credit. The training data mixes cleaned public corpora with targeted synthetic pages. At the default dynamic-resolution setting, Jina-OCR-v1 scores 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench, and reaches the highest page throughput in our comparison at 2.57 pages per second. On a low-budget GPU such as the NVIDIA L4, FastMTP doubles decoding speed over greedy autoregressive decoding. The model is publicly available at https://huggingface.co/jinaai/jina-ocr-v1.
展开完整摘要收起摘要↓
We present Jina-OCR-v1, an end-to-end document parsing model built to serve on low-budget GPUs. It combines the compressed-vision encoder and the 3B mixture-of-experts decoder of DeepSeek-OCR, which activates about 570M parameters per token, with a FastMTP speculative decoding head that shares a single draft block recursively across K=3 prediction steps. Greedy verification makes decoding lossless. Post-training combines instruction alignment, robustness fine-tuning on difficult documents, and GRPO under dense verifiable rewards: deterministic formula, table, and structural checks that award partial credit. The training data mixes cleaned public corpora with targeted synthetic pages. At the default dynamic-resolution setting, Jina-OCR-v1 scores 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench, and reaches the highest page throughput in our comparison at 2.57 pages per second. On a low-budget GPU such as the NVIDIA L4, FastMTP doubles decoding speed over greedy autoregressive decoding. The model is publicly available at https://huggingface.co/jinaai/jina-ocr-v1.
作者Ankur Singh, Ashish Gautam, Shruti R. Kulkarni, Guojing Cong
Softmax is a key operation in Transformer attention, but its exponentiation and normalization add significant overhead in compute-in-memory (CIM) accelerators, especially when analog attention scores must first be converted to the digital domain. This work presents a tunable-temperature analog softmax circuit in GlobalFoundries 22-nm fully depleted silicon-on-insulator (FDSOI) technology that operates directly on CIM-generated score voltages without intermediate analog-to-digital conversion. Each input score is converted into a time-domain event using a shared falling ramp. The corresponding comparator transition samples an RC-decaying reference to generate an exponential weight, which is then processed by an in-circuit normalization stage. In contrast to analog softmax circuits that rely on transistor weak-inversion behavior for exponentiation, the proposed architecture controls the softmax response through the ramp slope and RC time constant, enabling programmable effective temperature. The 128-element architecture is evaluated using transistor-level and post-layout extracted simulations, including multi-level input vectors, capacitance variation and mismatch, process and temperature variation, monte carlo analysis, and shared-interconnect parasitics. The complete 128-element implementation occupies 9453.42~$μ\mathrm{m}^{2}$ including the shared global ramp circuitry, while each replicated softmax element occupies 70.2~$μ\mathrm{m}^{2}$. The circuit achieves a 242.97-ns evaluation latency at 13.44~mW total power, corresponding to 25.5~pJ per output element. The simultaneous 128-element evaluation achieves an RMSE of 24.46~mV relative to the ideal softmax response. The extracted circuit characteristics are further incorporated into a MemTorch-based hardware-aware Transformer model, where the proposed softmax achieves a validation loss within 2.5% of the ideal-softmax baseline.
展开完整摘要收起摘要↓
Softmax is a key operation in Transformer attention, but its exponentiation and normalization add significant overhead in compute-in-memory (CIM) accelerators, especially when analog attention scores must first be converted to the digital domain. This work presents a tunable-temperature analog softmax circuit in GlobalFoundries 22-nm fully depleted silicon-on-insulator (FDSOI) technology that operates directly on CIM-generated score voltages without intermediate analog-to-digital conversion. Each input score is converted into a time-domain event using a shared falling ramp. The corresponding comparator transition samples an RC-decaying reference to generate an exponential weight, which is then processed by an in-circuit normalization stage. In contrast to analog softmax circuits that rely on transistor weak-inversion behavior for exponentiation, the proposed architecture controls the softmax response through the ramp slope and RC time constant, enabling programmable effective temperature. The 128-element architecture is evaluated using transistor-level and post-layout extracted simulations, including multi-level input vectors, capacitance variation and mismatch, process and temperature variation, monte carlo analysis, and shared-interconnect parasitics. The complete 128-element implementation occupies 9453.42~$μ\mathrm{m}^{2}$ including the shared global ramp circuitry, while each replicated softmax element occupies 70.2~$μ\mathrm{m}^{2}$. The circuit achieves a 242.97-ns evaluation latency at 13.44~mW total power, corresponding to 25.5~pJ per output element. The simultaneous 128-element evaluation achieves an RMSE of 24.46~mV relative to the ideal softmax response. The extracted circuit characteristics are further incorporated into a MemTorch-based hardware-aware Transformer model, where the proposed softmax achieves a validation loss within 2.5% of the ideal-softmax baseline.
作者To Truong An, Jie Zhang, Guolin Yin, Junqing Zhang, Yanjiao Li, Trung Q. Duong, Simon L. Cotton
Wi-Fi-based human activity recognition (HAR) has become an important part of integrated sensing and communications, paving the way for a range of context-aware services. However, most existing Wi-Fi-based HAR systems rely on deep learning (DL) models that are computationally and memory intensive in both training and inference, which poses significant challenges for real-world deployment. Conventional training requires simultaneous updates of millions of parameters, leading to prohibitive memory consumption. In this paper, we propose a novel quantum-assisted memory-efficient training framework (Q-MET) designed to improve efficiency in both training and inference. Q-MET utilizes a hybrid quantum classical neural network to indirectly generate parameters for HAR models, significantly reducing the trainable parameter count compared to direct optimization. To further support the deployment on resource-constrained devices, we integrate structured pruning during the training phase. Experimental results demonstrate that Q-MET achieves a 90% to 95% reduction in trainable parameters compared with conventional backpropagation-based DL training while maintaining or even exceeding classical classification accuracy. Additionally, Q-MET supports lightweight inference through structured pruning, achieving 75% to 85% model sparsity with less than 2% loss in classification accuracy. To the best of our knowledge, this work represents the first quantum-assisted approach to simultaneously tackle memory inefficiencies in both the training and inference stages of HAR systems.
