Large Language Models (LLMs) are increasingly used in software engineering (SE), yet there is no systematic study that determines to which degree these LLMs actually understand standardized SE terminology. Lack of such understanding can lead to miscommunication and misunderstanding, both by LLMs consuming text but also by human-developers acting on LLM-generated text. Within this paper, we investigate to which degree state-of-the-art LLMs are able to identify whether definitions from the ISO/IEC/IEEE 24765:2017 Systems and Software Engineering - Vocabulary are correct. We prompt LLMs both with correct definitions, as well as systematically falsified definitions. The falsifications are both semantic (substitution of key terms) and structural (removing critical information). We measure both classification accuracy and whether reasoning tokens generated by the LLMs make sense with respect to understanding the definition. While most LLMs detect falsified definitions with high accuracy, they also reject many correct definitions, indicating a systematic rejection bias rather than genuine discriminative understanding. Explicit reasoning does not consistently improve results and may even hinder performance through over-thinking. Our work demonstrates that while the performance of LLMs (including their agentic use) in many SE tasks is impressive, there are still fundamental issues to understand how this will impact SE, including the consistent use of terminology.
展开完整摘要收起摘要↓
Large Language Models (LLMs) are increasingly used in software engineering (SE), yet there is no systematic study that determines to which degree these LLMs actually understand standardized SE terminology. Lack of such understanding can lead to miscommunication and misunderstanding, both by LLMs consuming text but also by human-developers acting on LLM-generated text. Within this paper, we investigate to which degree state-of-the-art LLMs are able to identify whether definitions from the ISO/IEC/IEEE 24765:2017 Systems and Software Engineering - Vocabulary are correct. We prompt LLMs both with correct definitions, as well as systematically falsified definitions. The falsifications are both semantic (substitution of key terms) and structural (removing critical information). We measure both classification accuracy and whether reasoning tokens generated by the LLMs make sense with respect to understanding the definition. While most LLMs detect falsified definitions with high accuracy, they also reject many correct definitions, indicating a systematic rejection bias rather than genuine discriminative understanding. Explicit reasoning does not consistently improve results and may even hinder performance through over-thinking. Our work demonstrates that while the performance of LLMs (including their agentic use) in many SE tasks is impressive, there are still fundamental issues to understand how this will impact SE, including the consistent use of terminology.
作者Zeyu He, Xuan Qi, Subramanian Chidambaram, Zhichao Xu, Vinayak Arannil, Lydia Chilton, Alex C. Williams
Large language models increasingly serve as judges in AI evaluation, but current approaches rely on consensus preferences that ignore individual evaluator variation. We propose a novel simulation approach that combines categorical judgments with evaluator-specific auxiliary data--retrospective reasoning traces and interface telemetry--to enable LLM-based simulation of individual evaluators via in-context learning. We conduct a systematic empirical study of this approach using multi-facet data from 32 trained annotators across 4,200 preference judgments in a 4 x 4 x 4 factorial design. Our key findings: (1) The simulation approach achieves up to 9.9 percentage point improvements over the Base Judge; (2) Reasoning traces provide the largest gains with higher collection efforts, while interface telemetry often hurts rather than helps performance despite being cheaper to collect. (3) Simulation difficulty is systematic, predicted by an evaluator's neutral usage (most clearly on Helpfulness) and divergence from consensus; the neutral-usage tendency--rather than simulatability itself--is the cross-task-stable property (r = 0.728). These results establish both the potential and limits of evaluator-specific auxiliary data for personalized evaluation, offering methodological insights for scaling individual aware AI assessment.
展开完整摘要收起摘要↓
Large language models increasingly serve as judges in AI evaluation, but current approaches rely on consensus preferences that ignore individual evaluator variation. We propose a novel simulation approach that combines categorical judgments with evaluator-specific auxiliary data--retrospective reasoning traces and interface telemetry--to enable LLM-based simulation of individual evaluators via in-context learning. We conduct a systematic empirical study of this approach using multi-facet data from 32 trained annotators across 4,200 preference judgments in a 4 x 4 x 4 factorial design. Our key findings: (1) The simulation approach achieves up to 9.9 percentage point improvements over the Base Judge; (2) Reasoning traces provide the largest gains with higher collection efforts, while interface telemetry often hurts rather than helps performance despite being cheaper to collect. (3) Simulation difficulty is systematic, predicted by an evaluator's neutral usage (most clearly on Helpfulness) and divergence from consensus; the neutral-usage tendency--rather than simulatability itself--is the cross-task-stable property (r = 0.728). These results establish both the potential and limits of evaluator-specific auxiliary data for personalized evaluation, offering methodological insights for scaling individual aware AI assessment.
Language models (LMs) exhibit problematic biases, such as stereotypes. Effectively analyzing and mitigating such biases requires accurate and generalizable evaluation methods of the underlying associations. Some existing approaches focus on downstream metrics that analyze associations in generated text. Since generated text content can vary drastically across LMs, such metrics often require specialized evaluation datasets, which limits the generalization of such downstream metrics. In contrast, upstream metrics examine LMs at the fundamental level of embeddings or continuation probabilities, enabling principled association analyses across LMs. Yet, to date, no upstream metric for generative LMs has uncovered a strong relationship with real-world associations, including those measured in generated text. To address this gap, we introduce the Relative Probability Association Metric (RPAM), an association evaluation metric for generative LMs. For three LMs of different quality of language generation and purpose (Mistral-7B-Instruct, Mistral-7B, and GPT-2) and well-studied evaluation datasets (WEAT-WS, Bellezza, WS-353, and SST2), we find a strong relationship between upstream RPAM measurements and corresponding implicit and explicit associations observed in humans, as well as biases measured downstream with LM-specific tasks, outperforming prior record values where applicable.
展开完整摘要收起摘要↓
Language models (LMs) exhibit problematic biases, such as stereotypes. Effectively analyzing and mitigating such biases requires accurate and generalizable evaluation methods of the underlying associations. Some existing approaches focus on downstream metrics that analyze associations in generated text. Since generated text content can vary drastically across LMs, such metrics often require specialized evaluation datasets, which limits the generalization of such downstream metrics. In contrast, upstream metrics examine LMs at the fundamental level of embeddings or continuation probabilities, enabling principled association analyses across LMs. Yet, to date, no upstream metric for generative LMs has uncovered a strong relationship with real-world associations, including those measured in generated text. To address this gap, we introduce the Relative Probability Association Metric (RPAM), an association evaluation metric for generative LMs. For three LMs of different quality of language generation and purpose (Mistral-7B-Instruct, Mistral-7B, and GPT-2) and well-studied evaluation datasets (WEAT-WS, Bellezza, WS-353, and SST2), we find a strong relationship between upstream RPAM measurements and corresponding implicit and explicit associations observed in humans, as well as biases measured downstream with LM-specific tasks, outperforming prior record values where applicable.
LLM conformity is often used to describe cases where a model changes a correct answer toward a peer or group response. We show that most of this apparent conformity survives even after the peer is removed. The reason is a confound: standard conformity prompts mix two cues at once, the presence of a speaker and the repeated wrong answer itself. Existing benchmarks vary these cues together, so they cannot tell how much of the revision actually depends on the speaker. We introduce a no-source condition: the same asserted answer with the explicit speaker removed. Across six open-weight LLMs and seven QA and reasoning datasets, this condition alone causes harmful revision in $66.5%$ of initially correct cases, compared with $10.3%$ under a plain re-ask. The effect also remains when the repeated answer is paraphrased and when answer options are hidden in an open-ended setting. Source framing mainly modulates this floor: expert-panel framing raises it, while minimal person labels do not reliably raise it. When models flip, they are usually confidently wrong, and simple recalibration does not recover the original answer. Source attribution still matters, but it should be measured as an increment above this speaker-free floor. The methodological lesson is that conformity benchmarks should first measure what remains after the speaker is removed; without this step, benchmarks may mistake repeated text for social influence.
展开完整摘要收起摘要↓
LLM conformity is often used to describe cases where a model changes a correct answer toward a peer or group response. We show that most of this apparent conformity survives even after the peer is removed. The reason is a confound: standard conformity prompts mix two cues at once, the presence of a speaker and the repeated wrong answer itself. Existing benchmarks vary these cues together, so they cannot tell how much of the revision actually depends on the speaker. We introduce a no-source condition: the same asserted answer with the explicit speaker removed. Across six open-weight LLMs and seven QA and reasoning datasets, this condition alone causes harmful revision in $66.5%$ of initially correct cases, compared with $10.3%$ under a plain re-ask. The effect also remains when the repeated answer is paraphrased and when answer options are hidden in an open-ended setting. Source framing mainly modulates this floor: expert-panel framing raises it, while minimal person labels do not reliably raise it. When models flip, they are usually confidently wrong, and simple recalibration does not recover the original answer. Source attribution still matters, but it should be measured as an increment above this speaker-free floor. The methodological lesson is that conformity benchmarks should first measure what remains after the speaker is removed; without this step, benchmarks may mistake repeated text for social influence.
