Post-training quantization (PTQ) methods typically use sequential quantization that partitions a pre-trained LLM into a series of units (e.g., transformer blocks), with one unit quantized at each stage. State-of-the-art PTQ methods are predominantly learning-based, optimizing auxiliary quantization parameters (e.g., scaling factors, rotation matrices, clipping thresholds, and adapters) via gradient descent to minimize a reconstruction loss. A common practice is to use mean squared error (MSE) as the reconstruction loss function, yet its induced optimization behavior remains largely unexplored. In this work, we take a holistic view of sequential quantization and systematically investigate how optimization evolves from the first quantization stage to the last, aiming for a deep understanding of optimization in learning-based PTQ schemes. Through extensive empirical studies spanning representative learning-based PTQ methods, LLM families, model scales, architectures, quantization settings and various tasks, we consistently uncover Optimization Imbalance: reconstruction loss magnitudes vary dramatically across stages, accompanied by highly uneven gradient magnitudes and parameter updates under MSE. We term the cross-stage range of loss magnitudes the reconstruction loss scale, and reveal that MSE translates the unexpectedly large reconstruction loss scale into highly uneven gradient magnitudes, which in turn lead to uneven optimization strength across quantization stages. This finding suggests a general principle for improving learning-based PTQ: optimization strength across stages should be decoupled from the reconstruction loss scale. Theoretically, we show that root mean squared error (RMSE) variants defined at the sample, channel, token, and element levels naturally realize this principle through implicit gradient normalization, outperforming MSE significantly as a drop-in replacement.
展开完整摘要收起摘要↓
Post-training quantization (PTQ) methods typically use sequential quantization that partitions a pre-trained LLM into a series of units (e.g., transformer blocks), with one unit quantized at each stage. State-of-the-art PTQ methods are predominantly learning-based, optimizing auxiliary quantization parameters (e.g., scaling factors, rotation matrices, clipping thresholds, and adapters) via gradient descent to minimize a reconstruction loss. A common practice is to use mean squared error (MSE) as the reconstruction loss function, yet its induced optimization behavior remains largely unexplored. In this work, we take a holistic view of sequential quantization and systematically investigate how optimization evolves from the first quantization stage to the last, aiming for a deep understanding of optimization in learning-based PTQ schemes. Through extensive empirical studies spanning representative learning-based PTQ methods, LLM families, model scales, architectures, quantization settings and various tasks, we consistently uncover Optimization Imbalance: reconstruction loss magnitudes vary dramatically across stages, accompanied by highly uneven gradient magnitudes and parameter updates under MSE. We term the cross-stage range of loss magnitudes the reconstruction loss scale, and reveal that MSE translates the unexpectedly large reconstruction loss scale into highly uneven gradient magnitudes, which in turn lead to uneven optimization strength across quantization stages. This finding suggests a general principle for improving learning-based PTQ: optimization strength across stages should be decoupled from the reconstruction loss scale. Theoretically, we show that root mean squared error (RMSE) variants defined at the sample, channel, token, and element levels naturally realize this principle through implicit gradient normalization, outperforming MSE significantly as a drop-in replacement.
Reducing LLM serving energy does not by itself guarantee lower deployment cost when electricity procurement exposes operators to unfavorable deviations from preset commitments. We study hourly commitments with positive, potentially asymmetric costs for overuse and underuse, and formulate energy-Performance-Aware Commitment Tracking: minimize deviation costs subject to request-level service requirements. We implement ePACT, a two-level controller that adjusts serving capacity and GPU clocks as requests arrive. A global planner updates interval energy targets from measured consumption and the remaining hourly commitment. A local decision maker predicts candidate configurations' energy and completion times, checks predicted deadline misses, and selects among admitted configurations by asymmetric target-deviation cost, with a service-first fallback. Coarse-to-fine action search runs asynchronously with serving. We evaluate ePACT through single-hour comparisons, controller ablations, and full-day trace simulations for H20 and H200 GPU pools. In the 24-hour simulations, ePACT reduces the asymmetric deviation cost by $73.8%$ and $75.7%$ relative to vLLM while retaining near-vLLM SLO attainment. Mean absolute hourly deviations are $2.16%$ and $2.31%$, respectively.
展开完整摘要收起摘要↓
Reducing LLM serving energy does not by itself guarantee lower deployment cost when electricity procurement exposes operators to unfavorable deviations from preset commitments. We study hourly commitments with positive, potentially asymmetric costs for overuse and underuse, and formulate energy-Performance-Aware Commitment Tracking: minimize deviation costs subject to request-level service requirements. We implement ePACT, a two-level controller that adjusts serving capacity and GPU clocks as requests arrive. A global planner updates interval energy targets from measured consumption and the remaining hourly commitment. A local decision maker predicts candidate configurations' energy and completion times, checks predicted deadline misses, and selects among admitted configurations by asymmetric target-deviation cost, with a service-first fallback. Coarse-to-fine action search runs asynchronously with serving. We evaluate ePACT through single-hour comparisons, controller ablations, and full-day trace simulations for H20 and H200 GPU pools. In the 24-hour simulations, ePACT reduces the asymmetric deviation cost by $73.8%$ and $75.7%$ relative to vLLM while retaining near-vLLM SLO attainment. Mean absolute hourly deviations are $2.16%$ and $2.31%$, respectively.
作者Boyang Li, Bingyu Shen, Weihao Hong, Zhiyuan Jiang, Xinlei Guan, Yan Ma, Miles Q. Li, Yi Sheng, Ruiyang Qin
Quantized large language models are increasingly deployed on edge devices for their low latency and energy efficiency. However, model quantization weakens alignment safeguards, leaving qLLMs (quantized large language models) highly vulnerable to jailbreak attacks. To address this challenge, we present MOMAT (Mixture of Multiple Atlases), a hardware-enhanced safety framework that combines structured knowledge retrieval with low-power defense acceleration. Each atlas represents a semantic cluster of harmful or benign sample sets and policy templates, enabling domain-localized Retrieval-Augmented Generation guarding that mitigates the curse of dimensionality and the resulting semantic sparsity problem in large, heterogeneous safety databases. MOMAT retrieves top-$k$ similarity features from all atlases for each prompt and evaluates them using a lightweight MoE (Mixture of Experts) detector, while a CiM (Compute-in-Memory)-accelerated similarity engine performs fast, low-power atlas-local retrieval. MOMAT's CiM-based retrieval accelerates a 100-query batch from 15,052.44 ms to 3,207.21 ns (a $4.69 \times 10^6\times$ speedup) and reduces energy from $8.1 \times 10^7$ $μ$J to 3.32 $μ$J, yielding an approximately $2.5 \times 10^5\times$ energy reduction over DRAM-based (Raspberry Pi) baselines. Red-team evaluations across standard benchmarks show that MOMAT matches the defense performance of state-of-the-art methods while avoiding benign overkill and providing substantial efficiency gains, demonstrating that CiM-based modular defenses can make edge-deployed qLLMs both safer and more energy-efficient. We will release the full 223.2k-sample dataset to foster future research.
展开完整摘要收起摘要↓
Quantized large language models are increasingly deployed on edge devices for their low latency and energy efficiency. However, model quantization weakens alignment safeguards, leaving qLLMs (quantized large language models) highly vulnerable to jailbreak attacks. To address this challenge, we present MOMAT (Mixture of Multiple Atlases), a hardware-enhanced safety framework that combines structured knowledge retrieval with low-power defense acceleration. Each atlas represents a semantic cluster of harmful or benign sample sets and policy templates, enabling domain-localized Retrieval-Augmented Generation guarding that mitigates the curse of dimensionality and the resulting semantic sparsity problem in large, heterogeneous safety databases. MOMAT retrieves top-$k$ similarity features from all atlases for each prompt and evaluates them using a lightweight MoE (Mixture of Experts) detector, while a CiM (Compute-in-Memory)-accelerated similarity engine performs fast, low-power atlas-local retrieval. MOMAT's CiM-based retrieval accelerates a 100-query batch from 15,052.44 ms to 3,207.21 ns (a $4.69 \times 10^6\times$ speedup) and reduces energy from $8.1 \times 10^7$ $μ$J to 3.32 $μ$J, yielding an approximately $2.5 \times 10^5\times$ energy reduction over DRAM-based (Raspberry Pi) baselines. Red-team evaluations across standard benchmarks show that MOMAT matches the defense performance of state-of-the-art methods while avoiding benign overkill and providing substantial efficiency gains, demonstrating that CiM-based modular defenses can make edge-deployed qLLMs both safer and more energy-efficient. We will release the full 223.2k-sample dataset to foster future research.
Retrieval-based speculative decoding (SD) drafts tokens by copying continuations from existing text, which suits coding agents that repeatedly reproduce code, logs, and earlier attempts. Yet existing methods fall short in agent pipelines: much of the reusable text is missing from their corpora or stored in a form that differs from what the agent emits, and their draft lengths ignore that accept length varies across agents and drifts over turns. We present AgSpec, a framework that supplies the corpus and draft-length policies that existing retrieval engines lack in coding-agent pipelines. AgSpec retrieves from session, workspace, and global corpora, retaining the ongoing session trajectory and indexing opened files in the agent's emission format. It bounds each agent's draft length with an offline-profiled cap and adapts the length online from verification feedback. On two repository-level multi-agent coding benchmarks, AgSpec outperforms five retrieval-based drafters and EAGLE-3 in most evaluated settings, raising generation throughput over autoregressive decoding up to 4.37$\times$ at batch size 1 and 4.76$\times$ at batch size 16. AgSpec also remains effective on benchmarks without a repository or a multi-agent pipeline, showing that its gains generalize to coding agents broadly.
