Optimization-based latent reasoning improves large language model outputs by optimizing instance-specific continuous states at test time while keeping model parameters frozen. Existing methods, however, typically connect these states to the reasoning trajectory through decoded tokens, making sequence-level credit assignment indirect and obscuring how latent updates shape subsequent reasoning. We introduce GradCuit (gradient through circuit), which inserts optimizable latent states at a selected Transformer layer between the hidden representations of the prompt and the generated continuation. Causal self-attention provides every continuation-token log-probability with a differentiable path to every preceding latent state through the remaining Transformer blocks, enabling reward-weighted gradients from the entire continuation to be assigned directly to the latents. Across five instruction-tuned backbones, three reasoning benchmarks, and two answer formats, GradCuit achieves an average accuracy of 64.5%, outperforming chain-of-thought prompting by 6.6 percentage points and the strongest competing method by 2.4 points. GradCuit also demonstrates greater robustness: across seven learning-rate settings, it consistently outperforms LatentSeek while reducing the standard deviation of accuracy from 1.53 to 0.82, and even its random-walk variant remains competitive with LatentSeek. For interpretability, token-level gradient attribution reveals that latent influence concentrates on reasoning-connector tokens, while layer analysis identifies early-to-middle Transformer layers as the most effective optimization space. By directly optimizing internal reasoning from outcome feedback, GradCuit opens a new axis of robust and interpretable test-time scaling, where LLMs adapt how they reason rather than merely regenerate, sample, or rerank outputs.
展开完整摘要收起摘要↓
Optimization-based latent reasoning improves large language model outputs by optimizing instance-specific continuous states at test time while keeping model parameters frozen. Existing methods, however, typically connect these states to the reasoning trajectory through decoded tokens, making sequence-level credit assignment indirect and obscuring how latent updates shape subsequent reasoning. We introduce GradCuit (gradient through circuit), which inserts optimizable latent states at a selected Transformer layer between the hidden representations of the prompt and the generated continuation. Causal self-attention provides every continuation-token log-probability with a differentiable path to every preceding latent state through the remaining Transformer blocks, enabling reward-weighted gradients from the entire continuation to be assigned directly to the latents. Across five instruction-tuned backbones, three reasoning benchmarks, and two answer formats, GradCuit achieves an average accuracy of 64.5%, outperforming chain-of-thought prompting by 6.6 percentage points and the strongest competing method by 2.4 points. GradCuit also demonstrates greater robustness: across seven learning-rate settings, it consistently outperforms LatentSeek while reducing the standard deviation of accuracy from 1.53 to 0.82, and even its random-walk variant remains competitive with LatentSeek. For interpretability, token-level gradient attribution reveals that latent influence concentrates on reasoning-connector tokens, while layer analysis identifies early-to-middle Transformer layers as the most effective optimization space. By directly optimizing internal reasoning from outcome feedback, GradCuit opens a new axis of robust and interpretable test-time scaling, where LLMs adapt how they reason rather than merely regenerate, sample, or rerank outputs.
Large language models often show users a final response and a short reasoning summary while the full reasoning trace stays hidden. We introduce an observability ladder that holds each completed run fixed and varies only what a reader inspects to judge whether the answer is correct: the response, a self-summary the model writes from the trace, the trace itself, and internal signals, each with and without the prompt. Across three benchmarks and five open-weight Qwen3 and gpt-oss models, we train matched linear correctness predictors on each access level. Without the prompt, summaries carry most of the trace's ranking signal (mean AUROC 0.774 versus 0.813) and add +0.156 over the response alone. With the prompt visible, the summary's gain collapses to +0.019, while the trace still adds +0.041. Even at equal length, the trace's last words predict correctness as well as summaries, or slightly better, and carry denser and more discriminative uncertainty and self-correction cues. On MMLU-Pro questions with both correct and incorrect runs, linear summary readers are near chance and trace readers retain only modest signal, both with and without the prompt (prompt-withheld AUROC 0.503-0.545 versus 0.544-0.590). With the prompt withheld, a GPT-5-mini reader recovers substantially more signal from both summaries and traces on gpt-oss-20b, and even then the trace keeps a small +0.034 advantage. Much of the linear readers' trace signal is associated with length. In the common case where users already hold the prompt, summaries are less helpful than the full trace for monitoring correctness. Monitorability is thus a joint property of the display and the reader, so any monitorability claim, including for faithfulness, should specify both.