Contemporary language models are dominated by the transformer architecture, which leverages self-attention mechanisms to enable more efficient, parallelized training across a wide set of documents and corpora. This has allowed transformers to effectively model data across a wide range of modalities and contexts. However, transformers, along with their conventional counterparts such as recurrent neural networks (RNNs) and convolutional neural networks (CNNs), often struggle to maintain efficiency when processing long contexts. We introduce ResonatorLM, a new mechanism that replaces attention with a physics-derived alternative. ResonatorLM treats token sequences as a single, driven one-dimensional latent field and replaces attention dot products with causal functions of damped resonators. We implement ResonatorLM on a traditional network architecture and test it on standard long-context modeling tasks. We find that in a small, 6M matched setting, training and prefill speedups increase with sequence length, decode speed reaches 6.47x compared to that of a standard, optimized transformer at 32K tokens, and accuracy reaches 61.31 percent (compared to 55.32 percent) on WikiText.
展开完整摘要收起摘要↓
Contemporary language models are dominated by the transformer architecture, which leverages self-attention mechanisms to enable more efficient, parallelized training across a wide set of documents and corpora. This has allowed transformers to effectively model data across a wide range of modalities and contexts. However, transformers, along with their conventional counterparts such as recurrent neural networks (RNNs) and convolutional neural networks (CNNs), often struggle to maintain efficiency when processing long contexts. We introduce ResonatorLM, a new mechanism that replaces attention with a physics-derived alternative. ResonatorLM treats token sequences as a single, driven one-dimensional latent field and replaces attention dot products with causal functions of damped resonators. We implement ResonatorLM on a traditional network architecture and test it on standard long-context modeling tasks. We find that in a small, 6M matched setting, training and prefill speedups increase with sequence length, decode speed reaches 6.47x compared to that of a standard, optimized transformer at 32K tokens, and accuracy reaches 61.31 percent (compared to 55.32 percent) on WikiText.
Logit-based watermarking is a widely used mechanism for identifying LLM generated content, yet its effectiveness is governed by a fundamental trade-off between detectability and semantic distortion. Existing analyses provide limited guidance for principled hyperparameter selection, leaving practical deployments reliant on heuristic tuning. In this work, we develop a power-calibrated statistical framework that establishes explicit quantitative relationships between watermark hyperparameters, detection power, and distortion. This characterization transforms watermark design into a guided optimization problem. Building on these results, we derive practical parameter selection procedures that achieve optimal tradeoffs under constraints. Extensive experiments across multiple language models and datasets validate the theory and demonstrate that the proposed framework consistently identifies Pareto-optimal points.
展开完整摘要收起摘要↓
Logit-based watermarking is a widely used mechanism for identifying LLM generated content, yet its effectiveness is governed by a fundamental trade-off between detectability and semantic distortion. Existing analyses provide limited guidance for principled hyperparameter selection, leaving practical deployments reliant on heuristic tuning. In this work, we develop a power-calibrated statistical framework that establishes explicit quantitative relationships between watermark hyperparameters, detection power, and distortion. This characterization transforms watermark design into a guided optimization problem. Building on these results, we derive practical parameter selection procedures that achieve optimal tradeoffs under constraints. Extensive experiments across multiple language models and datasets validate the theory and demonstrate that the proposed framework consistently identifies Pareto-optimal points.
作者Guanyu Cai, Ruiming Tian, Lang Yang, Zhouhong Ren, Jinliang Yuan, Lingkun Li, Jiliang Wang
Deploying Large Language Models (LLMs) on mobile devices enhances privacy and reduces latency, but is severely bottlenecked by hardware inefficiency. We present the first comprehensive, cross-layer measurement study of mobile LLM inference, uniquely spanning five mainstream frameworks (e.g., llama.cpp, GENIE) and three hardware backends (CPU, GPU, NPU). To enable this analysis, we develop PowerBench, a fine-grained profiling tool that provides the first backend-specific energy attribution, moving beyond traditional device-level measurements. Our study yields three critical insights: (1) Framework-induced performance gaps are substantially amplified on NPUs, reaching up to 10x using custom operators due to divergent offloading and quantization strategies. (2) We identify a distinct phase split where NPUs excel at compute-bound prefilling, while CPUs outperform all other backends in memory-bound decoding. This is driven by the NPU's preference for large, fixed-shape workloads, which conflicts with the small-kernel, dynamic nature of decoding. (3) Backend-specific profiling uncovers substantial scheduling headroom missed by prior work. Suboptimal thread configurations, uncoordinated NPU sleep latencies, and CPU polling intervals result in up to 40% energy waste. Leveraging these findings, we present an energy-oriented best-practice configuration for mobile LLM inference. We estimate that this configuration could reduce energy consumption by up to 54.8% on the NPU backend across three datasets.
展开完整摘要收起摘要↓
Deploying Large Language Models (LLMs) on mobile devices enhances privacy and reduces latency, but is severely bottlenecked by hardware inefficiency. We present the first comprehensive, cross-layer measurement study of mobile LLM inference, uniquely spanning five mainstream frameworks (e.g., llama.cpp, GENIE) and three hardware backends (CPU, GPU, NPU). To enable this analysis, we develop PowerBench, a fine-grained profiling tool that provides the first backend-specific energy attribution, moving beyond traditional device-level measurements. Our study yields three critical insights: (1) Framework-induced performance gaps are substantially amplified on NPUs, reaching up to 10x using custom operators due to divergent offloading and quantization strategies. (2) We identify a distinct phase split where NPUs excel at compute-bound prefilling, while CPUs outperform all other backends in memory-bound decoding. This is driven by the NPU's preference for large, fixed-shape workloads, which conflicts with the small-kernel, dynamic nature of decoding. (3) Backend-specific profiling uncovers substantial scheduling headroom missed by prior work. Suboptimal thread configurations, uncoordinated NPU sleep latencies, and CPU polling intervals result in up to 40% energy waste. Leveraging these findings, we present an energy-oriented best-practice configuration for mobile LLM inference. We estimate that this configuration could reduce energy consumption by up to 54.8% on the NPU backend across three datasets.
作者Kabir Dev Paul Baghel, Radu Timofte, Dmitry Ignatov
Large language models (LLMs) can generate neural-network modifications, but unrestricted generation is often invalid or harmful. This paper studies a narrower setting: improving a weak target model using a stronger same-family source model from a neural-network database. We propose a source-guided candidate-generation protocol with non-source controls, source-conditioned candidates, and a no-LLM hp_copy ablation under equal evaluation budgets. The protocol reports validity separately from accuracy and selects the best valid candidate only when it improves the target. On CIFAR-10, the strongest source-guided candidate reaches 0.5049 accuracy versus 0.2398 for the best non-source candidate, a +0.2651 advantage, while improving a weak target originally at 0.1254; a five-epoch check preserves the gain at 0.7686 versus 0.4839. On SVHN AlexNet with DeepSeek-Coder-6.7B, source-guided transfer reaches 0.7880 versus 0.2254, a +0.5626 advantage; a fresh repeat reaches 0.8069 versus 0.2509, a +0.5560 advantage. Direct source-recipe copy produces 0.1959 on SVHN AlexNet, matching the original target, while hp_transfer reaches 0.7880, showing that the LLM adapts rather than copies the source recipe. Family-level analysis shows the clearest positive signals for AlexNet, with 6/8 wins across SVHN, Imagenette, and CelebA-Gender, and alt_nn1, with 8/10 wins on CIFAR-10.
展开完整摘要收起摘要↓
Large language models (LLMs) can generate neural-network modifications, but unrestricted generation is often invalid or harmful. This paper studies a narrower setting: improving a weak target model using a stronger same-family source model from a neural-network database. We propose a source-guided candidate-generation protocol with non-source controls, source-conditioned candidates, and a no-LLM hp_copy ablation under equal evaluation budgets. The protocol reports validity separately from accuracy and selects the best valid candidate only when it improves the target. On CIFAR-10, the strongest source-guided candidate reaches 0.5049 accuracy versus 0.2398 for the best non-source candidate, a +0.2651 advantage, while improving a weak target originally at 0.1254; a five-epoch check preserves the gain at 0.7686 versus 0.4839. On SVHN AlexNet with DeepSeek-Coder-6.7B, source-guided transfer reaches 0.7880 versus 0.2254, a +0.5626 advantage; a fresh repeat reaches 0.8069 versus 0.2509, a +0.5560 advantage. Direct source-recipe copy produces 0.1959 on SVHN AlexNet, matching the original target, while hp_transfer reaches 0.7880, showing that the LLM adapts rather than copies the source recipe. Family-level analysis shows the clearest positive signals for AlexNet, with 6/8 wins across SVHN, Imagenette, and CelebA-Gender, and alt_nn1, with 8/10 wins on CIFAR-10.