展开完整摘要收起摘要↓
Retrieval-based speculative decoding (SD) drafts tokens by copying continuations from existing text, which suits coding agents that repeatedly reproduce code, logs, and earlier attempts. Yet existing methods fall short in agent pipelines: much of the reusable text is missing from their corpora or stored in a form that differs from what the agent emits, and their draft lengths ignore that accept length varies across agents and drifts over turns. We present AgSpec, a framework that supplies the corpus and draft-length policies that existing retrieval engines lack in coding-agent pipelines. AgSpec retrieves from session, workspace, and global corpora, retaining the ongoing session trajectory and indexing opened files in the agent's emission format. It bounds each agent's draft length with an offline-profiled cap and adapts the length online from verification feedback. On two repository-level multi-agent coding benchmarks, AgSpec outperforms five retrieval-based drafters and EAGLE-3 in most evaluated settings, raising generation throughput over autoregressive decoding up to 4.37$\times$ at batch size 1 and 4.76$\times$ at batch size 16. AgSpec also remains effective on benchmarks without a repository or a multi-agent pipeline, showing that its gains generalize to coding agents broadly.
LLM agents revise running tasks when users change instructions, tools fail, or new information changes a plan. Today's servers express a revision as aborting old requests and submitting replacements. Yet the old execution's buffered output and outstanding work must stop affecting the application, while completed KV state may still be useful to its replacement. Handling these obligations separately can leave obsolete effects publishable and force the successor to rebuild valid state. We present \retire, a serving control-plane redesign around versioned execution. Requests own scheduling and memory resources; execution versions own authority, the permission to publish output or install state for the current execution. \retire first revokes obsolete work, then bounds its remaining execution and certifies the completed prefix its successor can inherit. The successor runs from that state while isolated old resources are reclaimed asynchronously. This unifies fast invalidation and selective preservation in one version transition. We implement \retire in vLLM across output publication, GPU execution, KV handoff, tiered recovery, and distributed and multi-tenant serving. Correctness experiments verify current-version output and valid state inheritance across these paths. Combining invalidation with inheritance reduces revision-to-successor time-to-first-token by a median 17.1% in controlled paired experiments. A replay of recorded coding-agent interruption arrivals emits no obsolete output and keeps every final version progressing through repeated revisions. \retire turns abort-and-restart into a coordinated handoff that stops obsolete work quickly and preserves useful work for its successor.
展开完整摘要收起摘要↓
LLM agents revise running tasks when users change instructions, tools fail, or new information changes a plan. Today's servers express a revision as aborting old requests and submitting replacements. Yet the old execution's buffered output and outstanding work must stop affecting the application, while completed KV state may still be useful to its replacement. Handling these obligations separately can leave obsolete effects publishable and force the successor to rebuild valid state. We present \retire, a serving control-plane redesign around versioned execution. Requests own scheduling and memory resources; execution versions own authority, the permission to publish output or install state for the current execution. \retire first revokes obsolete work, then bounds its remaining execution and certifies the completed prefix its successor can inherit. The successor runs from that state while isolated old resources are reclaimed asynchronously. This unifies fast invalidation and selective preservation in one version transition. We implement \retire in vLLM across output publication, GPU execution, KV handoff, tiered recovery, and distributed and multi-tenant serving. Correctness experiments verify current-version output and valid state inheritance across these paths. Combining invalidation with inheritance reduces revision-to-successor time-to-first-token by a median 17.1% in controlled paired experiments. A replay of recorded coding-agent interruption arrivals emits no obsolete output and keeps every final version progressing through repeated revisions. \retire turns abort-and-restart into a coordinated handoff that stops obsolete work quickly and preserves useful work for its successor.
作者Wenxun Wang, Likai Ma, Zongle Huang, Chen Tang, Yongpan Liu
The widespread adoption of Mixture-of-Experts (MoE) has created a growing need for deployment on heterogeneous platforms. However, it exposes a fundamental mismatch between the algorithmic demands of large-scale MoE and the disparate characteristics of hardware.Existing CPU-GPU hybrid inference systems fail to resolve this as they either encounter PCIe bandwidth bottlenecks when loading experts to GPUs, or rely heavily on CPU computation. Consequently, this leads to low resource utilization and inevitable violations of fixed latency budgets as parameters scale. In this paper, we identify and exploit Cross-Asymmetry--a structural alignment between the algorithmic workload skew of MoE routing and the physical disparity of heterogeneous hardware. To this end, we introduce RapidMoE, a residual offloading system for efficient large-scale MoE inference. We propose how RapidMoE leverages a residual-split framework to enable offloading paradigm shift from expert-level to bit-level, which unfolds across three key dimensions: (1) data representation, enabling compact and decoupled storage; (2) routing strategy, partitioning computation into dual paths aligned with hardware capabilities; (3) execution parallelism, scheduling a balanced storage-compute workload across devices. We further employ a novel Unified Multi-Level Importance Arbitration to adaptively adjust the critical expert set at runtime, ensuring the accuracy-latency Pareto frontier. These innovations exploit inherent cross-asymmetry, fundamentally breaking the algorithm-hardware misalignment. Experimental results show that RapidMoE achieves up to 3.5x speedup in decoding and 2.1x speedup in prefill compared to state-of-the-art (SOTA) offloading systems.
展开完整摘要收起摘要↓
The widespread adoption of Mixture-of-Experts (MoE) has created a growing need for deployment on heterogeneous platforms. However, it exposes a fundamental mismatch between the algorithmic demands of large-scale MoE and the disparate characteristics of hardware.Existing CPU-GPU hybrid inference systems fail to resolve this as they either encounter PCIe bandwidth bottlenecks when loading experts to GPUs, or rely heavily on CPU computation. Consequently, this leads to low resource utilization and inevitable violations of fixed latency budgets as parameters scale. In this paper, we identify and exploit Cross-Asymmetry--a structural alignment between the algorithmic workload skew of MoE routing and the physical disparity of heterogeneous hardware. To this end, we introduce RapidMoE, a residual offloading system for efficient large-scale MoE inference. We propose how RapidMoE leverages a residual-split framework to enable offloading paradigm shift from expert-level to bit-level, which unfolds across three key dimensions: (1) data representation, enabling compact and decoupled storage; (2) routing strategy, partitioning computation into dual paths aligned with hardware capabilities; (3) execution parallelism, scheduling a balanced storage-compute workload across devices. We further employ a novel Unified Multi-Level Importance Arbitration to adaptively adjust the critical expert set at runtime, ensuring the accuracy-latency Pareto frontier. These innovations exploit inherent cross-asymmetry, fundamentally breaking the algorithm-hardware misalignment. Experimental results show that RapidMoE achieves up to 3.5x speedup in decoding and 2.1x speedup in prefill compared to state-of-the-art (SOTA) offloading systems.
Parallel drafting reduces the drafting overhead of speculative decoding for large language models (LLMs), but its gains remain limited by the accepted prefix length. Even when the correct token is present in the candidate pool, a single early selection error prevents subsequent predictions from being used. We propose DRelay, which uses global information from the entire draft block to perform prefix-aware selective repair of candidate selections before target-model verification. DRelay bases its decisions on candidate correlations and the selected path: a global reader extracts predictive information across positions for each candidate. While a causal selector combines candidate-level information extracted by the global read with the tokens selected at preceding positions to determine whether the native choice at the current position is consistent with the global evidence and the selected prefix. It then decides whether to retain or replace the token, thereby repairing early errors and extending the accepted prefix. We further jointly train the draft backbone and the selector, combining candidate-support learning with a repair objective, while weighting the repair loss according to each block position's potential contribution to the consecutive accepted prefix. Across eight diverse benchmarks on an H800 GPU, DRelay consistently improves both average acceptance length and end-to-end decoding performance over DFlash, Domino, and DSpark. Under SGLang serving, DRelay improves average end-to-end speedup over DFlash, Domino, and DSpark by 14.7%-16.8%, 8.7%-9.3%, and 8.1%-9.3%, respectively.
展开完整摘要收起摘要↓
Parallel drafting reduces the drafting overhead of speculative decoding for large language models (LLMs), but its gains remain limited by the accepted prefix length. Even when the correct token is present in the candidate pool, a single early selection error prevents subsequent predictions from being used. We propose DRelay, which uses global information from the entire draft block to perform prefix-aware selective repair of candidate selections before target-model verification. DRelay bases its decisions on candidate correlations and the selected path: a global reader extracts predictive information across positions for each candidate. While a causal selector combines candidate-level information extracted by the global read with the tokens selected at preceding positions to determine whether the native choice at the current position is consistent with the global evidence and the selected prefix. It then decides whether to retain or replace the token, thereby repairing early errors and extending the accepted prefix. We further jointly train the draft backbone and the selector, combining candidate-support learning with a repair objective, while weighting the repair loss according to each block position's potential contribution to the consecutive accepted prefix. Across eight diverse benchmarks on an H800 GPU, DRelay consistently improves both average acceptance length and end-to-end decoding performance over DFlash, Domino, and DSpark. Under SGLang serving, DRelay improves average end-to-end speedup over DFlash, Domino, and DSpark by 14.7%-16.8%, 8.7%-9.3%, and 8.1%-9.3%, respectively.