展开完整摘要收起摘要↓
Large language models often show users a final response and a short reasoning summary while the full reasoning trace stays hidden. We introduce an observability ladder that holds each completed run fixed and varies only what a reader inspects to judge whether the answer is correct: the response, a self-summary the model writes from the trace, the trace itself, and internal signals, each with and without the prompt. Across three benchmarks and five open-weight Qwen3 and gpt-oss models, we train matched linear correctness predictors on each access level. Without the prompt, summaries carry most of the trace's ranking signal (mean AUROC 0.774 versus 0.813) and add +0.156 over the response alone. With the prompt visible, the summary's gain collapses to +0.019, while the trace still adds +0.041. Even at equal length, the trace's last words predict correctness as well as summaries, or slightly better, and carry denser and more discriminative uncertainty and self-correction cues. On MMLU-Pro questions with both correct and incorrect runs, linear summary readers are near chance and trace readers retain only modest signal, both with and without the prompt (prompt-withheld AUROC 0.503-0.545 versus 0.544-0.590). With the prompt withheld, a GPT-5-mini reader recovers substantially more signal from both summaries and traces on gpt-oss-20b, and even then the trace keeps a small +0.034 advantage. Much of the linear readers' trace signal is associated with length. In the common case where users already hold the prompt, summaries are less helpful than the full trace for monitoring correctness. Monitorability is thus a joint property of the display and the reader, so any monitorability claim, including for faithfulness, should specify both.
Understanding how computational effort is allocated across individual chain-of-thought (CoT) reasoning steps remains an open challenge: existing interpretability methods rely on output-level signals or collapse processing depth into a single trajectory-level scalar, leaving step-wise effort opaque. We propose Step-Aware Reasoning Energy (SARE), a geometric framework that quantifies effort at the granularity of individual CoT steps via Centered Kernel Alignment (CKA) between Gram matrices of token hidden states across adjacent transformer layers, capturing inter-token relational structure without requiring eigenvector alignment or cluster correspondence. SARE further contextualizes this energy within reasoning's semantic progression by modeling CoT trajectories as transitions among latent semantic states. Across six reasoning benchmarks and three open-weight LLMs, we find that reasoning energy is highly non-uniform across step types, exhibiting phase-like transitions invisible to trajectory-level metrics; incorrect trajectories show systematically lower energy at critical reasoning junctions; and SARE-based features match or outperform output-based confidence baselines in most settings, indicating that internal geometric dynamics encode predictive information beyond surface-level signals.
展开完整摘要收起摘要↓
Understanding how computational effort is allocated across individual chain-of-thought (CoT) reasoning steps remains an open challenge: existing interpretability methods rely on output-level signals or collapse processing depth into a single trajectory-level scalar, leaving step-wise effort opaque. We propose Step-Aware Reasoning Energy (SARE), a geometric framework that quantifies effort at the granularity of individual CoT steps via Centered Kernel Alignment (CKA) between Gram matrices of token hidden states across adjacent transformer layers, capturing inter-token relational structure without requiring eigenvector alignment or cluster correspondence. SARE further contextualizes this energy within reasoning's semantic progression by modeling CoT trajectories as transitions among latent semantic states. Across six reasoning benchmarks and three open-weight LLMs, we find that reasoning energy is highly non-uniform across step types, exhibiting phase-like transitions invisible to trajectory-level metrics; incorrect trajectories show systematically lower energy at critical reasoning junctions; and SARE-based features match or outperform output-based confidence baselines in most settings, indicating that internal geometric dynamics encode predictive information beyond surface-level signals.