展开完整摘要收起摘要↓
Wi-Fi-based human activity recognition (HAR) has become an important part of integrated sensing and communications, paving the way for a range of context-aware services. However, most existing Wi-Fi-based HAR systems rely on deep learning (DL) models that are computationally and memory intensive in both training and inference, which poses significant challenges for real-world deployment. Conventional training requires simultaneous updates of millions of parameters, leading to prohibitive memory consumption. In this paper, we propose a novel quantum-assisted memory-efficient training framework (Q-MET) designed to improve efficiency in both training and inference. Q-MET utilizes a hybrid quantum classical neural network to indirectly generate parameters for HAR models, significantly reducing the trainable parameter count compared to direct optimization. To further support the deployment on resource-constrained devices, we integrate structured pruning during the training phase. Experimental results demonstrate that Q-MET achieves a 90% to 95% reduction in trainable parameters compared with conventional backpropagation-based DL training while maintaining or even exceeding classical classification accuracy. Additionally, Q-MET supports lightweight inference through structured pruning, achieving 75% to 85% model sparsity with less than 2% loss in classification accuracy. To the best of our knowledge, this work represents the first quantum-assisted approach to simultaneously tackle memory inefficiencies in both the training and inference stages of HAR systems.
作者Tianhao Peng, Ho Man Kwan, Fan Zhang, Shan Liu, David Bull
Overfitted codecs have demonstrated promising performance for image and video compression. In particular, for image compression, the Cool-chic family of models has shown competitive performance against scene-agnostic models, with orders of magnitude lower decoding complexity at the cost of a longer overfitting process. However, these overfitted image codecs are not fully optimized toward the rate-distortion objective: their network weights remain in full precision during training, and the associated quantization parameters are selected in a separate post-training stage. Furthermore, their synthesis operates at a single scale, which overlooks cross-scale redundancy. In this paper, we propose MIRC, an overfitted image codec in which every coded component, including the latents, the synthesis network, and the entropy models, is quantized and entropy coded under a single rate-distortion objective, adopting the end-to-end compression pipeline of the neural video representation codec NVRC. We further introduce a multi-scale representation with cross-stage parameter sharing, which improves coding efficiency at a small transmitted overhead. On the CLIC2020 professional validation set, MIRC achieves a 10.5% BD-rate saving against VVC (VTM 22.0). Moreover, MIRC offers a family of configurations spanning 1.2 to 2.9 kMAC per pixel, so the decoding budget can be selected to match the deployment target.
展开完整摘要收起摘要↓
Overfitted codecs have demonstrated promising performance for image and video compression. In particular, for image compression, the Cool-chic family of models has shown competitive performance against scene-agnostic models, with orders of magnitude lower decoding complexity at the cost of a longer overfitting process. However, these overfitted image codecs are not fully optimized toward the rate-distortion objective: their network weights remain in full precision during training, and the associated quantization parameters are selected in a separate post-training stage. Furthermore, their synthesis operates at a single scale, which overlooks cross-scale redundancy. In this paper, we propose MIRC, an overfitted image codec in which every coded component, including the latents, the synthesis network, and the entropy models, is quantized and entropy coded under a single rate-distortion objective, adopting the end-to-end compression pipeline of the neural video representation codec NVRC. We further introduce a multi-scale representation with cross-stage parameter sharing, which improves coding efficiency at a small transmitted overhead. On the CLIC2020 professional validation set, MIRC achieves a 10.5% BD-rate saving against VVC (VTM 22.0). Moreover, MIRC offers a family of configurations spanning 1.2 to 2.9 kMAC per pixel, so the decoding budget can be selected to match the deployment target.
Low-bit key--value (KV) caches reduce the memory required for autoregressive decoding, but the resulting quality loss depends on the model and quantizer. We keep the quantizer fixed and distill the floating-cache model's behavior into low-rank Q/K/V projection updates while the student executes a physically packed incremental cache. Across three seeds, 4-bit affine-cache adapters recover $54.24%\pm2.47%$ of the held-out perplexity gap on TinyLlama-1.1B and $75.96%\pm4.04%$ on Gemma-4-12B. On the same frozen NF4 Llama-3.1-8B base, one validation-selected run per quantizer recovers $60.42%$ under KIVI K2V2 and $37.61%$ under KVarN K4V2, while preserving 180-case associative retrieval. Gemma's score on an official 4K/8K RULER subset rises from 42.80 with the unadapted 4-bit cache to 48.33 after adaptation (46.15 floating), with substantial task heterogeneity. Finally, a 2-bit rank--token sweep reduces TinyLlama's 2-bit PPL from 576.10 to $11.4000\pm0.0059$ across three seeds, versus 10.3988 floating, but restores only 11--12 of 180 retrieval cases. These results show that low-rank projection adaptation can recover held-out quality across fixed cache formats, while perplexity recovery need not restore long-context retrieval.
展开完整摘要收起摘要↓
Low-bit key--value (KV) caches reduce the memory required for autoregressive decoding, but the resulting quality loss depends on the model and quantizer. We keep the quantizer fixed and distill the floating-cache model's behavior into low-rank Q/K/V projection updates while the student executes a physically packed incremental cache. Across three seeds, 4-bit affine-cache adapters recover $54.24%\pm2.47%$ of the held-out perplexity gap on TinyLlama-1.1B and $75.96%\pm4.04%$ on Gemma-4-12B. On the same frozen NF4 Llama-3.1-8B base, one validation-selected run per quantizer recovers $60.42%$ under KIVI K2V2 and $37.61%$ under KVarN K4V2, while preserving 180-case associative retrieval. Gemma's score on an official 4K/8K RULER subset rises from 42.80 with the unadapted 4-bit cache to 48.33 after adaptation (46.15 floating), with substantial task heterogeneity. Finally, a 2-bit rank--token sweep reduces TinyLlama's 2-bit PPL from 576.10 to $11.4000\pm0.0059$ across three seeds, versus 10.3988 floating, but restores only 11--12 of 180 retrieval cases. These results show that low-rank projection adaptation can recover held-out quality across fixed cache formats, while perplexity recovery need not restore long-context retrieval.
This study examines the ability of large language models (LLMs) to predict the risk of weather-related forced outages in the distribution grid in a zero-shot framework, without labeled training data. The problem is formulated as a binary severity classification task across three forecast horizons (3h, 6h, 12h), using six years of outage records and high-resolution weather data for a utility service area in central Texas. Four zero-shot LLMs are benchmarked against two supervised classifiers across two input configurations: one using current weather observations and the other using weather forecast data. Results show that supervised models outperform LLMs on macro-F1 and precision, while newer LLM generations achieve competitive scores. Beyond accuracy, LLMs offer complementary strengths in actionable reasoning and geographic scalability, suggesting that combining them with supervised models may be the best practice.