Multimodal large language models (MLLMs) incur substantial inference costs when processing long visual-textual sequences. While existing operation compression methods exploit modality-level redundancy, they largely treat computation within attention heads and shared feed-forward network (FFN) channels as unified units, leaving finer-grained redundancy underexplored. We find that redundancy varies both across modality-interaction paths within the same attention head and across visual and textual executions of the same FFN channel. Based on these findings, we propose Modality-aware Width-wise Operation Pruning (MWOP), which independently prunes visual-to-visual (V2V), text-to-visual (T2V), and text-to-text (T2T) attention paths within each layer, and separately selects FFN channels for visual and textual inputs. A first-order Taylor criterion guides the pruning process, with FFN importance re-evaluated after attention pruning and LoRA-based recovery training. To translate the resulting fine-grained sparsity into practical acceleration, we further develop path-sparse Triton attention kernels and compact visual-side FFN execution. MWOP preserves the token sequence while reducing attention and FFN computation, making it complementary to token compression and enabling simultaneous reduction of sequence length and per-token computation. On LLaVA-OneVision-7B, MWOP alone achieves a $1.6\times$ prefill speedup with 99.7% average performance retention across 12 benchmarks. Combined with two representative token compression methods, it further increases their prefill speedups from $2.0\times$ and $1.9\times$ to $2.9\times$ and $2.7\times$, respectively. Results on Qwen2.5-VL-7B further demonstrate its applicability across architectures. The code is available at https://github.com/EIT-NLP/MWOP.
展开完整摘要收起摘要↓
Multimodal large language models (MLLMs) incur substantial inference costs when processing long visual-textual sequences. While existing operation compression methods exploit modality-level redundancy, they largely treat computation within attention heads and shared feed-forward network (FFN) channels as unified units, leaving finer-grained redundancy underexplored. We find that redundancy varies both across modality-interaction paths within the same attention head and across visual and textual executions of the same FFN channel. Based on these findings, we propose Modality-aware Width-wise Operation Pruning (MWOP), which independently prunes visual-to-visual (V2V), text-to-visual (T2V), and text-to-text (T2T) attention paths within each layer, and separately selects FFN channels for visual and textual inputs. A first-order Taylor criterion guides the pruning process, with FFN importance re-evaluated after attention pruning and LoRA-based recovery training. To translate the resulting fine-grained sparsity into practical acceleration, we further develop path-sparse Triton attention kernels and compact visual-side FFN execution. MWOP preserves the token sequence while reducing attention and FFN computation, making it complementary to token compression and enabling simultaneous reduction of sequence length and per-token computation. On LLaVA-OneVision-7B, MWOP alone achieves a $1.6\times$ prefill speedup with 99.7% average performance retention across 12 benchmarks. Combined with two representative token compression methods, it further increases their prefill speedups from $2.0\times$ and $1.9\times$ to $2.9\times$ and $2.7\times$, respectively. Results on Qwen2.5-VL-7B further demonstrate its applicability across architectures. The code is available at https://github.com/EIT-NLP/MWOP.
作者Ke Yang, Yongji Gao, Xushi Li, Kui Luo, Sicheng Zhang, Tianming Zhou, Keyi Liu, Shufang Lu, Aoxuan Chen, Jie Meng, Jingchun Gao, Dan Li, Xinkai You, Dan Li, Zhixiang Xia, Yan Shi, Yang Liu, Yanjia Zeng, Liangjun Feng
Sparse expert activation reduces MoE models' computation, yet expert weights can exceed limited device memory. Offloading makes inference feasible on a compact AI appliance but exposes host-to-device transfers to the inference path. We present MoE-CORE, a system that coordinates expert offloading and residency for memory-constrained MoE inference. It stages complete expert layers in alternating buffers during prefill. During decode, it combines nonuniform layer-wise cache capacity, domain-informed initialization, routing-history-aware replacement, and cross-layer prefetching. The main configuration executes router-selected experts exactly; an optional score-based substitution path handles eligible low-score misses. The main comparison uses 1K- and 128-token output caps for MoE-CORE and vLLM Prefetch, respectively. Across five workloads per model, MoE-CORE records a mean time per output token (TPOT) of 38.0-44.8 ms versus 1268.9-1269.1 ms for the evaluated vLLM Prefetch configuration on DeepSeek-V4-Flash-W4A8; the corresponding values on GLM-5.2-W4A8C8 are 206.6-220.5 and 5941.5-5941.8 ms. Under an 84-GB NPU-memory cap, the best measured DeepSeek GSM8K configuration achieves a TPOT of 21.5 ms with approximate expert substitution and multi-token prediction (MTP) at depth 2. These results support coordinated expert residency and transfer scheduling under a device-memory constraint. The code is here.
展开完整摘要收起摘要↓
Sparse expert activation reduces MoE models' computation, yet expert weights can exceed limited device memory. Offloading makes inference feasible on a compact AI appliance but exposes host-to-device transfers to the inference path. We present MoE-CORE, a system that coordinates expert offloading and residency for memory-constrained MoE inference. It stages complete expert layers in alternating buffers during prefill. During decode, it combines nonuniform layer-wise cache capacity, domain-informed initialization, routing-history-aware replacement, and cross-layer prefetching. The main configuration executes router-selected experts exactly; an optional score-based substitution path handles eligible low-score misses. The main comparison uses 1K- and 128-token output caps for MoE-CORE and vLLM Prefetch, respectively. Across five workloads per model, MoE-CORE records a mean time per output token (TPOT) of 38.0-44.8 ms versus 1268.9-1269.1 ms for the evaluated vLLM Prefetch configuration on DeepSeek-V4-Flash-W4A8; the corresponding values on GLM-5.2-W4A8C8 are 206.6-220.5 and 5941.5-5941.8 ms. Under an 84-GB NPU-memory cap, the best measured DeepSeek GSM8K configuration achieves a TPOT of 21.5 ms with approximate expert substitution and multi-token prediction (MTP) at depth 2. These results support coordinated expert residency and transfer scheduling under a device-memory constraint. The code is here.
As the capabilities and ubiquity of Large Language Models (LLMs) grow, so does their environmental footprint. Despite calls for responsible AI, the machine learning community lacks standardised practices for carbon accounting. Our automated literature review of the 5,285 papers accepted to NeurIPS 2025 reveals that reporting of environmental impact is nearly non-existent. To catalyse a shift toward sustainable AI, we define standardised sustainability metrics for evaluating model training efficiency, accompanied by simple heuristics to estimate the carbon cost of LLM inference. We implement these metrics in carbonbenchmark, a drop-in software solution for tracking and reporting emissions. Finally, to combat the pursuit of marginal accuracy gains at disproportionate environmental costs, we formalise the `Smallest Model that Achieves the Job' (SMAJ), a framework which challenges the field to prioritise computational efficiency and environmental accountability alongside traditional `State-of-the-Art' (SotA) accuracy.
展开完整摘要收起摘要↓
As the capabilities and ubiquity of Large Language Models (LLMs) grow, so does their environmental footprint. Despite calls for responsible AI, the machine learning community lacks standardised practices for carbon accounting. Our automated literature review of the 5,285 papers accepted to NeurIPS 2025 reveals that reporting of environmental impact is nearly non-existent. To catalyse a shift toward sustainable AI, we define standardised sustainability metrics for evaluating model training efficiency, accompanied by simple heuristics to estimate the carbon cost of LLM inference. We implement these metrics in carbonbenchmark, a drop-in software solution for tracking and reporting emissions. Finally, to combat the pursuit of marginal accuracy gains at disproportionate environmental costs, we formalise the `Smallest Model that Achieves the Job' (SMAJ), a framework which challenges the field to prioritise computational efficiency and environmental accountability alongside traditional `State-of-the-Art' (SotA) accuracy.
Layer interventions are widely used to probe the internal organization of language models, yet most analyses examine a single training checkpoint even though model representations and computations evolve throughout pretraining. This leaves open which depth-dependent intervention responses reflect persistent organization and which are transient consequences of training. We study this question using single-block identity bypass on fixed teacher-forced contexts across five released trajectories and 11 model-domain combinations. We find that block-bypass responses retain recognizable depth ordering while their magnitudes redistribute: nearby checkpoints preserve stronger rank correspondence than distant ones, and large changes concentrate at positions that recur across text samples and transfer across evaluation domains. Controlled experiments further show that changes in the natural bypass effect cannot be reduced to a single downstream sensitivity: in replicated Pythia runs, local missing-update magnitude grows while the pooled matched downstream response decreases, whereas OLMo-2 7B exhibits a different balance. These matched responses also depend on perturbation strength and direction, without identifying targeted compensation. Together, our results show that longitudinal layer sensitivity is structured but not static, and that single-checkpoint intervention responses should be interpreted in the context of how the underlying perturbation pathway evolves during training.