作者Yimi Wang, Hao Li, Shuo Yang, He Cao, Dechen Zhang, Ziang Wu, Zhiyuan Yan, Fanyang Mo, Li Yuan
While large language models excel in reasoning, these generalists often lack knowledge for specialized scientific domains. Conversely, domain models~(specialists), while knowledgeable, suffer from specialization side-effects including diminished logic and reduced robustness.To address this dilemma, we introduce Divergence Decoding, a training-free framework for capability fusion. It reconstructs the "draft-and-verify" skeleton of speculative decoding into an adaptive routing mechanism. The core is using Jensen-Shannon divergence to monitor the distributional disagreement between the two models at each token. When the specialist exhibits significant divergence, our method identifies it as a potential reasoning risk and instantaneously routes control to the generalist. This allows the dynamic injection of general reasoning while preserving domain expertise, achieving inference-time policy composition of the generalist and the specialist.We evaluate Divergence Decoding across diverse model families (Qwen and Llama series) on challenging scientific benchmarks (GPQA, ChemBench, and ChemCoTBench). Experimental results demonstrate that Divergence Decoding outperforms both the domain-specialized and general-purpose models, effectively surpassing the performance of most single-model baseline. This suggests that Divergence Decoding provides a general, training-free paradigm for fusing diverse LLM capabilities through adaptive inference-time collaboration.
展开完整摘要收起摘要↓
While large language models excel in reasoning, these generalists often lack knowledge for specialized scientific domains. Conversely, domain models~(specialists), while knowledgeable, suffer from specialization side-effects including diminished logic and reduced robustness.To address this dilemma, we introduce Divergence Decoding, a training-free framework for capability fusion. It reconstructs the "draft-and-verify" skeleton of speculative decoding into an adaptive routing mechanism. The core is using Jensen-Shannon divergence to monitor the distributional disagreement between the two models at each token. When the specialist exhibits significant divergence, our method identifies it as a potential reasoning risk and instantaneously routes control to the generalist. This allows the dynamic injection of general reasoning while preserving domain expertise, achieving inference-time policy composition of the generalist and the specialist.We evaluate Divergence Decoding across diverse model families (Qwen and Llama series) on challenging scientific benchmarks (GPQA, ChemBench, and ChemCoTBench). Experimental results demonstrate that Divergence Decoding outperforms both the domain-specialized and general-purpose models, effectively surpassing the performance of most single-model baseline. This suggests that Divergence Decoding provides a general, training-free paradigm for fusing diverse LLM capabilities through adaptive inference-time collaboration.
Action recognition in complex scenes often involves multiple concurrent fine-grained actions, making it challenging to model internal action structures. Most existing methods rely on holistic representations, which are insufficient for capturing subtle interactions and fine-grained semantics. While recent prompt-based approaches introduce disentanglement, they lack explicit semantic guidance, and methods based solely on visual or structured cues remain coarse-grained. In this paper, we propose Knowledge-guided Disentanglement with Atomic Actions (KDA), which leverages fine-grained semantic knowledge to enhance action representations and enable more precise disentanglement. Specifically, we use Large Language Models (LLMs) to decompose action labels into atomic actions, providing explicit spatial-temporal semantics. A Knowledge Injection Module (KIM) first integrates atomic action knowledge into video features. Based on this enhanced representation, a Knowledge Disentanglement Module (KDM) further disentangles atomic action knowledge to produce more precise semantic guidance for action disentanglement. A Knowledge Disentanglement Loss (KD Loss) is introduced to encourage clearer disentanglement of knowledge components within KDM. Extensive experiments demonstrate that KDA improves feature discriminability and achieves state-of-the-art performance on multi-label action recognition benchmarks. Moreover, KIM and KDM can be readily integrated into other methods, demonstrating strong generality.
展开完整摘要收起摘要↓
Action recognition in complex scenes often involves multiple concurrent fine-grained actions, making it challenging to model internal action structures. Most existing methods rely on holistic representations, which are insufficient for capturing subtle interactions and fine-grained semantics. While recent prompt-based approaches introduce disentanglement, they lack explicit semantic guidance, and methods based solely on visual or structured cues remain coarse-grained. In this paper, we propose Knowledge-guided Disentanglement with Atomic Actions (KDA), which leverages fine-grained semantic knowledge to enhance action representations and enable more precise disentanglement. Specifically, we use Large Language Models (LLMs) to decompose action labels into atomic actions, providing explicit spatial-temporal semantics. A Knowledge Injection Module (KIM) first integrates atomic action knowledge into video features. Based on this enhanced representation, a Knowledge Disentanglement Module (KDM) further disentangles atomic action knowledge to produce more precise semantic guidance for action disentanglement. A Knowledge Disentanglement Loss (KD Loss) is introduced to encourage clearer disentanglement of knowledge components within KDM. Extensive experiments demonstrate that KDA improves feature discriminability and achieves state-of-the-art performance on multi-label action recognition benchmarks. Moreover, KIM and KDM can be readily integrated into other methods, demonstrating strong generality.