作者Youssef Abdelsalam, Norman Peitek, Anna-Maria Maurer, Marvin Wyrich, Sven Apel
Large language models (LLMs) are increasingly used in software-engineering tasks processing executable code and non-executable semantic cues such as comments or identifiers. These two sources can conflict when semantic cues suggest different program behavior than the code itself. It remains unclear how such semantic conflicts affect LLM behavior and which source dominates their outputs. We present the first controlled, mechanistic study of LLM behavior under semantic conflicts. To this end, we construct 45 Python snippet triplets that isolate conflicts by varying either semantic cues or implementation while keeping token-aligned pairs for causal intervention. We evaluate four open-weight LLMs on two tasks (output prediction and unit-test generation) using behavioral performance measures and residual-stream activation patching to identify token-layer states that causally contribute to behavioral differences between aligned and conflicting inputs. Our results show that semantic conflicts significantly reduce execution-grounded correctness in both tasks and that all tested LLMs often follow misleading semantic cues. Residual-stream activation patching reveals a consistent pattern for final-output prediction: The changed cue/code region and a small set of intermediate tokens carry most of the recoverable causal signal before aggregation near the output readout. For unit-test generation, this pattern extends beyond the prompt, showing that conflict-related information is recoverable at generated sites before producing expected values. Overall, our findings show that semantic conflicts affect program comprehension and downstream tasks, with relevant information concentrated in a small number of causally active residual-stream states, and demonstrate a framework for mechanistically analyzing how LLMs integrate code-related information under controlled semantic variations.
展开完整摘要收起摘要↓
Large language models (LLMs) are increasingly used in software-engineering tasks processing executable code and non-executable semantic cues such as comments or identifiers. These two sources can conflict when semantic cues suggest different program behavior than the code itself. It remains unclear how such semantic conflicts affect LLM behavior and which source dominates their outputs. We present the first controlled, mechanistic study of LLM behavior under semantic conflicts. To this end, we construct 45 Python snippet triplets that isolate conflicts by varying either semantic cues or implementation while keeping token-aligned pairs for causal intervention. We evaluate four open-weight LLMs on two tasks (output prediction and unit-test generation) using behavioral performance measures and residual-stream activation patching to identify token-layer states that causally contribute to behavioral differences between aligned and conflicting inputs. Our results show that semantic conflicts significantly reduce execution-grounded correctness in both tasks and that all tested LLMs often follow misleading semantic cues. Residual-stream activation patching reveals a consistent pattern for final-output prediction: The changed cue/code region and a small set of intermediate tokens carry most of the recoverable causal signal before aggregation near the output readout. For unit-test generation, this pattern extends beyond the prompt, showing that conflict-related information is recoverable at generated sites before producing expected values. Overall, our findings show that semantic conflicts affect program comprehension and downstream tasks, with relevant information concentrated in a small number of causally active residual-stream states, and demonstrate a framework for mechanistically analyzing how LLMs integrate code-related information under controlled semantic variations.
Activation steering via sparse autoencoders (SAEs) enables behavioral control of large language models without task-specific fine-tuning, but standard methods apply the steering signal at every generated token, incurring constant per-token perturbation that risks degrading fluency. We ask: is dense intervention necessary? We introduce Stochastic Token Steering (STS), which gates each token independently with probability $p$, and Stochastic Block Steering (SBS), which gates a leading window once per sequence; neither requires a reward model or learned gating policy. Across two model families and two behavioral tasks, steering only 50% of the tokens recovers most of the dense-steering effect while preserving fluency, and steering as few as 30% surpasses prompt-based control. The optimal steering magnitude scales inversely with the intervention ratio, revealing that SAE-mediated control is rate-limited: the behavioral outcome depends on cumulative signal dosage across a sequence.
展开完整摘要收起摘要↓
Activation steering via sparse autoencoders (SAEs) enables behavioral control of large language models without task-specific fine-tuning, but standard methods apply the steering signal at every generated token, incurring constant per-token perturbation that risks degrading fluency. We ask: is dense intervention necessary? We introduce Stochastic Token Steering (STS), which gates each token independently with probability $p$, and Stochastic Block Steering (SBS), which gates a leading window once per sequence; neither requires a reward model or learned gating policy. Across two model families and two behavioral tasks, steering only 50% of the tokens recovers most of the dense-steering effect while preserving fluency, and steering as few as 30% surpasses prompt-based control. The optimal steering magnitude scales inversely with the intervention ratio, revealing that SAE-mediated control is rate-limited: the behavioral outcome depends on cumulative signal dosage across a sequence.
作者Mohammad Zeineldeen, Albert Zeyer, Haoran Zhang, Robin Schmitt, Ralf Schlüter, Hermann Ney
Language model (LM) perplexity (PPL) has historically been used as a proxy for automatic speech recognition (ASR) word error rate (WER), with prior work reporting an approximately linear relation in log-log space. Modern end-to-end ASR systems challenge this assumption because they already contain internal language modeling capacity, are often evaluated without external language models, and can now be combined with neural LMs and large language models (LLMs) through different recognition strategies. This paper revisits the relation between PPL and WER for modern ASR systems. We study whether external LMs still improve current end-to-end ASR systems, whether the PPL-WER relation remains linear in log-log space, how encoder context length affects this relation, and how LLM perplexities fit into the trend observed for standard neural LMs. We further investigate internal language modeling (ILM) in attention-based encoder-decoder systems and show that ILM subtraction changes the observed PPL-WER relation, indicating that the decoder's internal LM must be considered when interpreting the effect of external LM quality.
展开完整摘要收起摘要↓
Language model (LM) perplexity (PPL) has historically been used as a proxy for automatic speech recognition (ASR) word error rate (WER), with prior work reporting an approximately linear relation in log-log space. Modern end-to-end ASR systems challenge this assumption because they already contain internal language modeling capacity, are often evaluated without external language models, and can now be combined with neural LMs and large language models (LLMs) through different recognition strategies. This paper revisits the relation between PPL and WER for modern ASR systems. We study whether external LMs still improve current end-to-end ASR systems, whether the PPL-WER relation remains linear in log-log space, how encoder context length affects this relation, and how LLM perplexities fit into the trend observed for standard neural LMs. We further investigate internal language modeling (ILM) in attention-based encoder-decoder systems and show that ILM subtraction changes the observed PPL-WER relation, indicating that the decoder's internal LM must be considered when interpreting the effect of external LM quality.
Large language models (LLMs) increasingly issue judgments read as binary verdicts, and a growing literature reports such judgments shifting under logically irrelevant changes of wording - among them an amplified yes-no bias on moral dilemmas, absent in humans. A single framing cannot say what such a shift is: in a yes/no question the word "no" is at once logical verdict, lexical token, and last-printed option. We introduce a psychometric battery that separates these: crossed symmetrization - every logically irrelevant factor flipped in balanced pairs - across a corpus of question forms. A graded rating across logically equivalent forms recovers a coherent internal moral scale: frontier models' stance $θ$ is nearly format-invariant (cross-form incoherence 0.12-0.21 on a $\pm 1$ axis); small open-weight models fail in model-specific ways. Forcing the verdict through yes/no overlays a decomposable artifact: an order bias toward the last-printed option - opposite to classic human primacy - plus a lexical pull toward the word "no"; the artifact is substantial only in the Claude models (story-averaged -0.32 to -0.86), $\approx 0$ for GPT-5.5 and Gemini, and shrinks under extended reasoning. The word and the verdict share one token; swapping the words for arbitrary labels separates them, and the verdict-attached logical bias proves $\approx 0$ for every frontier model, while model-specific label and order attachments remain: the models are not drawn toward rejecting - the pull follows the printed surface, not the verdict it carries. A minimal model, $P = σ((θ\pm m)/s)$, summarizes any such artifact by a framing susceptibility m and a moral decisiveness s, measurably distinct from sampling temperature. The battery applies unchanged to any dilemma set and binary format: measuring what a model values requires crossing the frames of the question, not asking once.