展开完整摘要收起摘要↓
This study examines the ability of large language models (LLMs) to predict the risk of weather-related forced outages in the distribution grid in a zero-shot framework, without labeled training data. The problem is formulated as a binary severity classification task across three forecast horizons (3h, 6h, 12h), using six years of outage records and high-resolution weather data for a utility service area in central Texas. Four zero-shot LLMs are benchmarked against two supervised classifiers across two input configurations: one using current weather observations and the other using weather forecast data. Results show that supervised models outperform LLMs on macro-F1 and precision, while newer LLM generations achieve competitive scores. Beyond accuracy, LLMs offer complementary strengths in actionable reasoning and geographic scalability, suggesting that combining them with supervised models may be the best practice.
作者Zhiwei Zhang, Zechen Sun, Fei Zhao, Kang Peng, Bin Liang, Huayu Deng, Yao Hu, Kam-Fai Wong, Mu Chuan
On-policy distillation (OPD) accelerates post-training by providing dense token-level supervision from a frozen teacher on the student's own rollouts. Vanilla OPD applies this supervision uniformly across prompts, without checking whether the teacher is reliable for each prompt. Because reverse KL is mode-seeking, a confidently wrong teacher can induce a strong yet misleading update. Distributional proxies, such as entropy or teacher-student likelihood agreement, measure uncertainty or agreement but do not directly verify outcome correctness. We introduce Teacher-Gated On-Policy Distillation (TGOPD), built on the principle that teacher reliability should be verified at the prompt level before dense supervision is admitted. TGOPD estimates reliability from a small set of verifier-scored teacher probes and routes each prompt exclusively to dense OPD when the reliability check passes or to verifier-grounded GRPO otherwise. Across 4B and 35B students in mathematics, code, and instruction following, TGOPD outperforms Vanilla OPD in all six single-domain settings and achieves higher seven-benchmark averages at both scales under multi-domain training. By using otherwise-idle teacher capacity for reliability estimation, TGOPD also reduces teacher-side compute waste in asynchronous OPD, increasing teacher-node GPU utilization from 9.8% to 78.9% in the measured 4B single-domain run.
展开完整摘要收起摘要↓
On-policy distillation (OPD) accelerates post-training by providing dense token-level supervision from a frozen teacher on the student's own rollouts. Vanilla OPD applies this supervision uniformly across prompts, without checking whether the teacher is reliable for each prompt. Because reverse KL is mode-seeking, a confidently wrong teacher can induce a strong yet misleading update. Distributional proxies, such as entropy or teacher-student likelihood agreement, measure uncertainty or agreement but do not directly verify outcome correctness. We introduce Teacher-Gated On-Policy Distillation (TGOPD), built on the principle that teacher reliability should be verified at the prompt level before dense supervision is admitted. TGOPD estimates reliability from a small set of verifier-scored teacher probes and routes each prompt exclusively to dense OPD when the reliability check passes or to verifier-grounded GRPO otherwise. Across 4B and 35B students in mathematics, code, and instruction following, TGOPD outperforms Vanilla OPD in all six single-domain settings and achieves higher seven-benchmark averages at both scales under multi-domain training. By using otherwise-idle teacher capacity for reliability estimation, TGOPD also reduces teacher-side compute waste in asynchronous OPD, increasing teacher-node GPU utilization from 9.8% to 78.9% in the measured 4B single-domain run.
As services move online, trust institutions such as banks, lenders, and governments must verify the identity of remote users. Fraud detection tools are widely available, but evaluating and fine-tuning them remains difficult because identity documents are sensitive and therefore scarce. Synthetic data generation offers a path forward, and demand is clear: our prior work in this area has been downloaded over $11{,}000$ times (aggregated from eight parts). We introduce IDSpace, extending this line of research in three directions. First, we propose model-guided Bayesian optimization, which tunes generation parameters to maximize both visual similarity and prediction consistency with target-domain models given only a few samples from a target domain. Second, we decouple user-specified metadata (demographics, fraud patterns, capture device) from automatically tuned control parameters (font styles, noise levels, image quality), allowing users to configure evaluations without low-level expertise. Third, we expand beyond template images to support scanned and mobile-captured documents. Experiments show IDSpace improves evaluation consistency by $15-45%$ over baselines including CycleGAN, diffusion inpainting, and non-guided optimization, using only a few real samples, while improving training accuracy by up to $9%$ and SSIM similarity with the target domain by $10%$. We also released a new dataset consisting of $359{,}240$ high-quality synthetic documents across ten European ID types.
展开完整摘要收起摘要↓
As services move online, trust institutions such as banks, lenders, and governments must verify the identity of remote users. Fraud detection tools are widely available, but evaluating and fine-tuning them remains difficult because identity documents are sensitive and therefore scarce. Synthetic data generation offers a path forward, and demand is clear: our prior work in this area has been downloaded over $11{,}000$ times (aggregated from eight parts). We introduce IDSpace, extending this line of research in three directions. First, we propose model-guided Bayesian optimization, which tunes generation parameters to maximize both visual similarity and prediction consistency with target-domain models given only a few samples from a target domain. Second, we decouple user-specified metadata (demographics, fraud patterns, capture device) from automatically tuned control parameters (font styles, noise levels, image quality), allowing users to configure evaluations without low-level expertise. Third, we expand beyond template images to support scanned and mobile-captured documents. Experiments show IDSpace improves evaluation consistency by $15-45%$ over baselines including CycleGAN, diffusion inpainting, and non-guided optimization, using only a few real samples, while improving training accuracy by up to $9%$ and SSIM similarity with the target domain by $10%$. We also released a new dataset consisting of $359{,}240$ high-quality synthetic documents across ten European ID types.
作者Alessandro Coretti, Nico Unglert, Sebastian Falkner, Georg K. H. Madsen, Christoph Dellago
Nested sampling (NS) resolves the thermodynamics of an atomistic system from a single simulation, but its practical reach is limited by the Markov-chain updates needed to decorrelate walkers within each likelihood-constrained ensemble. Flow-based NS has removed this bottleneck for gravitational-wave (GW) inference, yet its transfer to atomistic systems is not merely a change of application. Comparing a GW150914-like binary-black-hole likelihood with an eight-particle two-dimensional Lennard-Jones (LJ) system of comparable dimensionality, we show that the two landscapes differ fundamentally: atomistic multimodality is discrete and combinatorial, generated by particle permutations separated by hard collision walls, and its coordinate coupling is dense and collective, whereas the GW posterior exhibits smooth degeneracies and localized parameter coupling. Guided by this diagnosis, we introduce NS-Flows: a single conditional normalizing flow, conditioned on the NS energy bound and trained on a sliding window of recent live sets, that replaces MCMC by direct parallel draws corrected by importance-weighted rejection resampling. Live sets supply data self-consistently, allowing flow training without structured priors or a pre-existing dataset. For LJ disks in PBC, the algorithm reduces energy evaluations by over two orders of magnitude and wall-clock time by roughly one third, an advantage that becomes increasingly favorable as the cost of the potential grows. The flow's generation efficiency further acts as a physical diagnostic: it varies non-monotonically along the annealing trajectory, is lowest in the dense disordered regime, and is quantitatively captured by the constrained ensemble's internal mode complexity together with target drift across the training window, identifying liquid-like ensembles, rather than prior-target separation, as the hard case for current flow architectures.