作者Marie Neubrander, Graham Tierney, Alexander Volfovsky
Estimating causal effects of linguistic properties from observational text is difficult because the same document can contain both the treatment of interest and the non-treatment textual attributes needed for adjustment. Existing approaches often learn representations from the full text to capture latent confounding, but when treatment status is itself encoded by words in the text, these representations can directly encode treatment. This creates a confounder trap: richer representations can make treated and control documents separable, inducing overlap violations even when the underlying causal problem satisfies overlap. We study latent text treatments that are encoded through lexicons or other treatment-defining lexical information, and propose masking-based adjustment representations that remove this lexical treatment signal before representation learning. We formalize representation-induced overlap failure, prove that deletion masking preserves overlap for bag-of-words/topic-model representations, and characterize replacement masking as a natural relaxation for large language models that hides treatment-defining tokens while preserving word order and context. Across simulations, masking improves overlap diagnostics, stabilizes treatment effect estimates, and reduces bias relative to adjustment methods that learn from the unmasked text.
展开完整摘要收起摘要↓
Estimating causal effects of linguistic properties from observational text is difficult because the same document can contain both the treatment of interest and the non-treatment textual attributes needed for adjustment. Existing approaches often learn representations from the full text to capture latent confounding, but when treatment status is itself encoded by words in the text, these representations can directly encode treatment. This creates a confounder trap: richer representations can make treated and control documents separable, inducing overlap violations even when the underlying causal problem satisfies overlap. We study latent text treatments that are encoded through lexicons or other treatment-defining lexical information, and propose masking-based adjustment representations that remove this lexical treatment signal before representation learning. We formalize representation-induced overlap failure, prove that deletion masking preserves overlap for bag-of-words/topic-model representations, and characterize replacement masking as a natural relaxation for large language models that hides treatment-defining tokens while preserving word order and context. Across simulations, masking improves overlap diagnostics, stabilizes treatment effect estimates, and reduces bias relative to adjustment methods that learn from the unmasked text.
Large language models (LLMs) are increasingly used as general-purpose translation systems, but their behavior is usually evaluated under a single prompt shape: translate one source sentence into one target language. In practice, users may ask for one target language, for several related languages at once, or for translations conditioned on examples. This paper studies prompt scope and demonstration selection as experimental variables for local LLM machine translation. We evaluate English-to-Romance and English-to-Germanic translation on the full FLORES devtest split for nine official European Union languages. We compare three local instruction-tuned LLMs, llama3.2:3b, mistral:latest, and qwen2.5:14b, against dedicated MT baselines from OPUS-MT and NLLB-200. We test zero-shot prompting and k=5 few-shot prompting with random, lexical-similarity, and embedding-similarity demonstration selection. We also compare single-target prompts with JSON-formatted family-scope prompts that request all languages in a family at once. Results show that dedicated MT systems remain strongest overall, especially for Germanic languages. Few-shot prompting helps mistral:latest and qwen2.5:14b, but hurts llama3.2:3b; embedding retrieval is best on average for the stronger LLMs, but its advantage over random and lexical examples is modest. Family-scope prompting is feasible for stronger local LLMs but exposes structured-output failures in smaller models. These findings motivate evaluating LLM translation not only by language pair and metric, but also by prompt scope, retrieval strategy, and multi-target compliance.