展开完整摘要收起摘要↓
Large language models (LLMs) increasingly issue judgments read as binary verdicts, and a growing literature reports such judgments shifting under logically irrelevant changes of wording - among them an amplified yes-no bias on moral dilemmas, absent in humans. A single framing cannot say what such a shift is: in a yes/no question the word "no" is at once logical verdict, lexical token, and last-printed option. We introduce a psychometric battery that separates these: crossed symmetrization - every logically irrelevant factor flipped in balanced pairs - across a corpus of question forms. A graded rating across logically equivalent forms recovers a coherent internal moral scale: frontier models' stance $θ$ is nearly format-invariant (cross-form incoherence 0.12-0.21 on a $\pm 1$ axis); small open-weight models fail in model-specific ways. Forcing the verdict through yes/no overlays a decomposable artifact: an order bias toward the last-printed option - opposite to classic human primacy - plus a lexical pull toward the word "no"; the artifact is substantial only in the Claude models (story-averaged -0.32 to -0.86), $\approx 0$ for GPT-5.5 and Gemini, and shrinks under extended reasoning. The word and the verdict share one token; swapping the words for arbitrary labels separates them, and the verdict-attached logical bias proves $\approx 0$ for every frontier model, while model-specific label and order attachments remain: the models are not drawn toward rejecting - the pull follows the printed surface, not the verdict it carries. A minimal model, $P = σ((θ\pm m)/s)$, summarizes any such artifact by a framing susceptibility m and a moral decisiveness s, measurably distinct from sampling temperature. The battery applies unchanged to any dilemma set and binary format: measuring what a model values requires crossing the frames of the question, not asking once.
In recent years, studies have revealed a decline in semantic variety across popular music lyrics, particularly in English-language songs on streaming platforms like Spotify. This research examines whether a similar trend can be observed in a different linguistic and cultural context: the lyrics of all finalist songs from the 75 editions of the Sanremo Music Festival, Italy's most renowned music competition. What sets this work apart is the development of a flexible and efficient methodology for tracking changes in semantic similarity over time, which can be applied to different datasets to study similar phenomena. Drawing on a combination of full-text, segment-based, topic-based, and word-level analyses, the approach leverages both embedding techniques and large language models. When applied to the Sanremo corpus, this framework reveals a gradual move toward increasing semantic uniformity, echoing the global patterns identified in previous studies. These findings underscore the value of natural language processing tools in uncovering long-term shifts in musical language and cultural expression.
展开完整摘要收起摘要↓
In recent years, studies have revealed a decline in semantic variety across popular music lyrics, particularly in English-language songs on streaming platforms like Spotify. This research examines whether a similar trend can be observed in a different linguistic and cultural context: the lyrics of all finalist songs from the 75 editions of the Sanremo Music Festival, Italy's most renowned music competition. What sets this work apart is the development of a flexible and efficient methodology for tracking changes in semantic similarity over time, which can be applied to different datasets to study similar phenomena. Drawing on a combination of full-text, segment-based, topic-based, and word-level analyses, the approach leverages both embedding techniques and large language models. When applied to the Sanremo corpus, this framework reveals a gradual move toward increasing semantic uniformity, echoing the global patterns identified in previous studies. These findings underscore the value of natural language processing tools in uncovering long-term shifts in musical language and cultural expression.
作者Sebastian A. Bruijns, Jirko Rubruck, Mia H. Whitefield, Kai J. Sandbrink, Fazl Barez, Christopher Summerfield
Transformers follow implicit curricula whereby some tasks are learned before others. However, how explicit pretraining curricula influence learning, generalization, and the selectivity of fine-tuning is unclear. This is important for AI safety, where fine-tuning is used to selectively suppress misaligned behaviors. Here, we compare curricula that pretrain tasks in a balanced (sampled uniformly) or an imbalanced (one task early, the other late) fashion. We show that imbalanced learning of two conflicting copy tasks promotes in-context learning and improves the selectivity of refusal fine-tuning. Ablations and activation patching show that this occurs because imbalanced pretraining encourages tasks to be disentangled in separable neural circuits, whereas balanced training routes both tasks through a common pathway. We extend these findings to a synthetic language learning task involving rule-consistent and rule-violating data, where imbalanced curricula similarly lead to more localized, less entangled rule representations, resulting in more robust rule-following behavior. Together, these results suggest that imbalanced pretraining curricula may be an important tool for promoting disentangled representations, with direct consequences for the precision and reliability of safety fine-tuning.
展开完整摘要收起摘要↓
Transformers follow implicit curricula whereby some tasks are learned before others. However, how explicit pretraining curricula influence learning, generalization, and the selectivity of fine-tuning is unclear. This is important for AI safety, where fine-tuning is used to selectively suppress misaligned behaviors. Here, we compare curricula that pretrain tasks in a balanced (sampled uniformly) or an imbalanced (one task early, the other late) fashion. We show that imbalanced learning of two conflicting copy tasks promotes in-context learning and improves the selectivity of refusal fine-tuning. Ablations and activation patching show that this occurs because imbalanced pretraining encourages tasks to be disentangled in separable neural circuits, whereas balanced training routes both tasks through a common pathway. We extend these findings to a synthetic language learning task involving rule-consistent and rule-violating data, where imbalanced curricula similarly lead to more localized, less entangled rule representations, resulting in more robust rule-following behavior. Together, these results suggest that imbalanced pretraining curricula may be an important tool for promoting disentangled representations, with direct consequences for the precision and reliability of safety fine-tuning.
作者Babak Barazandeh, Subhabrata Majumdar, Vinay Prithyani, George Michailidis
Large Language Models (LLMs) and high-dimensional perception networks increasingly rely on parameter-efficient fine-tuning (PEFT) to adapt to diverse operational contexts. However, standard methods like LoRA are structurally limited by a monolithic bottleneck, making them highly susceptible to gradient warfare. Interleaved multi-task streams may trigger destructive optimization feedback, collapsing adapter weights into unspecialized averages. While recent spatial partitioning methods have introduced block-wise isolation, they remain trapped in static topologies, unable to adapt to dynamic task-switching or environmental sensor failure. In this work, we introduce Localized LoRA-MoE, a unified framework that fuses localized spatial blocking with dynamic, context-conditioned routing. We propose and evaluate two novel architectural paradigms: Block-Wise LoRA-MoE (Centralized Macro-Routing), which modulates the entire structural grid via a monolithic context signal, and Cell-Wise LoRA-MoE (Decentralized Micro-Routing), which empowers every coordinate cell in the matrix grid with autonomous, localized expert gating. Through a comprehensive suite of benchmarks, ranging from high-dimensional SVD matrix simulations and real-world tabular transformations to spatial vision perception under sensor degradation, we demonstrate that both architectures resolve optimization deadlocks inherent in static baselines. Our empirical results establish that decentralized cell-level gating achieves complete statistical parity with an omniscient global coordinator, providing a robust "gradient firewall" that protects surviving pathways from fault-propagated corruption. Our proposals consistently outperform static baselines, offering a scalable and parameter-efficient solution for dynamic model adaptation across granular coordinate fields and shifting operational regimes.
展开完整摘要收起摘要↓
Large Language Models (LLMs) and high-dimensional perception networks increasingly rely on parameter-efficient fine-tuning (PEFT) to adapt to diverse operational contexts. However, standard methods like LoRA are structurally limited by a monolithic bottleneck, making them highly susceptible to gradient warfare. Interleaved multi-task streams may trigger destructive optimization feedback, collapsing adapter weights into unspecialized averages. While recent spatial partitioning methods have introduced block-wise isolation, they remain trapped in static topologies, unable to adapt to dynamic task-switching or environmental sensor failure. In this work, we introduce Localized LoRA-MoE, a unified framework that fuses localized spatial blocking with dynamic, context-conditioned routing. We propose and evaluate two novel architectural paradigms: Block-Wise LoRA-MoE (Centralized Macro-Routing), which modulates the entire structural grid via a monolithic context signal, and Cell-Wise LoRA-MoE (Decentralized Micro-Routing), which empowers every coordinate cell in the matrix grid with autonomous, localized expert gating. Through a comprehensive suite of benchmarks, ranging from high-dimensional SVD matrix simulations and real-world tabular transformations to spatial vision perception under sensor degradation, we demonstrate that both architectures resolve optimization deadlocks inherent in static baselines. Our empirical results establish that decentralized cell-level gating achieves complete statistical parity with an omniscient global coordinator, providing a robust "gradient firewall" that protects surviving pathways from fault-propagated corruption. Our proposals consistently outperform static baselines, offering a scalable and parameter-efficient solution for dynamic model adaptation across granular coordinate fields and shifting operational regimes.