展开完整摘要收起摘要↓
Nested sampling (NS) resolves the thermodynamics of an atomistic system from a single simulation, but its practical reach is limited by the Markov-chain updates needed to decorrelate walkers within each likelihood-constrained ensemble. Flow-based NS has removed this bottleneck for gravitational-wave (GW) inference, yet its transfer to atomistic systems is not merely a change of application. Comparing a GW150914-like binary-black-hole likelihood with an eight-particle two-dimensional Lennard-Jones (LJ) system of comparable dimensionality, we show that the two landscapes differ fundamentally: atomistic multimodality is discrete and combinatorial, generated by particle permutations separated by hard collision walls, and its coordinate coupling is dense and collective, whereas the GW posterior exhibits smooth degeneracies and localized parameter coupling. Guided by this diagnosis, we introduce NS-Flows: a single conditional normalizing flow, conditioned on the NS energy bound and trained on a sliding window of recent live sets, that replaces MCMC by direct parallel draws corrected by importance-weighted rejection resampling. Live sets supply data self-consistently, allowing flow training without structured priors or a pre-existing dataset. For LJ disks in PBC, the algorithm reduces energy evaluations by over two orders of magnitude and wall-clock time by roughly one third, an advantage that becomes increasingly favorable as the cost of the potential grows. The flow's generation efficiency further acts as a physical diagnostic: it varies non-monotonically along the annealing trajectory, is lowest in the dense disordered regime, and is quantitatively captured by the constrained ensemble's internal mode complexity together with target drift across the training window, identifying liquid-like ensembles, rather than prior-target separation, as the hard case for current flow architectures.
作者Meriem Yacoubi, Pia Schmidt, Nenad Petrovic, Ahmed Frikha, Martin Kirchhoff, Alois Knoll
Long-term LLM agents must preserve information across interactions while distinguishing repeated evidence, historical states, updates, and unresolved contradictions. Existing textual memory systems retrieve semantically relevant memories efficiently but often leave these relationships implicit, whereas richer structured approaches model them through global graphs, hierarchical abstractions, or reflection at greater complexity. We introduce MemoryLACE (MemLACE), a lightweight memory framework that explicitly models the lifecycle of textual evidence through sparse merge, supersession, and contradiction relations while preserving atomic natural-language memories and their provenance. Rather than retrieving memories independently, MemLACE reconstructs relation-aware evidence units that expose current, historical, supporting, and conflicting evidence for downstream reasoning. Across BEAM and StructMemEval, using open-weight and proprietary LLM backbones, MemLACE achieves the highest overall performance in same-backbone comparisons while reducing end-to-end runtime on BEAM by 66.6% relative to Hindsight, the strongest reported reflective-memory baseline. Ablation studies identify lifecycle expansion and temporal awareness as the principal contributors to these gains. Together, the results demonstrate that explicitly modeling the local lifecycle of textual evidence is sufficient to substantially improve long-term memory reasoning without requiring comprehensive knowledge graphs or global reflection.
展开完整摘要收起摘要↓
Long-term LLM agents must preserve information across interactions while distinguishing repeated evidence, historical states, updates, and unresolved contradictions. Existing textual memory systems retrieve semantically relevant memories efficiently but often leave these relationships implicit, whereas richer structured approaches model them through global graphs, hierarchical abstractions, or reflection at greater complexity. We introduce MemoryLACE (MemLACE), a lightweight memory framework that explicitly models the lifecycle of textual evidence through sparse merge, supersession, and contradiction relations while preserving atomic natural-language memories and their provenance. Rather than retrieving memories independently, MemLACE reconstructs relation-aware evidence units that expose current, historical, supporting, and conflicting evidence for downstream reasoning. Across BEAM and StructMemEval, using open-weight and proprietary LLM backbones, MemLACE achieves the highest overall performance in same-backbone comparisons while reducing end-to-end runtime on BEAM by 66.6% relative to Hindsight, the strongest reported reflective-memory baseline. Ablation studies identify lifecycle expansion and temporal awareness as the principal contributors to these gains. Together, the results demonstrate that explicitly modeling the local lifecycle of textual evidence is sufficient to substantially improve long-term memory reasoning without requiring comprehensive knowledge graphs or global reflection.
作者Athi Ram R S, Alwin A, S. G. Sreejeesh, J. U. Kidav
Computation of adaptive beamforming weights in Minimum Variance Distortionless Response (MVDR) processing is a latency-critical operation that poses significant challenges for real-time hardware implementation. This paper presents an FPGA implementation of a systolic Givens-rotation QR decomposition pipeline for MVDR beamforming on a simulated 32-element ultrasound transducer array, using single-precision floating-point arithmetic. The design is deployed on a Zynq UltraScale+ FPGA at 100 MHz with three parallel kernel instances operating concurrently, achieving a measured throughput of 31,123 weight vectors/s at 90.9% parallel efficiency relative to the measured single-instance rate. At an estimated 2.451 W of programmable-logic power, and 5.286 W including the processing system, this corresponds to 12,698 and 5,888 weight vectors/s/W, respectively. Under a matched three-way dispatch, a 24-core Intel Xeon Gold 5220R at 2.20 GHz achieves 465,699 weight vectors/s at a measured 83.08 W package power, corresponding to 5,606 weight vectors/s/W. The FPGA therefore attains 2.3x the power-normalised throughput of the processor on a programmable-logic basis and 1.05x on a total on-chip basis. In contrast, the processor retains a raw throughput advantage of approximately 15x at this operating point. Numerical precision is validated against MATLAB float32 reference outputs from a Field II cyst phantom simulation, achieving a 100% pass rate with a root mean square error of 5.10x10^-7, confirming near-theoretical finite-precision behaviour without systematic bias.