展开完整摘要收起摘要↓
Large language models (LLMs) are increasingly used as general-purpose translation systems, but their behavior is usually evaluated under a single prompt shape: translate one source sentence into one target language. In practice, users may ask for one target language, for several related languages at once, or for translations conditioned on examples. This paper studies prompt scope and demonstration selection as experimental variables for local LLM machine translation. We evaluate English-to-Romance and English-to-Germanic translation on the full FLORES devtest split for nine official European Union languages. We compare three local instruction-tuned LLMs, llama3.2:3b, mistral:latest, and qwen2.5:14b, against dedicated MT baselines from OPUS-MT and NLLB-200. We test zero-shot prompting and k=5 few-shot prompting with random, lexical-similarity, and embedding-similarity demonstration selection. We also compare single-target prompts with JSON-formatted family-scope prompts that request all languages in a family at once. Results show that dedicated MT systems remain strongest overall, especially for Germanic languages. Few-shot prompting helps mistral:latest and qwen2.5:14b, but hurts llama3.2:3b; embedding retrieval is best on average for the stronger LLMs, but its advantage over random and lexical examples is modest. Family-scope prompting is feasible for stronger local LLMs but exposes structured-output failures in smaller models. These findings motivate evaluating LLM translation not only by language pair and metric, but also by prompt scope, retrieval strategy, and multi-target compliance.
Public tests are widely used to guide large language model code generation, but whether models treat them as executable specifications or merely as extra prompt context remains unclear. We study test-driven code generation on HumanEval+, MBPP+, and recent LiveCodeBench tasks using Qwen2.5-Coder-7B and Qwen3.6-27B. We compare natural-language-only prompts with relevant visible tests, shuffled outputs, irrelevant tests, assertion-only tests, and stronger-model-generated synthetic tests. Evaluation combines hidden or private test pass rates with task-level behavior flips, linear probes, and layer-wise hidden-state shifts. Visible tests substantially improve Qwen2.5 performance on MBPP+ but have little or unstable effect on HumanEval+ and LiveCodeBench. For Qwen3.6 on LiveCodeBench, the natural-language-only pass rate rises from 13.1% to 39.4%, yet relevant original tests add only 2.9 percentage points (p = .458), and synthetic high5 tests add 1.7 points over their matched baseline (p = .701). More tests produce larger representation shifts, but accuracy gains quickly saturate or disappear; assertion-only and irrelevant-test prompts can strongly alter hidden states without improving correctness. After controlling for prompt length, task-level shifts are associated with both beneficial and harmful flips, with a substantially stronger association for harms. These results show that tests influence code models through both semantic guidance and prompt-context perturbation, and that representational change alone does not demonstrate effective test utilization.
展开完整摘要收起摘要↓
Public tests are widely used to guide large language model code generation, but whether models treat them as executable specifications or merely as extra prompt context remains unclear. We study test-driven code generation on HumanEval+, MBPP+, and recent LiveCodeBench tasks using Qwen2.5-Coder-7B and Qwen3.6-27B. We compare natural-language-only prompts with relevant visible tests, shuffled outputs, irrelevant tests, assertion-only tests, and stronger-model-generated synthetic tests. Evaluation combines hidden or private test pass rates with task-level behavior flips, linear probes, and layer-wise hidden-state shifts. Visible tests substantially improve Qwen2.5 performance on MBPP+ but have little or unstable effect on HumanEval+ and LiveCodeBench. For Qwen3.6 on LiveCodeBench, the natural-language-only pass rate rises from 13.1% to 39.4%, yet relevant original tests add only 2.9 percentage points (p = .458), and synthetic high5 tests add 1.7 points over their matched baseline (p = .701). More tests produce larger representation shifts, but accuracy gains quickly saturate or disappear; assertion-only and irrelevant-test prompts can strongly alter hidden states without improving correctness. After controlling for prompt length, task-level shifts are associated with both beneficial and harmful flips, with a substantially stronger association for harms. These results show that tests influence code models through both semantic guidance and prompt-context perturbation, and that representational change alone does not demonstrate effective test utilization.