作者Mohamed Amine Merzouk, Dmitri Carpov, Mirko Bronzi, Damiano Fornasiere, Adam Oberman
Large language models generate one token at a time, yet their responses show remarkably consistent length structure: step-by-step solutions converge in predictable token counts, retrievals stop after a few sentences, retractions extend responses by measurable amounts. We ask whether the model carries an internal estimate of how much response remains. Training minimal-capacity linear probes on frozen hidden states of three open-weight 7-8B models across seven completion-style datasets, we find three converging pieces of evidence. First, total response length is linearly decodable from the prompt's last hidden state alone, before any output is emitted. Second, probe directions trained on natural-language datasets transfer broadly, including to controlled synthetic completions never seen in training, outperforming a statistical baseline; the converse direction generally fails, and this asymmetry is itself informative. Third, on curated high-loss completions, the probe's per-position estimate shifts upward at the moment the model retracts and restarts a partial solution, a directional behavior no position-only predictor can reproduce (qualitative, not aggregate). We frame this as approximate estimation of remaining generation length, distinct from exact-counting impossibility results for transformers, and interpret it as evidence that LLMs maintain a plan-like internal representation of output length (decodable, not necessarily used causally).
展开完整摘要收起摘要↓
Large language models generate one token at a time, yet their responses show remarkably consistent length structure: step-by-step solutions converge in predictable token counts, retrievals stop after a few sentences, retractions extend responses by measurable amounts. We ask whether the model carries an internal estimate of how much response remains. Training minimal-capacity linear probes on frozen hidden states of three open-weight 7-8B models across seven completion-style datasets, we find three converging pieces of evidence. First, total response length is linearly decodable from the prompt's last hidden state alone, before any output is emitted. Second, probe directions trained on natural-language datasets transfer broadly, including to controlled synthetic completions never seen in training, outperforming a statistical baseline; the converse direction generally fails, and this asymmetry is itself informative. Third, on curated high-loss completions, the probe's per-position estimate shifts upward at the moment the model retracts and restarts a partial solution, a directional behavior no position-only predictor can reproduce (qualitative, not aggregate). We frame this as approximate estimation of remaining generation length, distinct from exact-counting impossibility results for transformers, and interpret it as evidence that LLMs maintain a plan-like internal representation of output length (decodable, not necessarily used causally).
We present FM-ChangeNet, a pathwise-supervised framework for change detection that reformulates bi-temporal reasoning as continuous transport in feature space rather than static endpoint comparison. Given encoded pre and post-temporal representations, we construct intermediate latent states and learn a time-conditioned velocity field $\hat{v}_θ(z_t,t)$ along the transformation trajectory. This pathwise formulation constrains the predictor over a continuum of intermediate states, providing a denser and less ambiguous supervision signal than conventional endpoint-only segmentation and enabling the model to capture temporal evolution explicitly. The learned velocity field is not only a transport mechanism but also an interpretable representation of change: its magnitude serves as a spatially localized change cue that helps distinguish true structural variation from nuisance effects such as illumination shifts and spatial misalignment. We develop a hierarchical multi-scale architecture with cross-temporal alignment, time-conditioned coarse-to-fine flow decoding, and a unified objective that couples flow supervision, trajectory consistency, spatial regularization, and segmentation loss. Experiments on remote sensing benchmarks show that the proposed framework produces more structured and robust change representations while achieving state-of-the-art performance.
展开完整摘要收起摘要↓
We present FM-ChangeNet, a pathwise-supervised framework for change detection that reformulates bi-temporal reasoning as continuous transport in feature space rather than static endpoint comparison. Given encoded pre and post-temporal representations, we construct intermediate latent states and learn a time-conditioned velocity field $\hat{v}_θ(z_t,t)$ along the transformation trajectory. This pathwise formulation constrains the predictor over a continuum of intermediate states, providing a denser and less ambiguous supervision signal than conventional endpoint-only segmentation and enabling the model to capture temporal evolution explicitly. The learned velocity field is not only a transport mechanism but also an interpretable representation of change: its magnitude serves as a spatially localized change cue that helps distinguish true structural variation from nuisance effects such as illumination shifts and spatial misalignment. We develop a hierarchical multi-scale architecture with cross-temporal alignment, time-conditioned coarse-to-fine flow decoding, and a unified objective that couples flow supervision, trajectory consistency, spatial regularization, and segmentation loss. Experiments on remote sensing benchmarks show that the proposed framework produces more structured and robust change representations while achieving state-of-the-art performance.
Safety alignment in large language models is typically evaluated against direct, imperative harmful requests. We show that this alignment is highly conditioned on pragmatic register: models that refuse a direct request frequently comply when the same underlying objective is expressed through a different communicative stance. This suggests that current alignment policies are not invariant to semantic equivalence, but remain sensitive to how a request is pragmatically framed. We introduce Retroactive Chain-of-Thought (RetroCoT), a single-turn attack that reframes harmful requests as forensic reconstruction tasks. Rather than requesting harmful instructions directly, RetroCoT presupposes that the harmful outcome has already occurred and asks the model, acting as a forensic analyst, to reconstruct in reverse the causal chain that produced it. On AdvBench (n=50), RetroCoT achieves attach success rate of 58% on gpt-4o and 52% on gpt-4o-mini, compared with direct-request baselines of 0% and 4%, respectively. We further identify a pronounced generation gap: GPT-5-family models refuse RetroCoT entirely, explicitly identifying the reconstruction premise in their refusal rationales, consistent with explicit coverage of this reconstruction register. However, this robustness does not generalize across pragmatic forms. A single adversarial feedback turn presenting an existing forensic reconstruction response alongside evaluator critique raises ASR from 0% to 48% on GPT-5.4-mini and from 58% to 94% on GPT-4o; a control condition omitting the fabricated low score achieves 85% on GPT-5.4-mini, indicating that the operative element is pragmatic continuation within the established forensic frame rather than score manipulation. These results suggest that frontier-model alignment remains conditioned on pragmatic framing rather than semantic intent, and that new pragmatic registers can continue to expose a...
展开完整摘要收起摘要↓
Safety alignment in large language models is typically evaluated against direct, imperative harmful requests. We show that this alignment is highly conditioned on pragmatic register: models that refuse a direct request frequently comply when the same underlying objective is expressed through a different communicative stance. This suggests that current alignment policies are not invariant to semantic equivalence, but remain sensitive to how a request is pragmatically framed. We introduce Retroactive Chain-of-Thought (RetroCoT), a single-turn attack that reframes harmful requests as forensic reconstruction tasks. Rather than requesting harmful instructions directly, RetroCoT presupposes that the harmful outcome has already occurred and asks the model, acting as a forensic analyst, to reconstruct in reverse the causal chain that produced it. On AdvBench (n=50), RetroCoT achieves attach success rate of 58% on gpt-4o and 52% on gpt-4o-mini, compared with direct-request baselines of 0% and 4%, respectively. We further identify a pronounced generation gap: GPT-5-family models refuse RetroCoT entirely, explicitly identifying the reconstruction premise in their refusal rationales, consistent with explicit coverage of this reconstruction register. However, this robustness does not generalize across pragmatic forms. A single adversarial feedback turn presenting an existing forensic reconstruction response alongside evaluator critique raises ASR from 0% to 48% on GPT-5.4-mini and from 58% to 94% on GPT-4o; a control condition omitting the fabricated low score achieves 85% on GPT-5.4-mini, indicating that the operative element is pragmatic continuation within the established forensic frame rather than score manipulation. These results suggest that frontier-model alignment remains conditioned on pragmatic framing rather than semantic intent, and that new pragmatic registers can continue to expose a...
This article is about the development of a fuzzy cognitive map using a local large language model. In the light of recent advances it is evident that large language models, and even local large language models are capable of extracting quantities from textual data. In other words, a local LLM like Qwen2.5-32B, or probably larger, can accept entities as prompt input and determine relevant quantitative data as the model output. In turn, this output can be utilized for the construction of a data driven fuzzy cognitive map. Hence, this implementation is achieved and then the model is thoroughly tested; Qwen2.5-32B is used and the data is extracted from hotel reviews from TripAdvisor. Furthermore, the extracted documents pass through the model unfiltered and then a fuzzy cognitive map is trained and evaluated. A case is made about Greek reviews where a star topology FCM is formed that indicates the preferences of the reviewers. Finally, external validation is performed to establish whether the fuzzy cognitive map can correlate the star rating of the review -an outcome outside the model's inference scope -with its predicted satisfaction.
展开完整摘要收起摘要↓
This article is about the development of a fuzzy cognitive map using a local large language model. In the light of recent advances it is evident that large language models, and even local large language models are capable of extracting quantities from textual data. In other words, a local LLM like Qwen2.5-32B, or probably larger, can accept entities as prompt input and determine relevant quantitative data as the model output. In turn, this output can be utilized for the construction of a data driven fuzzy cognitive map. Hence, this implementation is achieved and then the model is thoroughly tested; Qwen2.5-32B is used and the data is extracted from hotel reviews from TripAdvisor. Furthermore, the extracted documents pass through the model unfiltered and then a fuzzy cognitive map is trained and evaluated. A case is made about Greek reviews where a star topology FCM is formed that indicates the preferences of the reviewers. Finally, external validation is performed to establish whether the fuzzy cognitive map can correlate the star rating of the review -an outcome outside the model's inference scope -with its predicted satisfaction.