展开完整摘要收起摘要↓
Computation of adaptive beamforming weights in Minimum Variance Distortionless Response (MVDR) processing is a latency-critical operation that poses significant challenges for real-time hardware implementation. This paper presents an FPGA implementation of a systolic Givens-rotation QR decomposition pipeline for MVDR beamforming on a simulated 32-element ultrasound transducer array, using single-precision floating-point arithmetic. The design is deployed on a Zynq UltraScale+ FPGA at 100 MHz with three parallel kernel instances operating concurrently, achieving a measured throughput of 31,123 weight vectors/s at 90.9% parallel efficiency relative to the measured single-instance rate. At an estimated 2.451 W of programmable-logic power, and 5.286 W including the processing system, this corresponds to 12,698 and 5,888 weight vectors/s/W, respectively. Under a matched three-way dispatch, a 24-core Intel Xeon Gold 5220R at 2.20 GHz achieves 465,699 weight vectors/s at a measured 83.08 W package power, corresponding to 5,606 weight vectors/s/W. The FPGA therefore attains 2.3x the power-normalised throughput of the processor on a programmable-logic basis and 1.05x on a total on-chip basis. In contrast, the processor retains a raw throughput advantage of approximately 15x at this operating point. Numerical precision is validated against MATLAB float32 reference outputs from a Field II cyst phantom simulation, achieving a 100% pass rate with a root mean square error of 5.10x10^-7, confirming near-theoretical finite-precision behaviour without systematic bias.
High-fidelity TCAD simulation of drift-diffusion transport remains the workhorse of emerging FinFET device design, but it is computationally expensive, especially for 3D structures where runtime escalates steeply with mesh complexity. This sharply limits multi-objective design space exploration. Existing machine-learning surrogates map a fixed set of design parameters to a few scalar device metrics, discarding the underlying physics and losing transferability across device geometries and families. A physics-informed graph attention network (GAT) surrogate is proposed. It operates directly on the tetrahedral TCAD mesh and predicts, at every mesh node, the electrostatic potential together with the electron and hole quasi-Fermi levels, the fundamental unknowns of the drift-diffusion system. Training combines a data loss with finite-volume current-continuity residuals, embedding carrier-transport physics into the objective. Operating on the mesh as a graph, the surrogate inherits size generalization: a model trained on few-fin meshes applies unchanged to substantially larger arrays, bounded at inference only by GPU memory. Per-node uncertainty from a deep ensemble drives an active-learning loop that screens large candidate pools in seconds and forwards only the most informative designs for full simulation. Benchmarked against Sentaurus Device on multi-fin tri-gate FinFETs, the surrogate reproduces the three drift-diffusion fields with sub-volt per-field RMSE and reaches a per-design throughput orders of magnitude higher than the full simulator. The advantage grows with device size: on large multi-fin arrays that are prohibitively slow to simulate directly, inference still completes in under a second per device, enabling Pareto-front exploration across device scales infeasible for direct TCAD sweeps.
展开完整摘要收起摘要↓
High-fidelity TCAD simulation of drift-diffusion transport remains the workhorse of emerging FinFET device design, but it is computationally expensive, especially for 3D structures where runtime escalates steeply with mesh complexity. This sharply limits multi-objective design space exploration. Existing machine-learning surrogates map a fixed set of design parameters to a few scalar device metrics, discarding the underlying physics and losing transferability across device geometries and families. A physics-informed graph attention network (GAT) surrogate is proposed. It operates directly on the tetrahedral TCAD mesh and predicts, at every mesh node, the electrostatic potential together with the electron and hole quasi-Fermi levels, the fundamental unknowns of the drift-diffusion system. Training combines a data loss with finite-volume current-continuity residuals, embedding carrier-transport physics into the objective. Operating on the mesh as a graph, the surrogate inherits size generalization: a model trained on few-fin meshes applies unchanged to substantially larger arrays, bounded at inference only by GPU memory. Per-node uncertainty from a deep ensemble drives an active-learning loop that screens large candidate pools in seconds and forwards only the most informative designs for full simulation. Benchmarked against Sentaurus Device on multi-fin tri-gate FinFETs, the surrogate reproduces the three drift-diffusion fields with sub-volt per-field RMSE and reaches a per-design throughput orders of magnitude higher than the full simulator. The advantage grows with device size: on large multi-fin arrays that are prohibitively slow to simulate directly, inference still completes in under a second per device, enabling Pareto-front exploration across device scales infeasible for direct TCAD sweeps.
Visual token pruning reduces the inference cost of vision-language models (VLMs), but most methods only ask which tokens to keep. This retained-token view can keep redundant high-scoring tokens while leaving discarded evidence without a close representative. We propose CoverPruner, a training-free pruner that asks the complementary demand-side question: after a token is removed, which surviving original token represents it for the target VLM? CoverPruner formulates pruning as Representational Coverage Maximization (RCM), covering the full projected visual-token set with query-weighted demand. It instantiates RCM with projector-space coverage and a lightweight first-layer attention probe. Across multiple VLM architectures and compression rates, CoverPruner achieves the best average accuracy among all compared methods, with the largest gains usually appearing under aggressive compression.
展开完整摘要收起摘要↓
Visual token pruning reduces the inference cost of vision-language models (VLMs), but most methods only ask which tokens to keep. This retained-token view can keep redundant high-scoring tokens while leaving discarded evidence without a close representative. We propose CoverPruner, a training-free pruner that asks the complementary demand-side question: after a token is removed, which surviving original token represents it for the target VLM? CoverPruner formulates pruning as Representational Coverage Maximization (RCM), covering the full projected visual-token set with query-weighted demand. It instantiates RCM with projector-space coverage and a lightweight first-layer attention probe. Across multiple VLM architectures and compression rates, CoverPruner achieves the best average accuracy among all compared methods, with the largest gains usually appearing under aggressive compression.
On-device LLM inference is attractive for privacy and responsiveness, but remains challenging on mobile and embedded devices because model weights far exceed available DRAM. Prior systems exploit activation sparsity and offload weights to SSD or flash storage, but face a fundamental systems trade-off: accurate sparse execution decisions require the latest context, whereas efficient computation-I/O overlap requires early prediction. As a result, existing designs either serialize execution or incur redundant weight fetches, extra computation, and large cache overheads. We present LeanStream, a streaming speculate-and-refine framework for efficient on-device LLM inference. LeanStream progressively refines computation, loading, and cache-retention priorities using partial GPU results, enabling fine-grained overlap between GPU execution and storage I/O. We implement LeanStream on both mobile and embedded platforms. Compared with prior on-device LLM inference systems, LeanStream reduces memory usage by 4.8$\times$ to 7.5$\times$ at the best throughput achieved by prior work, while further improving token generation throughput by 1.6$\times$ to 2.1$\times$.