展开完整摘要收起摘要↓
Layer interventions are widely used to probe the internal organization of language models, yet most analyses examine a single training checkpoint even though model representations and computations evolve throughout pretraining. This leaves open which depth-dependent intervention responses reflect persistent organization and which are transient consequences of training. We study this question using single-block identity bypass on fixed teacher-forced contexts across five released trajectories and 11 model-domain combinations. We find that block-bypass responses retain recognizable depth ordering while their magnitudes redistribute: nearby checkpoints preserve stronger rank correspondence than distant ones, and large changes concentrate at positions that recur across text samples and transfer across evaluation domains. Controlled experiments further show that changes in the natural bypass effect cannot be reduced to a single downstream sensitivity: in replicated Pythia runs, local missing-update magnitude grows while the pooled matched downstream response decreases, whereas OLMo-2 7B exhibits a different balance. These matched responses also depend on perturbation strength and direction, without identifying targeted compensation. Together, our results show that longitudinal layer sensitivity is structured but not static, and that single-checkpoint intervention responses should be interpreted in the context of how the underlying perturbation pathway evolves during training.
作者Andreea Dutulescu, Stefan Ruseti, Mihai Masala, Traian Rebedea, Mihai Dascalu
Most preference optimization methods, such as Direct Preference Optimization (DPO), apply preference supervision at the response level, although autoregressive language models are optimized token by token. As a result, all tokens in a rejected response contribute to the negative training signal, including tokens that may encode behavior that is useful for the preferred response. We introduce GAW-PO, a gradient-aligned token reweighting method for DPO that estimates, for each rejected token, whether penalizing it would interfere with the preferred update directions. Tokens whose gradients are strongly aligned with the preferred behavior receive a weaker negative contribution, while conflicting tokens retain a stronger penalty. Our method achieves the highest average performance among the evaluated preference-optimization methods, improving by 0.97 points over standard DPO and 0.65 points over the strongest competing baseline across 11 benchmarks spanning mathematics, reasoning, coding, and question answering. We further show that gradient-aligned weighting is substantially more robust to aggressive preference optimization: as the DPO regularization parameter $β$ decreases, standard DPO degrades sharply, whereas GAW-PO continues to improve. These results suggest that accounting for the interaction between rejected-token updates and preferred behavior provides an effective form of token-level credit assignment for preference optimization.
展开完整摘要收起摘要↓
Most preference optimization methods, such as Direct Preference Optimization (DPO), apply preference supervision at the response level, although autoregressive language models are optimized token by token. As a result, all tokens in a rejected response contribute to the negative training signal, including tokens that may encode behavior that is useful for the preferred response. We introduce GAW-PO, a gradient-aligned token reweighting method for DPO that estimates, for each rejected token, whether penalizing it would interfere with the preferred update directions. Tokens whose gradients are strongly aligned with the preferred behavior receive a weaker negative contribution, while conflicting tokens retain a stronger penalty. Our method achieves the highest average performance among the evaluated preference-optimization methods, improving by 0.97 points over standard DPO and 0.65 points over the strongest competing baseline across 11 benchmarks spanning mathematics, reasoning, coding, and question answering. We further show that gradient-aligned weighting is substantially more robust to aggressive preference optimization: as the DPO regularization parameter $β$ decreases, standard DPO degrades sharply, whereas GAW-PO continues to improve. These results suggest that accounting for the interaction between rejected-token updates and preferred behavior provides an effective form of token-level credit assignment for preference optimization.
Adaptive preconditioners accelerate model training, but heterogeneous client geometries can bias federated updates even when gradients are evaluated at the same model. Round-start synchronization alone cannot prevent this mismatch from reappearing during local training. We propose FedMIX-P, which mixes shared and local preconditioners at every local step, retaining local adaptation while reducing mean-squared operator mismatch by a factor of $λ^2$. For smooth nonconvex objectives with stochastic gradients and partial participation, we establish an $O(R^{-1/2})$ stationarity bound using suitable stepsizes and a horizon-dependent mixing weight, without requiring local preconditioners to converge to one another. A two-client counterexample shows that fixed positive mixing can preserve a nonstationary fixed point. The theory covers bounded linear symmetric positive-definite preconditioners. Experiments with SOAP, Sophia, and Muon variants across vision and language tasks show improvements over corresponding local optimizers, including accuracy gains of up to $19.47$ percentage points and lower validation loss for 60M--350M language models. Full nonlinear and momentum-based updates require separate analysis.
展开完整摘要收起摘要↓
Adaptive preconditioners accelerate model training, but heterogeneous client geometries can bias federated updates even when gradients are evaluated at the same model. Round-start synchronization alone cannot prevent this mismatch from reappearing during local training. We propose FedMIX-P, which mixes shared and local preconditioners at every local step, retaining local adaptation while reducing mean-squared operator mismatch by a factor of $λ^2$. For smooth nonconvex objectives with stochastic gradients and partial participation, we establish an $O(R^{-1/2})$ stationarity bound using suitable stepsizes and a horizon-dependent mixing weight, without requiring local preconditioners to converge to one another. A two-client counterexample shows that fixed positive mixing can preserve a nonstationary fixed point. The theory covers bounded linear symmetric positive-definite preconditioners. Experiments with SOAP, Sophia, and Muon variants across vision and language tasks show improvements over corresponding local optimizers, including accuracy gains of up to $19.47$ percentage points and lower validation loss for 60M--350M language models. Full nonlinear and momentum-based updates require separate analysis.
作者Laura van Weesep, Riccardo Tedoldi, Jens Sjölund, Hossein Azizpour, Susanne Winiwarter, Ola Engkvist, Jon Paul Janet, Samuel Genheden, Juan Viguera Diez
The emergence of foundation models for molecular property prediction requires a high degree of AI data readiness, including reliable metadata annotation. However, both public repositories and industrial screening databases suffer from missing, inconsistent, or conflated assay annotations. In this work, we quantify the extent of missing annotations in PubChem for the BioAssay Ontology (BAO) assay format and physical detection method fields and investigate whether open-source and proprietary large language models (LLMs) can reliably predict and audit metadata annotations directly from the assay text. In our assessment, we found that the annotation coverage across PubChem's $\sim$2 million bioassays is critically sparse, 36% lacking an assay format, 89% a BioAssay type, and >99.9% any BAO-mapped assay format or detection technology term. This motivates the need for automated test-metadata curation. Using evaluation sets derived from PubChem and ChEMBL, we assess the agreement of seven open-source and proprietary LLMs with existing silver labels. Recall is at least 0.96 for biochemical and cell-based assay formats, with a similar pattern for detection technology, although disagreements increase on under-represented classes. Manual inspection shows that many of these disagreements trace back to inconsistencies between silver sources rather than to LLM error. Moreover, in a qualitative study with a senior industrial curator, LLM-generated evidence prompted the expert to revise some of their own labels, showing LLMs can flag potentially mislabeled assays. Across the study, performance differences between proprietary and open-source models were small. Together, these results suggest LLMs can support the large-scale annotation and auditing of assay metadata, though per-class reliability estimates and targeted human review remain necessary before such labels enter downstream ML pipelines.
展开完整摘要收起摘要↓
The emergence of foundation models for molecular property prediction requires a high degree of AI data readiness, including reliable metadata annotation. However, both public repositories and industrial screening databases suffer from missing, inconsistent, or conflated assay annotations. In this work, we quantify the extent of missing annotations in PubChem for the BioAssay Ontology (BAO) assay format and physical detection method fields and investigate whether open-source and proprietary large language models (LLMs) can reliably predict and audit metadata annotations directly from the assay text. In our assessment, we found that the annotation coverage across PubChem's $\sim$2 million bioassays is critically sparse, 36% lacking an assay format, 89% a BioAssay type, and >99.9% any BAO-mapped assay format or detection technology term. This motivates the need for automated test-metadata curation. Using evaluation sets derived from PubChem and ChEMBL, we assess the agreement of seven open-source and proprietary LLMs with existing silver labels. Recall is at least 0.96 for biochemical and cell-based assay formats, with a similar pattern for detection technology, although disagreements increase on under-represented classes. Manual inspection shows that many of these disagreements trace back to inconsistencies between silver sources rather than to LLM error. Moreover, in a qualitative study with a senior industrial curator, LLM-generated evidence prompted the expert to revise some of their own labels, showing LLMs can flag potentially mislabeled assays. Across the study, performance differences between proprietary and open-source models were small. Together, these results suggest LLMs can support the large-scale annotation and auditing of assay metadata, though per-class reliability estimates and targeted human review remain necessary before such labels enter downstream ML pipelines.