Can large language models participate in the production and validation of ERTMS/ETCS data without undermining the certification arguments required by CENELEC EN 50128/50716? ERTMS/ETCS is a distributed safety-critical system (trackside, onboard, radio-block centre) whose behaviour is parameterised by large volumes of data drawn from the UNISIG Subsets; errors in that data propagate through the distributed architecture. This paper reports the current status of an ongoing industrial research effort at CLEARSY, ValidAItion, that bridges the ERTMS Operational Simulator to the CLEARSY Data Solver and applies rules expressed in the B mathematical language to that trackside data. During construction, a large language model (Claude) has authored the rule corpus and the parsers through a Model Context Protocol server; every proposal is adjudicated by the downstream toolchain and by systematic human review, and the toolchain has already rejected a syntactically valid but semantically illegal generated scenario. The contribution is architectural and industrial, not algorithmic: the work combines frameworks already in use at CLEARSY (CLEARSY Data Solver, ERTMS Operational Simulator) with a conversational authoring loop, rather than proposing a new formal method. It is a progress report: rule coverage is growing, the human-review campaign is underway, and the quantitative results will be published separately. The paper argues, on the evidence gathered so far, that formal rules in the mathematical language of B must remain the source of truth, while the language model serves as the fenced assistant in a distributed safety-critical railway system: AI proposes, the formal oracle disposes, the human confirms.
展开完整摘要收起摘要↓
Can large language models participate in the production and validation of ERTMS/ETCS data without undermining the certification arguments required by CENELEC EN 50128/50716? ERTMS/ETCS is a distributed safety-critical system (trackside, onboard, radio-block centre) whose behaviour is parameterised by large volumes of data drawn from the UNISIG Subsets; errors in that data propagate through the distributed architecture. This paper reports the current status of an ongoing industrial research effort at CLEARSY, ValidAItion, that bridges the ERTMS Operational Simulator to the CLEARSY Data Solver and applies rules expressed in the B mathematical language to that trackside data. During construction, a large language model (Claude) has authored the rule corpus and the parsers through a Model Context Protocol server; every proposal is adjudicated by the downstream toolchain and by systematic human review, and the toolchain has already rejected a syntactically valid but semantically illegal generated scenario. The contribution is architectural and industrial, not algorithmic: the work combines frameworks already in use at CLEARSY (CLEARSY Data Solver, ERTMS Operational Simulator) with a conversational authoring loop, rather than proposing a new formal method. It is a progress report: rule coverage is growing, the human-review campaign is underway, and the quantitative results will be published separately. The paper argues, on the evidence gathered so far, that formal rules in the mathematical language of B must remain the source of truth, while the language model serves as the fenced assistant in a distributed safety-critical railway system: AI proposes, the formal oracle disposes, the human confirms.
作者Xin Xin, Jincheng Lou, Junhui Li, Jinglin Yan, Panda Xiao, Di Wu, Haixiao Li, Weicong Lu, Weijian Fan, Xinyu Qu, Yuxiang Zhao, Min Yu, Zhixiong Di, Yibo Lin
Functional verification dominates integrated circuit (IC) front-end engineering effort, and a single missed bug that escapes to silicon can trigger a costly respin. Recent large language models (LLMs) offer new opportunities to automate this process, yet existing LLM-based approaches generate each component through independent single-turn calls with no shared context, leaving interface mismatches undetected and reported coverage disconnected from specification requirements. To address these challenges, we present GoGoTB, an agentic framework that achieves end-to-end verification closure through three subsystems: an agentic execution control layer, an evolvable knowledge system, and specification-grounded coverage closure. The execution control layer separates deterministic enforcement from LLM reasoning at every tool and stage boundary. The knowledge system dispatches methodology and design-specific expertise on demand. The coverage framework anchors every bin to a named specification behavior so that each residual gap has a diagnosable root cause and a targeted remedy. Tested on 8 register transfer level (RTL) designs without any human intervention, GoGoTB achieves 100% environment generation success and averages 98.4% line, 97.2% branch, 97.0% toggle, and 83.2% functional coverage. No prior work successfully generates a complete verification environment or achieves meaningful coverage on the same benchmarks.