展开完整摘要收起摘要↓
On-device LLM inference is attractive for privacy and responsiveness, but remains challenging on mobile and embedded devices because model weights far exceed available DRAM. Prior systems exploit activation sparsity and offload weights to SSD or flash storage, but face a fundamental systems trade-off: accurate sparse execution decisions require the latest context, whereas efficient computation-I/O overlap requires early prediction. As a result, existing designs either serialize execution or incur redundant weight fetches, extra computation, and large cache overheads. We present LeanStream, a streaming speculate-and-refine framework for efficient on-device LLM inference. LeanStream progressively refines computation, loading, and cache-retention priorities using partial GPU results, enabling fine-grained overlap between GPU execution and storage I/O. We implement LeanStream on both mobile and embedded platforms. Compared with prior on-device LLM inference systems, LeanStream reduces memory usage by 4.8$\times$ to 7.5$\times$ at the best throughput achieved by prior work, while further improving token generation throughput by 1.6$\times$ to 2.1$\times$.
作者Xiao Shi Huang, Chen-Yuan Lin, Bruce Kuwahara, Kin Kwan Leung, Jesse C. Cresswell
Many NLP tasks, such as summarization and extractive question answering, reduce to retrieving relevant content from documents under two constraints: coverage, retaining enough pertinent information to achieve some goal, and conciseness, removing as much irrelevant information as possible. Conformal prediction methods have been used to guarantee coverage, and must be optimized for conciseness through design of a score function. State-of-the-art scoring functions use hand-engineered LLM prompts asking the model to rate the importance of content, but manual prompt engineering is labor-intensive and task-specific. We introduce the Conformal Relevance framework which uses in-context learning example curation and ensembling to create a score function which maintains coverage while improving conciseness with minimal manual input. We demonstrate this framework's application on seven NLP tasks, and also theoretically study the impact of diversity for ensembled conformal scores, giving a complementarity condition that characterizes when ensembling improves worst-case sentence scores, and a saturation bound on ensemble improvement.
展开完整摘要收起摘要↓
Many NLP tasks, such as summarization and extractive question answering, reduce to retrieving relevant content from documents under two constraints: coverage, retaining enough pertinent information to achieve some goal, and conciseness, removing as much irrelevant information as possible. Conformal prediction methods have been used to guarantee coverage, and must be optimized for conciseness through design of a score function. State-of-the-art scoring functions use hand-engineered LLM prompts asking the model to rate the importance of content, but manual prompt engineering is labor-intensive and task-specific. We introduce the Conformal Relevance framework which uses in-context learning example curation and ensembling to create a score function which maintains coverage while improving conciseness with minimal manual input. We demonstrate this framework's application on seven NLP tasks, and also theoretically study the impact of diversity for ensembled conformal scores, giving a complementarity condition that characterizes when ensembling improves worst-case sentence scores, and a saturation bound on ensemble improvement.
Hybrid architectures combining Full Attention (FA) and Linear Attention (LA) are increasingly prominent, yet their allocation remains heuristic. We seek an evidence-grounded basis in head-level functional organization learned by RoPE-based Transformers. Behavioral probes do not yield a complete taxonomy, so we propose two intervention metrics: RoPE Frequency Importance Score (RFIS), measuring how each frequency affects a head's attention distribution, and RoPE Positional Dependence (RPD), isolating dependence on rotary positional modulation. On Qwen3-series models and Llama3.1, RFIS suggests and RPD verifies a complete taxonomy of retrieval and positional heads separated by a salient mid-low-frequency band. Controlled Transformers show that this boundary follows the training-length positional scale; we term it the Global Positional Band (GPBand). The analysis suggests a potential cause of zero-shot length-extrapolation failure and yields two principles: positional modeling should operate only locally, with global access through position-independent retrieval; and both functions should be assigned at head granularity with layer-specific allocation. We instantiate them in Head-wise Hybrid Architecture (HwH), using NoPE FA for global retrieval and LA for local positional modeling. With an FA-to-LA ratio below 1:3, HwH retains strong language modeling and commonsense reasoning while improving retrieval and substantially strengthening zero-shot long-context extrapolation over Transformer, LA, and a layer-wise hybrid baseline. Ablations validate both principles and component roles, highlighting principled hybrid architecture design as a promising route toward future foundation models.
展开完整摘要收起摘要↓
Hybrid architectures combining Full Attention (FA) and Linear Attention (LA) are increasingly prominent, yet their allocation remains heuristic. We seek an evidence-grounded basis in head-level functional organization learned by RoPE-based Transformers. Behavioral probes do not yield a complete taxonomy, so we propose two intervention metrics: RoPE Frequency Importance Score (RFIS), measuring how each frequency affects a head's attention distribution, and RoPE Positional Dependence (RPD), isolating dependence on rotary positional modulation. On Qwen3-series models and Llama3.1, RFIS suggests and RPD verifies a complete taxonomy of retrieval and positional heads separated by a salient mid-low-frequency band. Controlled Transformers show that this boundary follows the training-length positional scale; we term it the Global Positional Band (GPBand). The analysis suggests a potential cause of zero-shot length-extrapolation failure and yields two principles: positional modeling should operate only locally, with global access through position-independent retrieval; and both functions should be assigned at head granularity with layer-specific allocation. We instantiate them in Head-wise Hybrid Architecture (HwH), using NoPE FA for global retrieval and LA for local positional modeling. With an FA-to-LA ratio below 1:3, HwH retains strong language modeling and commonsense reasoning while improving retrieval and substantially strengthening zero-shot long-context extrapolation over Transformer, LA, and a layer-wise hybrid baseline. Ablations validate both principles and component roles, highlighting principled hybrid architecture design as a promising route toward future foundation models.
作者Ahmed Asaad, Amr Mohamed, Yang Zhang, Omneya Abdelsalam
Large Language Models (LLMs) increasingly use user context such as memory, profiles, and role prompts to personalize their responses. This personalization can affect evidence-based judgment: the same evidence may lead to different conclusions under different user contexts. Finance provides a high-stakes setting to study this problem because decisions often depend on interpreting long and complex documents. We test this using 3,575 SEC filings across twelve LLMs. We compare persona-conditioned retrieval, neutral retrieval, and memory-framed context to separate the effect of evidence selection from the effect of interpretation. We find that most user-context spillover comes from how models interpret the same evidence under different roles, rather than from retrieving different evidence. We then test two simple mitigation strategies: expressing the same investor mindset as a user profile instead of an assistant role, and separating evidence-based and personalized outputs. Both reduce spillover, but neither removes it completely, and their effectiveness varies substantially across models.