Skill evolution offers a promising way to improve large language model agents without updating their parameters, but its use in formal theorem proving remains underexplored. Existing methods mainly target natural-language reasoning, improving skills by analyzing successful and failed trajectories and incrementally revising solving strategies. Although the Lean verifier provides reliable execution feedback, when all sampled trajectories fail, existing skill evolution methods lack successful trajectories from which to infer effective update directions. Furthermore, these methods also focus mainly on the root instruction file, thus underexploring the evolution of reference knowledge including mathematical concepts and proving techniques. To address these limitations, we propose a mutation-enhanced skill self-evolution framework for building skill-augmented Lean provers. The framework jointly evolves a high-level solving policy and its reference knowledge through progressive and mutation-based updates. Progressive evolution derives local improvements from successful and failed trajectories, while mutation is triggered when no complete proof can be generated, sampling mathematical concepts to produce and select new skill candidates under verifier feedback. We evaluate our method on MiniF2F, PutnamBench, the 2025 International Mathematical Olympiad (IMO 2025), and the 2026 USA Mathematical Olympiad (USAMO 2026). Under the same backbone model, trajectorysampling budget, and test-time compute, our method achieves proof success rates of 100.0%, 90.6%, 4/6, and 4/6, respectively, with GPT-5.5, outperforming the baseline methods. Further analysis shows that concept-guided mutation outperforms random-text-guided mutation by 6.9 and 8.2 percentage points on MiniF2F and PutnamBench, respectively, while solving one additional problem on both IMO 2025 and USAMO 2026.
展开完整摘要收起摘要↓
Skill evolution offers a promising way to improve large language model agents without updating their parameters, but its use in formal theorem proving remains underexplored. Existing methods mainly target natural-language reasoning, improving skills by analyzing successful and failed trajectories and incrementally revising solving strategies. Although the Lean verifier provides reliable execution feedback, when all sampled trajectories fail, existing skill evolution methods lack successful trajectories from which to infer effective update directions. Furthermore, these methods also focus mainly on the root instruction file, thus underexploring the evolution of reference knowledge including mathematical concepts and proving techniques. To address these limitations, we propose a mutation-enhanced skill self-evolution framework for building skill-augmented Lean provers. The framework jointly evolves a high-level solving policy and its reference knowledge through progressive and mutation-based updates. Progressive evolution derives local improvements from successful and failed trajectories, while mutation is triggered when no complete proof can be generated, sampling mathematical concepts to produce and select new skill candidates under verifier feedback. We evaluate our method on MiniF2F, PutnamBench, the 2025 International Mathematical Olympiad (IMO 2025), and the 2026 USA Mathematical Olympiad (USAMO 2026). Under the same backbone model, trajectorysampling budget, and test-time compute, our method achieves proof success rates of 100.0%, 90.6%, 4/6, and 4/6, respectively, with GPT-5.5, outperforming the baseline methods. Further analysis shows that concept-guided mutation outperforms random-text-guided mutation by 6.9 and 8.2 percentage points on MiniF2F and PutnamBench, respectively, while solving one additional problem on both IMO 2025 and USAMO 2026.
Should low-precision transformer inference use stochastic rounding (SR) or round-to-nearest (RN)? The answer depends on where in the network you look. We isolate this effect by holding the numerical format fixed and varying only the rounding rule at individual operation sites. To enable experiments at freely chosen precisions, we extend the PRISM vectorized rounding library to arbitrary virtual precision via a variable-precision stochastic rounding (VPSR) algorithm, proving that the rounding decision is evaluated exactly in hardware floating point. We develop two analyses providing complementary insight into this site-level trade-off. First, a probabilistic forward-error bound for linear projections shows that SR's error envelope grows as $O(\sqrt{n} u)$ in reduction length $n$, versus $O(n u)$ for RN, a gap that widens rapidly at low precision and is most pronounced in the long multilayer perceptron (MLP) down-projection. Second, a second-order decomposition of expected cross-entropy loss change at the output softmax into signed drift, drift curvature, and a Fisher-weighted variance penalty reveals why the two sites behave oppositely: MLP noise is predominantly a uniform logit shift to which softmax is invariant, so SR's variance is largely discounted; head noise is non-uniform across the vocabulary and is not. On DistilGPT-2 at $t=6$ significand bits, observations match theory: SR in the MLP raises perplexity to 1.15x the full-precision reference, versus 2.21x for RN. At the language-model head, the ordering reverses because SR introduces non-uniform variance, whereas deterministic RN carries none. In a mixed-precision configuration (MLP output at $t=6$), assigning SR to the MLP and RN to the head brings perplexity within 1.10x of the full-precision reference, a 28% reduction over matched-bit RN.
展开完整摘要收起摘要↓
Should low-precision transformer inference use stochastic rounding (SR) or round-to-nearest (RN)? The answer depends on where in the network you look. We isolate this effect by holding the numerical format fixed and varying only the rounding rule at individual operation sites. To enable experiments at freely chosen precisions, we extend the PRISM vectorized rounding library to arbitrary virtual precision via a variable-precision stochastic rounding (VPSR) algorithm, proving that the rounding decision is evaluated exactly in hardware floating point. We develop two analyses providing complementary insight into this site-level trade-off. First, a probabilistic forward-error bound for linear projections shows that SR's error envelope grows as $O(\sqrt{n} u)$ in reduction length $n$, versus $O(n u)$ for RN, a gap that widens rapidly at low precision and is most pronounced in the long multilayer perceptron (MLP) down-projection. Second, a second-order decomposition of expected cross-entropy loss change at the output softmax into signed drift, drift curvature, and a Fisher-weighted variance penalty reveals why the two sites behave oppositely: MLP noise is predominantly a uniform logit shift to which softmax is invariant, so SR's variance is largely discounted; head noise is non-uniform across the vocabulary and is not. On DistilGPT-2 at $t=6$ significand bits, observations match theory: SR in the MLP raises perplexity to 1.15x the full-precision reference, versus 2.21x for RN. At the language-model head, the ordering reverses because SR introduces non-uniform variance, whereas deterministic RN carries none. In a mixed-precision configuration (MLP output at $t=6$), assigning SR to the MLP and RN to the head brings perplexity within 1.10x of the full-precision reference, a 28% reduction over matched-bit RN.
作者Shawn Bowers, Martin Caminada, Haoyang Liu, Bertram Ludäscher
ABDA-NL adds a natural-language interface to ABDA, a system for argument-based discussion using ASPIC- knowledge bases under grounded semantics. Users see which conclusions are accepted, rejected, or undecided, open an interactive rendering of the grounded discussion game to learn why, explore what-if alternatives by suspending assumptions and rules or changing preferences, ask questions that are answered from a scenario's reference documents, and author new facts, assumptions, and rules in plain English. A large language model provides the bridge between language and formalism: it answers questions from the documents and the current state of the scenario, and it translates plain-English edits into candidate formal statements. The deterministic ABDA engine remains the sole source of arguments, attacks, and acceptance labels, and every proposal of the model is validated and confirmed by the user before it takes effect.
展开完整摘要收起摘要↓
ABDA-NL adds a natural-language interface to ABDA, a system for argument-based discussion using ASPIC- knowledge bases under grounded semantics. Users see which conclusions are accepted, rejected, or undecided, open an interactive rendering of the grounded discussion game to learn why, explore what-if alternatives by suspending assumptions and rules or changing preferences, ask questions that are answered from a scenario's reference documents, and author new facts, assumptions, and rules in plain English. A large language model provides the bridge between language and formalism: it answers questions from the documents and the current state of the scenario, and it translates plain-English edits into candidate formal statements. The deterministic ABDA engine remains the sole source of arguments, attacks, and acceptance labels, and every proposal of the model is validated and confirmed by the user before it takes effect.
作者Peng Cui, Qiaoyuan Zheng, Rudolf Debelak, Mrinmaya Sachan
Difficulty is one of the most fundamental properties of a question: it determines whether the question can meaningfully discriminate between models of differing ability. Although a variety of methods can now estimate or predict difficulty automatically, they yield only a single descriptive number, with no account of the underlying factors that make a question difficult in the first place. In this work, we propose a data-driven approach that automatically generates and validates natural-language hypotheses explaining what makes one question harder than another. We first estimate each item's difficulty from the responses of a large pool of LLMs using Item Response Theory. We then sample contrasting sets of easy and hard questions and prompt an LLM to propose candidate explanations of the difference, which are subsequently validated and selected on held-out questions. Experimental results across three datasets spanning mathematical, logical, and commonsense reasoning show that our method produces interpretable and predictive hypotheses. On their own, they predict the difficulty of unseen questions competitively with, or better than, advanced black-box difficulty regressors; used as additional features, they further improve those regressors, implying that they discover difficulty signals that existing models fail to capture. Moreover, we demonstrate that editing questions according to a hypothesis can shift their measured difficulty in the expected direction, indicating that the discovered hypotheses are causally valid difficulty factors rather than post-hoc descriptions. Our approach thus turns a purely descriptive difficulty score into actionable statements.