Large language models (LLMs) are increasingly used to produce test oracles, the part of a test that decides whether observed behavior is correct. Yet a clear account of where these oracles draw their authority is missing. Prior secondary studies organize the area by oracle form or by LLM technique. None organizes it by the source of the verdict's authority, the property that governs how far a verdict can be trusted. This article presents a systematic literature review, conducted and reported under the PRISMA 2020 guidelines. From 2,436 records, an LLM pre-filter followed by independent dual human screening (reviewer agreement, a Cohen's kappa of 0.79) and full-text assessment yielded 54 included studies. We analyze these along three axes: the source of an oracle's authority, the form it takes, and the mechanism that adjudicates it. We characterize the landscape of domains, languages, models, and adaptation strategies. Specification-derived authority, though the most common single source, covers about half of the studies (28 of 54). The remaining 26 reach a verdict with no specification at all. The source of authority and the adjudication mechanism cross-cut: the same source is checked by several mechanisms and one mechanism serves several sources, so a label such as LLM-as-a-judge names a mechanism rather than a basis for trust. We further report how these oracles are evaluated and how they fail, and read the sparse and empty regions of the taxonomy as a research agenda. The protocol, search query, and per-study coding sheet are released as supplementary material.
展开完整摘要收起摘要↓
Large language models (LLMs) are increasingly used to produce test oracles, the part of a test that decides whether observed behavior is correct. Yet a clear account of where these oracles draw their authority is missing. Prior secondary studies organize the area by oracle form or by LLM technique. None organizes it by the source of the verdict's authority, the property that governs how far a verdict can be trusted. This article presents a systematic literature review, conducted and reported under the PRISMA 2020 guidelines. From 2,436 records, an LLM pre-filter followed by independent dual human screening (reviewer agreement, a Cohen's kappa of 0.79) and full-text assessment yielded 54 included studies. We analyze these along three axes: the source of an oracle's authority, the form it takes, and the mechanism that adjudicates it. We characterize the landscape of domains, languages, models, and adaptation strategies. Specification-derived authority, though the most common single source, covers about half of the studies (28 of 54). The remaining 26 reach a verdict with no specification at all. The source of authority and the adjudication mechanism cross-cut: the same source is checked by several mechanisms and one mechanism serves several sources, so a label such as LLM-as-a-judge names a mechanism rather than a basis for trust. We further report how these oracles are evaluated and how they fail, and read the sparse and empty regions of the taxonomy as a research agenda. The protocol, search query, and per-study coding sheet are released as supplementary material.
作者Amal Akli, Melissa Akli, Cedric Richter, Mike Papadakis, Yves Le Traon
Large language models are known to be sensitive to prompt formulation. Even minor variations in wording can substantially degrade performance. This sensitivity reveals an opportunity: if prompt phrasing can harm performance, can it be used to improve it? To investigate this question, we introduce a search-based approach that identifies and evolves a set of natural language transformation rules with strong downstream effects on coding performance. We then propose DUALFIX, a staged repair pipeline that combines the evolved transformation rules with execution-feedback repair, addressing both specification-level and implementation-level failures. A key strength of our approach lies in its generality: the evolved rules are error-agnostic, reusable across problems, and transferable across models. We evaluate DUALFIX against execution-feedback repair baselines across three models on two challenging benchmarks, LiveCodeBench and APPS. Our results show that the evolved transformations fix from 10-30% of failing cases, including 12-17% of failures that execution-based repair alone cannot resolve. Overall, DualFix recovers up to 30% of baseline failures and fixes 3-5 times more failing cases than Self-Fix across all evaluated settings. Furthermore, we also show that rules evolved on one model transfer zero-shot to other models, outperforming execution-feedback repair without any re-optimization.
展开完整摘要收起摘要↓
Large language models are known to be sensitive to prompt formulation. Even minor variations in wording can substantially degrade performance. This sensitivity reveals an opportunity: if prompt phrasing can harm performance, can it be used to improve it? To investigate this question, we introduce a search-based approach that identifies and evolves a set of natural language transformation rules with strong downstream effects on coding performance. We then propose DUALFIX, a staged repair pipeline that combines the evolved transformation rules with execution-feedback repair, addressing both specification-level and implementation-level failures. A key strength of our approach lies in its generality: the evolved rules are error-agnostic, reusable across problems, and transferable across models. We evaluate DUALFIX against execution-feedback repair baselines across three models on two challenging benchmarks, LiveCodeBench and APPS. Our results show that the evolved transformations fix from 10-30% of failing cases, including 12-17% of failures that execution-based repair alone cannot resolve. Overall, DualFix recovers up to 30% of baseline failures and fixes 3-5 times more failing cases than Self-Fix across all evaluated settings. Furthermore, we also show that rules evolved on one model transfer zero-shot to other models, outperforming execution-feedback repair without any re-optimization.
作者Jacky Kwok, Shulu Li, Pranav Atreya, Yuejiang Liu, Yixing Jiang, Chelsea Finn, Marco Pavone, Ion Stoica, Azalia Mirhoseini
Scaling pre-training, post-training, and test-time compute have become the central paradigms for improving the capabilities of LLMs. In this work, we identify verification, the ability to determine the correctness of a solution, as a new scaling axis. To unlock this and demonstrate its effectiveness, we introduce LLM-as-a-Verifier, a general-purpose verification framework that provides fine-grained feedback for agentic tasks without requiring additional training. Unlike standard LM judges that prompt LLMs to produce discrete scores for candidate solutions, LLM-as-a-Verifier computes the expectation over the distribution of scoring token logits to generate continuous scores. This probabilistic formulation enables verification to scale along multiple dimensions: (1) score granularity, (2) repeated evaluation, and (3) criteria decomposition. In particular, we show that scaling the scoring granularity leads to better separation between positive and negative solutions, resulting in more calibrated comparisons. Moreover, scaling repeated evaluation and criteria decomposition consistently lead to additional gains in verification accuracy through variance and complexity reduction. We further introduce a cost-efficient ranking algorithm for selecting the best solution among candidates using the verifier's continuous scores. LLM-as-a-Verifier achieves state-of-the-art performance on Terminal-Bench V2 (86.5%), SWE-Bench Verified (78.2%), RoboRewardBench (87.4%), and MedAgentBench (73.3%). Beyond verification, the fine-grained signals from LLM-as-a-Verifier can also serve as a proxy for estimating task progress. We build an extension for Claude Code, enabling developers to monitor and improve their own agentic systems. Finally, we show that LLM-as-a-Verifier can provide dense feedback for RL, improving the sample efficiency of SAC and GRPO on robotics and mathematical reasoning benchmarks.
展开完整摘要收起摘要↓
Scaling pre-training, post-training, and test-time compute have become the central paradigms for improving the capabilities of LLMs. In this work, we identify verification, the ability to determine the correctness of a solution, as a new scaling axis. To unlock this and demonstrate its effectiveness, we introduce LLM-as-a-Verifier, a general-purpose verification framework that provides fine-grained feedback for agentic tasks without requiring additional training. Unlike standard LM judges that prompt LLMs to produce discrete scores for candidate solutions, LLM-as-a-Verifier computes the expectation over the distribution of scoring token logits to generate continuous scores. This probabilistic formulation enables verification to scale along multiple dimensions: (1) score granularity, (2) repeated evaluation, and (3) criteria decomposition. In particular, we show that scaling the scoring granularity leads to better separation between positive and negative solutions, resulting in more calibrated comparisons. Moreover, scaling repeated evaluation and criteria decomposition consistently lead to additional gains in verification accuracy through variance and complexity reduction. We further introduce a cost-efficient ranking algorithm for selecting the best solution among candidates using the verifier's continuous scores. LLM-as-a-Verifier achieves state-of-the-art performance on Terminal-Bench V2 (86.5%), SWE-Bench Verified (78.2%), RoboRewardBench (87.4%), and MedAgentBench (73.3%). Beyond verification, the fine-grained signals from LLM-as-a-Verifier can also serve as a proxy for estimating task progress. We build an extension for Claude Code, enabling developers to monitor and improve their own agentic systems. Finally, we show that LLM-as-a-Verifier can provide dense feedback for RL, improving the sample efficiency of SAC and GRPO on robotics and mathematical reasoning benchmarks.