展开完整摘要收起摘要↓
Large Language Models (LLMs) increasingly use user context such as memory, profiles, and role prompts to personalize their responses. This personalization can affect evidence-based judgment: the same evidence may lead to different conclusions under different user contexts. Finance provides a high-stakes setting to study this problem because decisions often depend on interpreting long and complex documents. We test this using 3,575 SEC filings across twelve LLMs. We compare persona-conditioned retrieval, neutral retrieval, and memory-framed context to separate the effect of evidence selection from the effect of interpretation. We find that most user-context spillover comes from how models interpret the same evidence under different roles, rather than from retrieving different evidence. We then test two simple mitigation strategies: expressing the same investor mindset as a user profile instead of an assistant role, and separating evidence-based and personalized outputs. Both reduce spillover, but neither removes it completely, and their effectiveness varies substantially across models.
作者Alessandro Schena, Ilgiz Mustafin, Julia Kotovich
This article introduces Eiffel-tools, a language server protocol (LSP) implementation for the Eiffel programming language that uses Large Language Models (LLMs) to aid the development of statically verified software. The tool provides various interactive and non-interactive commands to produce code and specifications. It uses language and project specific knowledge to precisely direct the LLM and verifies the output using a static verifier. It crafts rich programmatic prompts for the input and corrects or rejects the output. Furthermore, it handles the retries until the program passes verification. The tool's bug fixing capability is evaluated on 2 public datasets using 3 models. The tool can fix 76% to 95% of bugs by combining LLMs and a formal verifier depending on the model and prompts used. The results show the trade-off between the number of fixing attempts and the success rate.
展开完整摘要收起摘要↓
This article introduces Eiffel-tools, a language server protocol (LSP) implementation for the Eiffel programming language that uses Large Language Models (LLMs) to aid the development of statically verified software. The tool provides various interactive and non-interactive commands to produce code and specifications. It uses language and project specific knowledge to precisely direct the LLM and verifies the output using a static verifier. It crafts rich programmatic prompts for the input and corrects or rejects the output. Furthermore, it handles the retries until the program passes verification. The tool's bug fixing capability is evaluated on 2 public datasets using 3 models. The tool can fix 76% to 95% of bugs by combining LLMs and a formal verifier depending on the model and prompts used. The results show the trade-off between the number of fixing attempts and the success rate.
作者Xiang Fu, Seungmin Cho, Yukyung Lee, Najoung Kim
Large language models (LLMs) exhibit in-context learning capabilities, where they can learn new tasks from prompt contexts without weight updates. We compare the learning efficacies of two prominent modes of in-context learning: (1) learning from descriptions of rules (instruction following); and (2) learning from examples of input-output demonstrations (few-shot prompting). Through five learning tasks that cover diverse domains (games, arithmetic, linguistic inferences), we compare two modes of learning (rules vs. examples) specifying the same underlying task. We furthermore explore model and task properties that modulate the learning efficacies. We find that models generally learn more reliably from rules than from examples alone, and additional examples on top of rules or simply scaling up the number of examples do not lead to consistent and significant gains. Instruction tuning amplifies the benefit of rule-based learning while keeping example-based learning capacities intact. Surprisingly, we find no privileged effect of example-based learning in base models, and rules still lead to gains in algebraic task domains. Overall, the comparative efficacy of rules over examples is larger when the task recruits algebraic abstractions and computations, and smaller when the task requires distributional sensitivity and/or recruits parametric knowledge.
展开完整摘要收起摘要↓
Large language models (LLMs) exhibit in-context learning capabilities, where they can learn new tasks from prompt contexts without weight updates. We compare the learning efficacies of two prominent modes of in-context learning: (1) learning from descriptions of rules (instruction following); and (2) learning from examples of input-output demonstrations (few-shot prompting). Through five learning tasks that cover diverse domains (games, arithmetic, linguistic inferences), we compare two modes of learning (rules vs. examples) specifying the same underlying task. We furthermore explore model and task properties that modulate the learning efficacies. We find that models generally learn more reliably from rules than from examples alone, and additional examples on top of rules or simply scaling up the number of examples do not lead to consistent and significant gains. Instruction tuning amplifies the benefit of rule-based learning while keeping example-based learning capacities intact. Surprisingly, we find no privileged effect of example-based learning in base models, and rules still lead to gains in algebraic task domains. Overall, the comparative efficacy of rules over examples is larger when the task recruits algebraic abstractions and computations, and smaller when the task requires distributional sensitivity and/or recruits parametric knowledge.
Large language models (LLMs) are increasingly used in software engineering pipelines for code generation, where production prompts often combine multiple constraints. This paper presents a full-factorial empirical study of how output formatting, persona assignment, and urgency framing jointly affect LLM code-generation reliability. We evaluate all 27 combinations in a controlled 3x3x3 design and decompose each compound condition into an additive prediction and a residual interaction term that captures super-additive degradation. The study uses all 164 HumanEval+ problems across five OpenAI models from the GPT-4o family, GPT-4.1 family, and o3-mini, yielding 22,140 greedy-decoding evaluations. A format-aware extraction pipeline separates formatting failures from reasoning failures, and significance is assessed with McNemar's test, odds ratios, and 95% confidence intervals. Results show that compound constraints can produce architecture-dependent degradation not predictable from single-factor experiments. The GPT-4o family exhibits consistent super-additive effects, with pass@1 reductions 3-12 percentage points beyond additive predictions; the largest interaction is -12.2 pp on GPT-4o-mini for JSON + expert persona + moderate urgency. JSON combinations generally produce larger interactions than XML. In contrast, the GPT-4.1 family is largely resistant, while o3-mini shows a qualitatively different pattern in which structured output constraints can improve performance. These findings show that vulnerability is architecture-dependent rather than size-dependent, that individually neutral or beneficial constraints can combine to cause substantial degradation, and that compound-prompt testing should be standard in reliability assessment for LLM-assisted engineering pipelines.