展开完整摘要收起摘要↓
Difficulty is one of the most fundamental properties of a question: it determines whether the question can meaningfully discriminate between models of differing ability. Although a variety of methods can now estimate or predict difficulty automatically, they yield only a single descriptive number, with no account of the underlying factors that make a question difficult in the first place. In this work, we propose a data-driven approach that automatically generates and validates natural-language hypotheses explaining what makes one question harder than another. We first estimate each item's difficulty from the responses of a large pool of LLMs using Item Response Theory. We then sample contrasting sets of easy and hard questions and prompt an LLM to propose candidate explanations of the difference, which are subsequently validated and selected on held-out questions. Experimental results across three datasets spanning mathematical, logical, and commonsense reasoning show that our method produces interpretable and predictive hypotheses. On their own, they predict the difficulty of unseen questions competitively with, or better than, advanced black-box difficulty regressors; used as additional features, they further improve those regressors, implying that they discover difficulty signals that existing models fail to capture. Moreover, we demonstrate that editing questions according to a hypothesis can shift their measured difficulty in the expected direction, indicating that the discovered hypotheses are causally valid difficulty factors rather than post-hoc descriptions. Our approach thus turns a purely descriptive difficulty score into actionable statements.
Many real-world systems exhibit higher-order interactions among groups of entities that cannot be captured by pairwise relationships alone. Graph Transformers and Graph Neural Networks increasingly rely on positional encodings to enrich graph representations, yet existing positional encodings are computed solely from the original graph and therefore cannot directly capture observed higher-order interactions. Topological Deep Learning addresses this limitation by lifting graphs to simplicial complexes, but typically requires performing message passing or attention on higher-order neural network representations. We introduce a representation learning paradigm that enriches graph representations with higher-order topology through positional encodings, enabling standard graph learning models to exploit lifted incidence structure without modifying the backbone. We derive a theoretical characterization of the expressivity of higher-order positional encodings, proving that node-level operators induced by higher-order lifts can mix graph Laplacian frequencies in ways that scalar graph spectral filters cannot. Guided by this theory, we instantiate higher-order positional encodings using Hodge Laplacians derived from clique complexes. Experiments with Graph Transformers on ZINC and controlled synthetic benchmarks demonstrate improvements in predictive performance, while a fixed-1-skeleton experiment shows that the pipeline can transmit higher-order information when cells are supplied independently of the graph. Together, our results establish higher-order positional encodings as a principled bridge between graph positional encodings and topological deep learning.
展开完整摘要收起摘要↓
Many real-world systems exhibit higher-order interactions among groups of entities that cannot be captured by pairwise relationships alone. Graph Transformers and Graph Neural Networks increasingly rely on positional encodings to enrich graph representations, yet existing positional encodings are computed solely from the original graph and therefore cannot directly capture observed higher-order interactions. Topological Deep Learning addresses this limitation by lifting graphs to simplicial complexes, but typically requires performing message passing or attention on higher-order neural network representations. We introduce a representation learning paradigm that enriches graph representations with higher-order topology through positional encodings, enabling standard graph learning models to exploit lifted incidence structure without modifying the backbone. We derive a theoretical characterization of the expressivity of higher-order positional encodings, proving that node-level operators induced by higher-order lifts can mix graph Laplacian frequencies in ways that scalar graph spectral filters cannot. Guided by this theory, we instantiate higher-order positional encodings using Hodge Laplacians derived from clique complexes. Experiments with Graph Transformers on ZINC and controlled synthetic benchmarks demonstrate improvements in predictive performance, while a fixed-1-skeleton experiment shows that the pipeline can transmit higher-order information when cells are supplied independently of the graph. Together, our results establish higher-order positional encodings as a principled bridge between graph positional encodings and topological deep learning.
Memory-augmented large language models must decide which memories to retain, and recent systems do so by estimating each memory's effect on task performance. However, these estimates rely entirely on retrieved memories. When a memory is never retrieved, store-level interventions produce identical outcomes, leaving its utility unidentified. This is a retrieval-level positivity violation, invisible to diagnostics that examine only memory operations. We introduce Causal Memory Policy (CMP), a causal framework that restores identification by intervening on retrieval itself, reserving a fixed number of context slots for memories sampled with known propensities. CMP estimates memory utility by self-normalized inverse propensity weighting under a balanced assignment design. We prove the causal factorization of memory utility through retrieval, the unbiasedness and exact variance of the estimator, and the optimal decision rule under irreversible operations. Empirically, identification fails for 54% of required memories on LongMemEval and 67% on LoCoMo, and the failure persists in a deployed memory system. CMP improves discrimination between required and non-required memories from 0.54 to 0.66 AUC. Finally, we show that identified memory utility alone is insufficient for retention decisions: per-query utility reaches 0.78 AUC on the query for which it is estimated, yet no aggregation available to a retention policy predicts a memory's value on unseen queries. Code is available at: https://anonymous.4open.science/r/cmp-release-D0C3/.
展开完整摘要收起摘要↓
Memory-augmented large language models must decide which memories to retain, and recent systems do so by estimating each memory's effect on task performance. However, these estimates rely entirely on retrieved memories. When a memory is never retrieved, store-level interventions produce identical outcomes, leaving its utility unidentified. This is a retrieval-level positivity violation, invisible to diagnostics that examine only memory operations. We introduce Causal Memory Policy (CMP), a causal framework that restores identification by intervening on retrieval itself, reserving a fixed number of context slots for memories sampled with known propensities. CMP estimates memory utility by self-normalized inverse propensity weighting under a balanced assignment design. We prove the causal factorization of memory utility through retrieval, the unbiasedness and exact variance of the estimator, and the optimal decision rule under irreversible operations. Empirically, identification fails for 54% of required memories on LongMemEval and 67% on LoCoMo, and the failure persists in a deployed memory system. CMP improves discrimination between required and non-required memories from 0.54 to 0.66 AUC. Finally, we show that identified memory utility alone is insufficient for retention decisions: per-query utility reaches 0.78 AUC on the query for which it is estimated, yet no aggregation available to a retention policy predicts a memory's value on unseen queries. Code is available at: https://anonymous.4open.science/r/cmp-release-D0C3/.
Activation sparsity speeds up large language model (LLM) inference by setting unimportant activations to zero so that the corresponding computations can be skipped. Existing training-free methods, however, make different trade-offs: threshold-based methods such as TEAL adapt the sparsity level to each token but do not tightly control the realised sparsity, while TopK-based methods such as WINA enforce a fixed sparsity level but use the same sparsity budget for every token. Both also apply the same budget across transformer blocks, despite large differences in block sensitivity. We introduce TopK-Guided, a training-free method that addresses both limitations by combining bounded token-level sparsity adaptation with sensitivity-aware block-level budget allocation. Across Llama-2 and Llama-3 models, TopK-Guided consistently improves perplexity and downstream accuracy over TEAL and WINA while preserving essentially the same sparsitydependent projection compute as WINA, with the largest gains at high sparsity. Ablations show that both components provide complementary improvements.
展开完整摘要收起摘要↓
Activation sparsity speeds up large language model (LLM) inference by setting unimportant activations to zero so that the corresponding computations can be skipped. Existing training-free methods, however, make different trade-offs: threshold-based methods such as TEAL adapt the sparsity level to each token but do not tightly control the realised sparsity, while TopK-based methods such as WINA enforce a fixed sparsity level but use the same sparsity budget for every token. Both also apply the same budget across transformer blocks, despite large differences in block sensitivity. We introduce TopK-Guided, a training-free method that addresses both limitations by combining bounded token-level sparsity adaptation with sensitivity-aware block-level budget allocation. Across Llama-2 and Llama-3 models, TopK-Guided consistently improves perplexity and downstream accuracy over TEAL and WINA while preserving essentially the same sparsitydependent projection compute as WINA, with the largest gains at high sparsity. Ablations show that both components provide complementary improvements.
作者Azal Ahmad Khan, Keshav Ramji, Tahira Naseem, Ali Anwar, Ramón Fernandez Astudillo
Hypernetworks have recently shown success in dynamically adapting the parameters of Large Language Models (LLMs) at runtime based on signals such as task descriptions or additional demostrations. Here we ask: how much adaptation signal can be obtained using only the input query to an LLM?. To answer this, we study query-conditioned Hypernetworks for LoRA estimation. Further, we introduce distributional Hypernetworks, able to produce not only point estimates of parameter adaptors, but also a distribution over possible LoRAs. For this we propose a simple end-to-end loss using a differentiable Monte Carlo approximation and explore multiple distribution parametrizations including regression and convex combination variants. Results show that even using the mean of the learned distribution can outperform deterministic hypernetworks. Crucially, the learned distribution enables a different form of test-time scaling: instead of spending additional compute only by sampling more token sequences from a fixed model, we sample weight updates, yielding multiple adapted models for the same query. Performance improves as more weight samples are considered and remains stronger than corresponding token-sampling adaptation baselines. Finally, we find that generated updates can transfer across queries, suggesting that the hypernetwork learns reusable structure in how the model should adapt. Together, these results show that query-conditioned distributions over weight updates can support both adaptation and test-time scaling.