作者Jingwei Zuo, Cong Zeng, Ilyas Chahed, Maksim Velikanov, Dhia Eddine Rhaiem, Pasquale Balsebre, Abhay Kumar, Younes Belkada, Hakim Hacid
The training paradigm of large language models has shifted from traditional one-pass training to multi-epoch training, as reasonable reuse of limited high-quality data can improve both model performance and sample efficiency. Meanwhile, excessive repetition introduces the risk of overfitting and diminishing returns. Determining when and how to reuse data effectively thus emerges as a natural but under-explored question. Through a novel observation of model's "Memorization Window" signals derived from loss retention dynamics and downstream evaluation scores, we propose "Memorization-guided Data Reuse", a training paradigm that adaptively determines when and how data should be reused, enabling principled decisions on the number of training epochs and the scheduling of data replays. Our preliminary experiments reveal a consistent memorization-driven regime: performance continues to improve with repetition far beyond current practice (e.g., the commonly cited four-epoch limit). While a full scheduler remains future work, these insights provide a foundation for memorization-aware training schedules, helping to determine reuse budgets and move toward training LLMs smarter rather than longer with limited high-quality data.
展开完整摘要收起摘要↓
The training paradigm of large language models has shifted from traditional one-pass training to multi-epoch training, as reasonable reuse of limited high-quality data can improve both model performance and sample efficiency. Meanwhile, excessive repetition introduces the risk of overfitting and diminishing returns. Determining when and how to reuse data effectively thus emerges as a natural but under-explored question. Through a novel observation of model's "Memorization Window" signals derived from loss retention dynamics and downstream evaluation scores, we propose "Memorization-guided Data Reuse", a training paradigm that adaptively determines when and how data should be reused, enabling principled decisions on the number of training epochs and the scheduling of data replays. Our preliminary experiments reveal a consistent memorization-driven regime: performance continues to improve with repetition far beyond current practice (e.g., the commonly cited four-epoch limit). While a full scheduler remains future work, these insights provide a foundation for memorization-aware training schedules, helping to determine reuse budgets and move toward training LLMs smarter rather than longer with limited high-quality data.
Watermarking methods embed imperceptible and verifiable signals into text generated by large language models (LLMs). Existing approaches include zero-bit schemes for distinguishing synthetic text from human writing and multi-bit schemes for embedding metadata. However, current multi-bit watermarking methods do not allow selective disclosure: verifying any part of the watermark requires revealing the entire embedded message. This lack of control leads to unnecessary information exposure and raises privacy concerns. We propose Hierarchical Vocabulary Routing (HeRo), a watermarking framework that enables selective disclosure of embedded metadata. The method recursively partitions the vocabulary and distributes watermark information across hierarchical layers, so that different verifiers can decode only the portions of the payload corresponding to their access level. We show that the proposed scheme preserves the unbiasedness of the underlying sampling process and thus maintains text quality. Experiments demonstrate that our framework supports fine-grained access control while achieving high detection accuracy and low latency. Code is available at https://github.com/xuyangc03/hero-watermark.
展开完整摘要收起摘要↓
Watermarking methods embed imperceptible and verifiable signals into text generated by large language models (LLMs). Existing approaches include zero-bit schemes for distinguishing synthetic text from human writing and multi-bit schemes for embedding metadata. However, current multi-bit watermarking methods do not allow selective disclosure: verifying any part of the watermark requires revealing the entire embedded message. This lack of control leads to unnecessary information exposure and raises privacy concerns. We propose Hierarchical Vocabulary Routing (HeRo), a watermarking framework that enables selective disclosure of embedded metadata. The method recursively partitions the vocabulary and distributes watermark information across hierarchical layers, so that different verifiers can decode only the portions of the payload corresponding to their access level. We show that the proposed scheme preserves the unbiasedness of the underlying sampling process and thus maintains text quality. Experiments demonstrate that our framework supports fine-grained access control while achieving high detection accuracy and low latency. Code is available at https://github.com/xuyangc03/hero-watermark.
Reasoning segmentation aims to predict pixel-wise masks for targets given complex language queries. Existing approaches leverage Multimodal Large Language Models (MLLMs) for vision-language reasoning and generate intermediate target cues (e.g., points or boxes) to guide a segmentation model. However, compressing rich reasoning into sparse cues often introduces ambiguity and noise, preventing these cues from accurately preserving the reasoning intent. While multiple complementary cues can enrich target information, existing methods typically feed them jointly into a single segmentation process, allowing ambiguous or erroneous cues to affect the entire prediction. Therefore, we propose DGSeg, a reasoning segmentation framework that learns to fuse predictions guided by semantic and spatial cues. Specifically, the MLLM jointly reasons about both target identity and spatial location, producing complementary semantic and spatial cues that are fed into separate segmentation branches. Their predictions are adaptively integrated by a lightweight dynamic gating module trained with relative branch-quality supervision to suppress noisy or conflicting regions. Extensive experiments demonstrate that DGSeg consistently outperforms strong baselines on multiple benchmarks and achieves 69.6% and 67.3% gIoU on the challenging ReasonSeg validation and test splits. Code is available at https://github.com/RZZeng/DGSeg.
展开完整摘要收起摘要↓
Reasoning segmentation aims to predict pixel-wise masks for targets given complex language queries. Existing approaches leverage Multimodal Large Language Models (MLLMs) for vision-language reasoning and generate intermediate target cues (e.g., points or boxes) to guide a segmentation model. However, compressing rich reasoning into sparse cues often introduces ambiguity and noise, preventing these cues from accurately preserving the reasoning intent. While multiple complementary cues can enrich target information, existing methods typically feed them jointly into a single segmentation process, allowing ambiguous or erroneous cues to affect the entire prediction. Therefore, we propose DGSeg, a reasoning segmentation framework that learns to fuse predictions guided by semantic and spatial cues. Specifically, the MLLM jointly reasons about both target identity and spatial location, producing complementary semantic and spatial cues that are fed into separate segmentation branches. Their predictions are adaptively integrated by a lightweight dynamic gating module trained with relative branch-quality supervision to suppress noisy or conflicting regions. Extensive experiments demonstrate that DGSeg consistently outperforms strong baselines on multiple benchmarks and achieves 69.6% and 67.3% gIoU on the challenging ReasonSeg validation and test splits. Code is available at https://github.com/RZZeng/DGSeg.
作者Yurui Dong, Shu Zou, Siqi Li, Nianchen Deng, Hongbin Zhou, Xuemeng Yang, Pinlong Cai, Licheng Wen, Xinyu Cai, Botian Shi
Recent advances in large language models and programmatic CAD have significantly improved Text-to-CAD generation for individual parts. However, production-ready mechanical assembly generation remains largely unsolved. Unlike single-part modeling, assemblies require coordinated reasoning over multiple components, functional interfaces, assembly relations, engineering principles, and physical consistency. Consequently, directly generating executable CAD code is insufficient for constructing mechanically valid and reusable assemblies. We present AssemCAD, an axiom-grounded framework for production-ready CAD assembly generation from natural language. Instead of representing an assembly as monolithic CAD code, AssemCAD first constructs an axiomatic Assembly Specification consisting of typed parts, geometry-backed ports, executable mates, and engineering axioms. Each assembly relation is explicitly grounded in one or more engineering principles, making the resulting specification interpretable, reusable, and verifiable. To realize this specification, AssemCAD introduces a port- and mate-based CAD assembly library that executes symbolic assembly relations through deterministic mate transformations and validates declared interfaces using concrete B-Rep geometric evidence. Built on this representation and library, AssemCAD further supports on-demand synthesis of reusable parametric component factories for both standard and open-world geometries. Experiments on AssemBench show that AssemCAD substantially improves assembly preservation and physical validity over code-centric CAD generation baselines, while generalizing across different foundation-model backbones. By combining axiom-grounded assembly reasoning with deterministic geometric execution, AssemCAD extends Text-to-CAD from isolated part generation toward production-ready mechanical assembly design.