展开完整摘要收起摘要↓
Large language models (LLMs) are increasingly used in software engineering pipelines for code generation, where production prompts often combine multiple constraints. This paper presents a full-factorial empirical study of how output formatting, persona assignment, and urgency framing jointly affect LLM code-generation reliability. We evaluate all 27 combinations in a controlled 3x3x3 design and decompose each compound condition into an additive prediction and a residual interaction term that captures super-additive degradation. The study uses all 164 HumanEval+ problems across five OpenAI models from the GPT-4o family, GPT-4.1 family, and o3-mini, yielding 22,140 greedy-decoding evaluations. A format-aware extraction pipeline separates formatting failures from reasoning failures, and significance is assessed with McNemar's test, odds ratios, and 95% confidence intervals. Results show that compound constraints can produce architecture-dependent degradation not predictable from single-factor experiments. The GPT-4o family exhibits consistent super-additive effects, with pass@1 reductions 3-12 percentage points beyond additive predictions; the largest interaction is -12.2 pp on GPT-4o-mini for JSON + expert persona + moderate urgency. JSON combinations generally produce larger interactions than XML. In contrast, the GPT-4.1 family is largely resistant, while o3-mini shows a qualitatively different pattern in which structured output constraints can improve performance. These findings show that vulnerability is architecture-dependent rather than size-dependent, that individually neutral or beneficial constraints can combine to cause substantial degradation, and that compound-prompt testing should be standard in reliability assessment for LLM-assisted engineering pipelines.
Frontier large language models (LLMs) have become attractive priors for optimization due to their large-scale pretraining that enables them to navigate a variety of optimization settings. However, the effectiveness of modern reasoning LLMs in batch optimization settings remains underexplored. Here we investigate the performance of the current generation of frontier LLMs as batch optimizers in both continuous and discrete settings. We find that while LLMs are competitive zero-shot batch optimizers for numerical test functions, their performance is brittle compared to classical non-LLM optimization approaches. However, LLM priors are significantly better in semantically rich settings, indicating that their batch optimization behavior is highly effective when navigating and reasoning over the discrete spaces most similar in structure to their pretraining data.
展开完整摘要收起摘要↓
Frontier large language models (LLMs) have become attractive priors for optimization due to their large-scale pretraining that enables them to navigate a variety of optimization settings. However, the effectiveness of modern reasoning LLMs in batch optimization settings remains underexplored. Here we investigate the performance of the current generation of frontier LLMs as batch optimizers in both continuous and discrete settings. We find that while LLMs are competitive zero-shot batch optimizers for numerical test functions, their performance is brittle compared to classical non-LLM optimization approaches. However, LLM priors are significantly better in semantically rich settings, indicating that their batch optimization behavior is highly effective when navigating and reasoning over the discrete spaces most similar in structure to their pretraining data.
Nastase et al. (2026) argue that large language models (LLMs) may illuminate language processing because both rely on distributed, context-sensitive representations shaped by statistical learning. Their rejection of simple cortical "boxology" is persuasive, and they articulate a strong case for the value of LLM-brain alignment research. The key question is what kind of inference LLM-brain alignment licenses. My claim here will be narrow: representational alignment can in principle constrain mechanistic hypotheses, but it does not by itself identify a mechanism. Nastase et al. acknowledge that an encoding model can capture features represented in neural activity without establishing a shared architecture or algorithm. Yet the authors sometime move from alignment to "shared computational principles" and ultimately to LLMs as mechanistic models of natural language. Indeed, their methodological caveat that alignment does not establish a shared architecture or algorithm sits uneasily with their conclusion that LLMs might instantiate the same computational principles as biological brains and provide a "fully mechanistic model" of language. I discuss what I consider to be problems of logical, causal, and computational underdetermination in Nastase et al.'s (2026) proposal.
展开完整摘要收起摘要↓
Nastase et al. (2026) argue that large language models (LLMs) may illuminate language processing because both rely on distributed, context-sensitive representations shaped by statistical learning. Their rejection of simple cortical "boxology" is persuasive, and they articulate a strong case for the value of LLM-brain alignment research. The key question is what kind of inference LLM-brain alignment licenses. My claim here will be narrow: representational alignment can in principle constrain mechanistic hypotheses, but it does not by itself identify a mechanism. Nastase et al. acknowledge that an encoding model can capture features represented in neural activity without establishing a shared architecture or algorithm. Yet the authors sometime move from alignment to "shared computational principles" and ultimately to LLMs as mechanistic models of natural language. Indeed, their methodological caveat that alignment does not establish a shared architecture or algorithm sits uneasily with their conclusion that LLMs might instantiate the same computational principles as biological brains and provide a "fully mechanistic model" of language. I discuss what I consider to be problems of logical, causal, and computational underdetermination in Nastase et al.'s (2026) proposal.
Long-context reasoning for large language models (LLMs) is becoming increasingly important, but training over long sequences remains challenging due to massive memory and communication requirements. Sequence parallelism has emerged as an essential technique for addressing bottlenecks in long sequence LLM training. However, we observe that existing sequence parallelism methods are batch-agnostic and apply uniform sequence partitioning across all batch sizes, resulting in inefficient communication. In this paper, we introduce Batch- Aware Sequence Parallelism (BASP), a sequence parallelism approach that leverages batch structure to reduce communication overhead. BASP exploits batch structure by partitioning GPUs into disjoint sequence-parallel groups according to the micro- batch size. This design reduces the all-to-all communication group size, thereby localizing communication and improving training efficiency. Experimental results on an NVIDIA A100 cluster show that BASP improves end-to-end training time by up to 1.17 - 1.31x in Llama and Qwen models compared to standard sequence parallel baselines, while preserving identical model accuracy and memory usage.
展开完整摘要收起摘要↓
Long-context reasoning for large language models (LLMs) is becoming increasingly important, but training over long sequences remains challenging due to massive memory and communication requirements. Sequence parallelism has emerged as an essential technique for addressing bottlenecks in long sequence LLM training. However, we observe that existing sequence parallelism methods are batch-agnostic and apply uniform sequence partitioning across all batch sizes, resulting in inefficient communication. In this paper, we introduce Batch- Aware Sequence Parallelism (BASP), a sequence parallelism approach that leverages batch structure to reduce communication overhead. BASP exploits batch structure by partitioning GPUs into disjoint sequence-parallel groups according to the micro- batch size. This design reduces the all-to-all communication group size, thereby localizing communication and improving training efficiency. Experimental results on an NVIDIA A100 cluster show that BASP improves end-to-end training time by up to 1.17 - 1.31x in Llama and Qwen models compared to standard sequence parallel baselines, while preserving identical model accuracy and memory usage.