展开完整摘要收起摘要↓
Hypernetworks have recently shown success in dynamically adapting the parameters of Large Language Models (LLMs) at runtime based on signals such as task descriptions or additional demostrations. Here we ask: how much adaptation signal can be obtained using only the input query to an LLM?. To answer this, we study query-conditioned Hypernetworks for LoRA estimation. Further, we introduce distributional Hypernetworks, able to produce not only point estimates of parameter adaptors, but also a distribution over possible LoRAs. For this we propose a simple end-to-end loss using a differentiable Monte Carlo approximation and explore multiple distribution parametrizations including regression and convex combination variants. Results show that even using the mean of the learned distribution can outperform deterministic hypernetworks. Crucially, the learned distribution enables a different form of test-time scaling: instead of spending additional compute only by sampling more token sequences from a fixed model, we sample weight updates, yielding multiple adapted models for the same query. Performance improves as more weight samples are considered and remains stronger than corresponding token-sampling adaptation baselines. Finally, we find that generated updates can transfer across queries, suggesting that the hypernetwork learns reusable structure in how the model should adapt. Together, these results show that query-conditioned distributions over weight updates can support both adaptation and test-time scaling.
Step-size selection remains a central challenge in large-scale neural network optimization; conservative steps slow convergence, while aggressive steps can destabilize it. We combine Zero-and-First-Order optimization~(ZFO) and propose a lightweight framework that decouples direction selection from step-size. ZFO uses a trusted first-order optimizer to determine the direction and performs zeroth-order evaluations only along this one-dimensional subspace to choose how far to move. Using the current {gradient information} and two additional objective function evaluations, ZFO instances construct a local model of the objective function along the proposed direction and select a curvature-aware step within a bounded search interval. This yields an adaptive step-selection mechanism that costs less than a full line search. We provide theoretical guarantees to show that shared-sample evaluations produce reliable finite-difference curvature estimates, that the induced local model selects a near-optimal step along the search interval, and that ZFO converges to a neighborhood of a stationary point. Across the evaluated settings, language models and datasets, ZFO frequently improves optimization and final performance relative to fixed-step first-order baselines, with the magnitude and preferred local model depending on the objective. Our code is publicly available at: https://github.com/nizswan/Zeroth-First-Order-Framework.
展开完整摘要收起摘要↓
Step-size selection remains a central challenge in large-scale neural network optimization; conservative steps slow convergence, while aggressive steps can destabilize it. We combine Zero-and-First-Order optimization~(ZFO) and propose a lightweight framework that decouples direction selection from step-size. ZFO uses a trusted first-order optimizer to determine the direction and performs zeroth-order evaluations only along this one-dimensional subspace to choose how far to move. Using the current {gradient information} and two additional objective function evaluations, ZFO instances construct a local model of the objective function along the proposed direction and select a curvature-aware step within a bounded search interval. This yields an adaptive step-selection mechanism that costs less than a full line search. We provide theoretical guarantees to show that shared-sample evaluations produce reliable finite-difference curvature estimates, that the induced local model selects a near-optimal step along the search interval, and that ZFO converges to a neighborhood of a stationary point. Across the evaluated settings, language models and datasets, ZFO frequently improves optimization and final performance relative to fixed-step first-order baselines, with the magnitude and preferred local model depending on the objective. Our code is publicly available at: https://github.com/nizswan/Zeroth-First-Order-Framework.
Alignment does not eliminate behavioral errors in language models. Models may still refuse benign requests, call unnecessary tools, or yield to false user claims. Current methods mitigate such errors as a computation problem, and rarely explore if the desired behavior is already encoded in the model's representation. Motivated by the observation that behavior-relevant information remains linearly decodable from the final hidden state even when the resulting logits produce the undesired behavior, we introduce HeadEdit, a gradient-free method that calibrates model behavior through the unembedding matrix. HeadEdit extracts a low-rank behavioral subspace from paired completions and uses each prompt's coordinates within it to generate a vocabulary-wide correction, thereby implementing implicitly adaptive steering without manually specified target tokens or parameter updates. HeadEdit improves all nine experimental settings across three tasks and three model families, with negligible inference overhead and no systematic loss of general capabilities. It also reveals a connection to gradient-based alignment. HeadEdit's low-dimensional representation partly predicts how preference tuning changes output logits on unseen prompts. The subspace learned from the model can also be reused after tuning, improving performance without re-extracting or retuning. These results show that HeadEdit provides a practical, lightweight, and interpretable way to calibrate model behavior through the unembedding matrix.
展开完整摘要收起摘要↓
Alignment does not eliminate behavioral errors in language models. Models may still refuse benign requests, call unnecessary tools, or yield to false user claims. Current methods mitigate such errors as a computation problem, and rarely explore if the desired behavior is already encoded in the model's representation. Motivated by the observation that behavior-relevant information remains linearly decodable from the final hidden state even when the resulting logits produce the undesired behavior, we introduce HeadEdit, a gradient-free method that calibrates model behavior through the unembedding matrix. HeadEdit extracts a low-rank behavioral subspace from paired completions and uses each prompt's coordinates within it to generate a vocabulary-wide correction, thereby implementing implicitly adaptive steering without manually specified target tokens or parameter updates. HeadEdit improves all nine experimental settings across three tasks and three model families, with negligible inference overhead and no systematic loss of general capabilities. It also reveals a connection to gradient-based alignment. HeadEdit's low-dimensional representation partly predicts how preference tuning changes output logits on unseen prompts. The subspace learned from the model can also be reused after tuning, improving performance without re-extracting or retuning. These results show that HeadEdit provides a practical, lightweight, and interpretable way to calibrate model behavior through the unembedding matrix.
With the increasing capabilities of Large-Language-Models (LLMs) and LLM-based agents, users are increasingly using them to solve everyday problems, such as answering e-mails or providing programming support. Existing work has extensively investigated security and privacy risks, such as prompt injections and the disclosure of sensitive data to chatbot providers. While various solutions were developed to address these risks, including input structuring to prevent prompt injections or deploying local LLMs to avoid sharing confidential data with chatbot operators, LLMs also pose the risk of leaking confidential data to third parties. In this paper, we demonstrate with LLMLeak a novel attack vector where malicious software that runs locally but cannot communicate directly with the internet abuses LLMs to establish a covert channel. While inputs that instruct the LLM to send data directly via generated code are easy to detect and network libraries are typically restricted, LLMLeak relies only on the LLM's tool to fetch websites for further information. A malicious software component on the client side embeds a secret into a URL. It presents the referenced website as providing information required for a benign task, such as migrating a software library. When the LLM accesses the URL, the attacker receives the encoded secret through an attacker-controlled DNS or web server. We perform an extensive evaluation on eleven open-parameter models, observe an attack success rate of 79.7%, and also conduct a case study on real-world chatbots, demonstrating the relevance of LLMLeak.
展开完整摘要收起摘要↓
With the increasing capabilities of Large-Language-Models (LLMs) and LLM-based agents, users are increasingly using them to solve everyday problems, such as answering e-mails or providing programming support. Existing work has extensively investigated security and privacy risks, such as prompt injections and the disclosure of sensitive data to chatbot providers. While various solutions were developed to address these risks, including input structuring to prevent prompt injections or deploying local LLMs to avoid sharing confidential data with chatbot operators, LLMs also pose the risk of leaking confidential data to third parties. In this paper, we demonstrate with LLMLeak a novel attack vector where malicious software that runs locally but cannot communicate directly with the internet abuses LLMs to establish a covert channel. While inputs that instruct the LLM to send data directly via generated code are easy to detect and network libraries are typically restricted, LLMLeak relies only on the LLM's tool to fetch websites for further information. A malicious software component on the client side embeds a secret into a URL. It presents the referenced website as providing information required for a benign task, such as migrating a software library. When the LLM accesses the URL, the attacker receives the encoded secret through an attacker-controlled DNS or web server. We perform an extensive evaluation on eleven open-parameter models, observe an attack success rate of 79.7%, and also conduct a case study on real-world chatbots, demonstrating the relevance of LLMLeak.
Some properties of languages, e.g., subject-object-verb (SOV) word order, are more prevalent than others among the thousands of attested natural languages (NLs). Such typological commonality is often attributed to learning biases. Computational simulations, recently with language models (LMs), have facilitated the exploration of this theory. In this paper, we extend existing analyses of the relationship between LMs' learning biases and typological commonality on both data and model sides, focusing on: (i) cross-serial dependencies, the upper limit of attested syntactic complexity, and (ii) stack-based LMs (SLMs), potentially facilitating learning of hierarchical patterns. We first evaluate generalization of SLMs on cross-serial dependencies across diverse artificial languages and confirm that they struggle with such constructions. However, SLMs with limited working memory generalize better suggesting a possible basis for such inductive bias and thus the typological commonality of some word order configurations.
展开完整摘要收起摘要↓
Some properties of languages, e.g., subject-object-verb (SOV) word order, are more prevalent than others among the thousands of attested natural languages (NLs). Such typological commonality is often attributed to learning biases. Computational simulations, recently with language models (LMs), have facilitated the exploration of this theory. In this paper, we extend existing analyses of the relationship between LMs' learning biases and typological commonality on both data and model sides, focusing on: (i) cross-serial dependencies, the upper limit of attested syntactic complexity, and (ii) stack-based LMs (SLMs), potentially facilitating learning of hierarchical patterns. We first evaluate generalization of SLMs on cross-serial dependencies across diverse artificial languages and confirm that they struggle with such constructions. However, SLMs with limited working memory generalize better suggesting a possible basis for such inductive bias and thus the typological commonality of some word order configurations.