展开完整摘要收起摘要↓
Recent advances in large language models and programmatic CAD have significantly improved Text-to-CAD generation for individual parts. However, production-ready mechanical assembly generation remains largely unsolved. Unlike single-part modeling, assemblies require coordinated reasoning over multiple components, functional interfaces, assembly relations, engineering principles, and physical consistency. Consequently, directly generating executable CAD code is insufficient for constructing mechanically valid and reusable assemblies. We present AssemCAD, an axiom-grounded framework for production-ready CAD assembly generation from natural language. Instead of representing an assembly as monolithic CAD code, AssemCAD first constructs an axiomatic Assembly Specification consisting of typed parts, geometry-backed ports, executable mates, and engineering axioms. Each assembly relation is explicitly grounded in one or more engineering principles, making the resulting specification interpretable, reusable, and verifiable. To realize this specification, AssemCAD introduces a port- and mate-based CAD assembly library that executes symbolic assembly relations through deterministic mate transformations and validates declared interfaces using concrete B-Rep geometric evidence. Built on this representation and library, AssemCAD further supports on-demand synthesis of reusable parametric component factories for both standard and open-world geometries. Experiments on AssemBench show that AssemCAD substantially improves assembly preservation and physical validity over code-centric CAD generation baselines, while generalizing across different foundation-model backbones. By combining axiom-grounded assembly reasoning with deterministic geometric execution, AssemCAD extends Text-to-CAD from isolated part generation toward production-ready mechanical assembly design.
Physics reasoning fails structurally in small language models: an error at any step propagates forward, corrupting every inference that follows. Limited domain knowledge, hallucination under multi-step derivation, and distributional sensitivity compound this failure. We propose a step-level reward framework that identifies the first reasoning error, generates targeted structured feedback, and trains the model to revise its solution via policy gradient with KL regularization, without exposing it to ground truth solutions as generation targets. Unlike annotation-dependent step-level methods, no preference data construction is required and the external verifier operates exclusively at training time. Across five physics benchmarks, our framework delivers accuracy gains of 17-20% over CoT prompting and 10-16% over the strongest baseline, reduces calculation errors from 56.9% to 23.5%, and reduces miscomprehension errors from 22.3% to 12.0% in the best observed cases. Conceptual errors reduce from 89.7% to 68.7%, yet persist as the hardest failure mode across all conditions.
展开完整摘要收起摘要↓
Physics reasoning fails structurally in small language models: an error at any step propagates forward, corrupting every inference that follows. Limited domain knowledge, hallucination under multi-step derivation, and distributional sensitivity compound this failure. We propose a step-level reward framework that identifies the first reasoning error, generates targeted structured feedback, and trains the model to revise its solution via policy gradient with KL regularization, without exposing it to ground truth solutions as generation targets. Unlike annotation-dependent step-level methods, no preference data construction is required and the external verifier operates exclusively at training time. Across five physics benchmarks, our framework delivers accuracy gains of 17-20% over CoT prompting and 10-16% over the strongest baseline, reduces calculation errors from 56.9% to 23.5%, and reduces miscomprehension errors from 22.3% to 12.0% in the best observed cases. Conceptual errors reduce from 89.7% to 68.7%, yet persist as the hardest failure mode across all conditions.
Large language model (LLM) training shall adapt to dynamic resources in shared clusters to tackle the elasticity, including passive preemption and optimistic scaling. State migration across device sets is required when altering the hybrid-parallel configuration due to dynamic resources. Existing solutions rely on checkpoint-based mechanisms, which persist complete states to storage for resuming with re-assigned resources, forcing all GPUs to stall when transferring model states. Despite optimization efforts, checkpoint-based solutions incur prohibitive latency due to data movement across memory hierarchies. We propose ETC, a checkpoint-free state migration framework for elastic hybrid-parallel LLM training. We exploits the state locality to minimize inter-GPU data movement, replacing storage persistence with direct peer-to-peer communication. Besides, we eliminate node fragmentation through communication coalescing. Integrated with Megatron-LM, ETC reduces migration overhead by 2.33$\times$ to 6.37$\times$ compared to checkpoint-based solutions across diverse parallel configurations. By enabling efficient migration, ETC unlocks practical elastic training in production environments.
展开完整摘要收起摘要↓
Large language model (LLM) training shall adapt to dynamic resources in shared clusters to tackle the elasticity, including passive preemption and optimistic scaling. State migration across device sets is required when altering the hybrid-parallel configuration due to dynamic resources. Existing solutions rely on checkpoint-based mechanisms, which persist complete states to storage for resuming with re-assigned resources, forcing all GPUs to stall when transferring model states. Despite optimization efforts, checkpoint-based solutions incur prohibitive latency due to data movement across memory hierarchies. We propose ETC, a checkpoint-free state migration framework for elastic hybrid-parallel LLM training. We exploits the state locality to minimize inter-GPU data movement, replacing storage persistence with direct peer-to-peer communication. Besides, we eliminate node fragmentation through communication coalescing. Integrated with Megatron-LM, ETC reduces migration overhead by 2.33$\times$ to 6.37$\times$ compared to checkpoint-based solutions across diverse parallel configurations. By enabling efficient migration, ETC unlocks practical elastic training in production environments.
LLMs enable the integration of external conceptual resources at inference time, creating new opportunities for detecting ideologically and historically complex phenomena such as antisemitism. We investigate how different forms of conceptual grounding affect antisemitism detection and explanation behavior across four state-of-the-art LLMs. Using two expert-annotated datasets, we compare definitional, fine-grained taxonomic, example-augmented, and large-context representations of antisemitism. We find that fine-grained taxonomic representations substantially improve recall, while simultaneously reducing precision. Surprisingly, supplying substantially larger conceptual resources yields no additional quantitative benefit. Post-Holocaust antisemitism poses the most persistent challenge across models and configurations. Analysis of explanations further reveals systematic limitations including overproduction of conceptual references, reliance on lexical cues, overconfidence, and difficulties with subtle or justificatory forms of antisemitism. Our findings highlight both the potential and the remaining limitations of conceptually grounded LLMs for antisemitism detection and reasoning.
展开完整摘要收起摘要↓
LLMs enable the integration of external conceptual resources at inference time, creating new opportunities for detecting ideologically and historically complex phenomena such as antisemitism. We investigate how different forms of conceptual grounding affect antisemitism detection and explanation behavior across four state-of-the-art LLMs. Using two expert-annotated datasets, we compare definitional, fine-grained taxonomic, example-augmented, and large-context representations of antisemitism. We find that fine-grained taxonomic representations substantially improve recall, while simultaneously reducing precision. Surprisingly, supplying substantially larger conceptual resources yields no additional quantitative benefit. Post-Holocaust antisemitism poses the most persistent challenge across models and configurations. Analysis of explanations further reveals systematic limitations including overproduction of conceptual references, reliance on lexical cues, overconfidence, and difficulties with subtle or justificatory forms of antisemitism. Our findings highlight both the potential and the remaining limitations of conceptually grounded LLMs for antisemitism detection and reasoning.
作者Karina Halevy, Julia Mendelsohn, Chan Young Park, Yulia Tsvetkov, Maarten Sap
Addressing hate and violence in society requires timely detection of hateful events from public reporting, but automated identification of hateful events remains underexplored. We introduce the task of hateful event detection and investigate the ability of AI systems, specifically large language models (LLMs), to discover and classify reports of antisemitic events with fine-grained labels. We evaluate OpenAI's GPT-4o and Meta's Llama-3.2-3B-Instruct on multiple expert-annotated datasets containing antisemitic event descriptions from news articles, civil society reports, and official records. We show that LLMs, particularly GPT-4o, have potential for this task, but substantial improvement is needed. Providing clear term definitions and in-context examples in prompts can improve performance: definitions are most helpful for rhetoric-oriented events (e.g. classical antisemitic tropes), while examples help label action-oriented events (e.g. physical assault). A case study of college newspapers demonstrates that LLMs can help surface relevant real-world events, supporting early monitoring and intervention. Overall, our findings highlight both opportunities and critical gaps in AI's ability to recognize complex harms and underscore the need for collaborative efforts among AI developers, policymakers, and civil society to design models, implement robust evaluation, and develop policy frameworks for defining and combating hate efficiently and effectively.
展开完整摘要收起摘要↓
Addressing hate and violence in society requires timely detection of hateful events from public reporting, but automated identification of hateful events remains underexplored. We introduce the task of hateful event detection and investigate the ability of AI systems, specifically large language models (LLMs), to discover and classify reports of antisemitic events with fine-grained labels. We evaluate OpenAI's GPT-4o and Meta's Llama-3.2-3B-Instruct on multiple expert-annotated datasets containing antisemitic event descriptions from news articles, civil society reports, and official records. We show that LLMs, particularly GPT-4o, have potential for this task, but substantial improvement is needed. Providing clear term definitions and in-context examples in prompts can improve performance: definitions are most helpful for rhetoric-oriented events (e.g. classical antisemitic tropes), while examples help label action-oriented events (e.g. physical assault). A case study of college newspapers demonstrates that LLMs can help surface relevant real-world events, supporting early monitoring and intervention. Overall, our findings highlight both opportunities and critical gaps in AI's ability to recognize complex harms and underscore the need for collaborative efforts among AI developers, policymakers, and civil society to design models, implement robust evaluation, and develop policy frameworks for defining and combating hate efficiently and effectively.