Data selection is critical for training large language models on massive and heterogeneous corpora. Meta-learning for Training-data Selection offers a principled alternative to heuristic scoring by learning data weights from a target validation objective, but existing methods face a trade-off between fine-grained valuation and transferability to unseen data. A natural solution is to replace per-sample weights with a selection network. However, we find that directly incorporating such a network into existing MTS objectives leads to unstable optimization and poor generalization, caused by weight suppression and persistent reliance on easy-to-learn features. To address these issues, we propose Transferable Example Scoring and Selection (TESS), a scalable data-selection framework built on a Pointwise Value Matching objective (PVM). Experiments on LLM safety and targeted instruction tuning demonstrate strong transfer across datasets, from subsets to full corpora, and from smaller to larger models.
展开完整摘要收起摘要↓
Data selection is critical for training large language models on massive and heterogeneous corpora. Meta-learning for Training-data Selection offers a principled alternative to heuristic scoring by learning data weights from a target validation objective, but existing methods face a trade-off between fine-grained valuation and transferability to unseen data. A natural solution is to replace per-sample weights with a selection network. However, we find that directly incorporating such a network into existing MTS objectives leads to unstable optimization and poor generalization, caused by weight suppression and persistent reliance on easy-to-learn features. To address these issues, we propose Transferable Example Scoring and Selection (TESS), a scalable data-selection framework built on a Pointwise Value Matching objective (PVM). Experiments on LLM safety and targeted instruction tuning demonstrate strong transfer across datasets, from subsets to full corpora, and from smaller to larger models.
作者Chuqin Geng, Li Zhang, Haolin Ye, Mark Zhang, Luke Zhang, Xujie Si
Mechanistic interpretability aims to recover the internal computations responsible for model behavior. Progress in automated circuit discovery is often framed as a search problem: better attribution or optimization should identify better mechanisms. This assumes that the evaluation objective can recognize a better circuit once it is found. Yet standard intervention-based evaluation replaces signals outside a circuit to isolate its effect, changing the computational environment in which the retained components operate. We call this change context distortion. This raises the question of whether faithfulness rewards recovery of model behavior or compatibility with the altered context. We show that intervention-based faithfulness can prefer an equally sized circuit that reproduces the model's behavior less well. We call this reversal misranking, an objective-level recovery gap. To isolate evaluation from discovery, we construct equally sized candidates through controlled edits to human-identified circuits in pretrained language models, including IOI and Greater-Than, and benchmark circuits with known mechanisms. We find misranking even without a discovery algorithm. Misranking also occurs in real automated circuit-discovery settings. Among outputs of EAP, EAP-IG, ACDC, and Edge-SP, KL faithfulness under ordinary resampling prefers the behaviorally worse candidate in 13.2%-24.4% of candidate pairs across these methods on four pretrained-language-model tasks. Finally, restoring selected intact-model signals reverses 96 of 100 selected persistent misrankings, demonstrating that context distortion contributes to these failures. In light of these findings, we call for rethinking both how circuit discovery is performed and how its results are evaluated.
展开完整摘要收起摘要↓
Mechanistic interpretability aims to recover the internal computations responsible for model behavior. Progress in automated circuit discovery is often framed as a search problem: better attribution or optimization should identify better mechanisms. This assumes that the evaluation objective can recognize a better circuit once it is found. Yet standard intervention-based evaluation replaces signals outside a circuit to isolate its effect, changing the computational environment in which the retained components operate. We call this change context distortion. This raises the question of whether faithfulness rewards recovery of model behavior or compatibility with the altered context. We show that intervention-based faithfulness can prefer an equally sized circuit that reproduces the model's behavior less well. We call this reversal misranking, an objective-level recovery gap. To isolate evaluation from discovery, we construct equally sized candidates through controlled edits to human-identified circuits in pretrained language models, including IOI and Greater-Than, and benchmark circuits with known mechanisms. We find misranking even without a discovery algorithm. Misranking also occurs in real automated circuit-discovery settings. Among outputs of EAP, EAP-IG, ACDC, and Edge-SP, KL faithfulness under ordinary resampling prefers the behaviorally worse candidate in 13.2%-24.4% of candidate pairs across these methods on four pretrained-language-model tasks. Finally, restoring selected intact-model signals reverses 96 of 100 selected persistent misrankings, demonstrating that context distortion contributes to these failures. In light of these findings, we call for rethinking both how circuit discovery is performed and how its results are evaluated.
作者Wanyang Hong, Zhaoning Zhang, Yi Chen, Libo Zhang, Baihui Liu, Linbo Qiao, Zhiliang Tian, Dongsheng Li
Large language models often lose track of persistent instructions and relevant information as multi-turn conversations grow. We study this cumulative contextual decay through three related failure modes: attention pollution, dilution, and drift. We propose REA (Role-aware Heuristic Episodic Attention), a context-management framework that assigns different persistence and representation policies to instructions and episodic interactions. Instructional Memory retains identified global constraints in a dedicated prefix. Episodic Memory preserves user inputs and compresses model replies, while heuristic retrieval selects raw text, compressed representations, or omission for each historical turn. On Long-MT-Bench+, REA improves the judge score from 6.32 to 7.36 on a 10-point scale, a 16.5% relative gain over the Vanilla baseline, and reduces average latency by 2.91$\times$. Additional evaluations show aggregate gains on three backbones spanning 1.7B-7B parameters and on Chinese and English role-playing tasks. These results support role-aware context management as a practical approach to maintaining conversational continuity and instruction adherence.
展开完整摘要收起摘要↓
Large language models often lose track of persistent instructions and relevant information as multi-turn conversations grow. We study this cumulative contextual decay through three related failure modes: attention pollution, dilution, and drift. We propose REA (Role-aware Heuristic Episodic Attention), a context-management framework that assigns different persistence and representation policies to instructions and episodic interactions. Instructional Memory retains identified global constraints in a dedicated prefix. Episodic Memory preserves user inputs and compresses model replies, while heuristic retrieval selects raw text, compressed representations, or omission for each historical turn. On Long-MT-Bench+, REA improves the judge score from 6.32 to 7.36 on a 10-point scale, a 16.5% relative gain over the Vanilla baseline, and reduces average latency by 2.91$\times$. Additional evaluations show aggregate gains on three backbones spanning 1.7B-7B parameters and on Chinese and English role-playing tasks. These results support role-aware context management as a practical approach to maintaining conversational continuity and instruction adherence.
作者Jimin Seo, Gyubok Lee, Yeonsik Jo, Kiwoong Yoo, Yeongoon Kim, Minhae Oh, Jin Woo Koo, Suhwan Kim, Nakyung Lee, Minsik Seol, Idris Nechnech, Jaehyeon Kim, Giho Lee, Jungwoo Lee
We study learning rate (LR) scaling for hybrid architectures combining Transformer and State-Space Model (SSM) blocks, a class adopted by several recent production language models. In particular, we focus on the gap between the theoretical scaling rules derived for SSMs under zero-order-hold (ZOH) discretization at infinite width with growing state size, and the field-standard practical implementations using simplified-ZOH Mamba at fixed state size. Surprisingly, in this practical regime hybrid architectures achieve a near-zero LR transfer gap across widths 256-2048 and depths 4-32 up to billion-parameter scale using only the original $μ$P prescription, even though SSM operations fall outside its Tensor Programs representability conditions and every parameterization we test fails the standard coordinate-check diagnostic of $μ$P correctness. We attribute this to a two-condition decomposition of LR transfer in hybrid architectures: a global update-to-weight invariance, enforced by $μ$P's initialization and LR scaling; and a local per-component balance, provided by AdamW's per-parameter normalization. Our observations show that the optimal LR is invariant to width up to 8$\times$, that this width invariance holds across depth, sequence length, batch size, and Transformer-to-SSM ratio, and that it transfers to Nemotron-H, a production hybrid outside our custom architecture set. We hope these findings fill the gap between theoretical scaling rules and practical hybrid implementations, and stimulate further research toward bridging it.
展开完整摘要收起摘要↓
We study learning rate (LR) scaling for hybrid architectures combining Transformer and State-Space Model (SSM) blocks, a class adopted by several recent production language models. In particular, we focus on the gap between the theoretical scaling rules derived for SSMs under zero-order-hold (ZOH) discretization at infinite width with growing state size, and the field-standard practical implementations using simplified-ZOH Mamba at fixed state size. Surprisingly, in this practical regime hybrid architectures achieve a near-zero LR transfer gap across widths 256-2048 and depths 4-32 up to billion-parameter scale using only the original $μ$P prescription, even though SSM operations fall outside its Tensor Programs representability conditions and every parameterization we test fails the standard coordinate-check diagnostic of $μ$P correctness. We attribute this to a two-condition decomposition of LR transfer in hybrid architectures: a global update-to-weight invariance, enforced by $μ$P's initialization and LR scaling; and a local per-component balance, provided by AdamW's per-parameter normalization. Our observations show that the optimal LR is invariant to width up to 8$\times$, that this width invariance holds across depth, sequence length, batch size, and Transformer-to-SSM ratio, and that it transfers to Nemotron-H, a production hybrid outside our custom architecture set. We hope these findings fill the gap between theoretical scaling rules and practical hybrid implementations, and stimulate further research toward bridging it.