展开完整摘要收起摘要↓
Functional verification dominates integrated circuit (IC) front-end engineering effort, and a single missed bug that escapes to silicon can trigger a costly respin. Recent large language models (LLMs) offer new opportunities to automate this process, yet existing LLM-based approaches generate each component through independent single-turn calls with no shared context, leaving interface mismatches undetected and reported coverage disconnected from specification requirements. To address these challenges, we present GoGoTB, an agentic framework that achieves end-to-end verification closure through three subsystems: an agentic execution control layer, an evolvable knowledge system, and specification-grounded coverage closure. The execution control layer separates deterministic enforcement from LLM reasoning at every tool and stage boundary. The knowledge system dispatches methodology and design-specific expertise on demand. The coverage framework anchors every bin to a named specification behavior so that each residual gap has a diagnosable root cause and a targeted remedy. Tested on 8 register transfer level (RTL) designs without any human intervention, GoGoTB achieves 100% environment generation success and averages 98.4% line, 97.2% branch, 97.0% toggle, and 83.2% functional coverage. No prior work successfully generates a complete verification environment or achieves meaningful coverage on the same benchmarks.
作者Vivek Shukla, Varun Shukla, Atul, Divya Mishra, Mehul Kumar Das
Mathematical chain of thought (CoT) evaluation is commonly reduced to whether the final answer matches a reference. This conflates producing a correct conclusion with producing a valid derivation an invalid chain can accidentally reach the right answer, while a valid calculation can be followed by a transcription error. We call this mismatch the reasoning answer consistency gap. This framework paper introduces the Reasoning Answer Faithfulness Score (RAFS), a reference free, instance level diagnostic of whether an emitted mathematical trace is locally credible, supports its answer, and is stable under resampling and targeted counterfactual interventions. RAFS combines step validity, reasoning to answer entailment and counterfactual sensitivity, answer consensus, and conditional reasoning stability. It evaluates transcript level agreement, not a models private computation and not factual correctness outside the tested mathematical setting. We retain a preregistered, results blind confirmatory study on GSM8K and MATH, with hypotheses, admissibility rules, calibration, and tests fixed before confirmatory outcomes are inspected. A separate feasibility pilot is specified to verify end to end execution and estimate interven tion coverage before that freeze numerical pilot claims are re ported only when trace level artifacts are available. We formalize four reasoning answer outcomes, justify the non compensatory aggregator, instantiate semantic trace distance, quantify compute and abstention tradeoffs, and define verifier independence and power analyses. RAFS is intended to complement mathematical answer accuracy with an auditable warning signal for silent reasoning failures and answer extraction errors
展开完整摘要收起摘要↓
Mathematical chain of thought (CoT) evaluation is commonly reduced to whether the final answer matches a reference. This conflates producing a correct conclusion with producing a valid derivation an invalid chain can accidentally reach the right answer, while a valid calculation can be followed by a transcription error. We call this mismatch the reasoning answer consistency gap. This framework paper introduces the Reasoning Answer Faithfulness Score (RAFS), a reference free, instance level diagnostic of whether an emitted mathematical trace is locally credible, supports its answer, and is stable under resampling and targeted counterfactual interventions. RAFS combines step validity, reasoning to answer entailment and counterfactual sensitivity, answer consensus, and conditional reasoning stability. It evaluates transcript level agreement, not a models private computation and not factual correctness outside the tested mathematical setting. We retain a preregistered, results blind confirmatory study on GSM8K and MATH, with hypotheses, admissibility rules, calibration, and tests fixed before confirmatory outcomes are inspected. A separate feasibility pilot is specified to verify end to end execution and estimate interven tion coverage before that freeze numerical pilot claims are re ported only when trace level artifacts are available. We formalize four reasoning answer outcomes, justify the non compensatory aggregator, instantiate semantic trace distance, quantify compute and abstention tradeoffs, and define verifier independence and power analyses. RAFS is intended to complement mathematical answer accuracy with an auditable warning signal for silent reasoning failures and answer extraction errors
Adapting Large Language Models (LLMs) to specialized domains often incurs an alignment tax, as fine-tuning on domain-specific tasks can cause catastrophic forgetting and substantially degrade performance on general tasks. We propose MemSFT, which mitigates the alignment tax by decoupling domain specialization from backbone parameter updates through a plug-and-play parametric memory. The memory is trained to imitate the behavior of a non-parametric retriever operating over domain data, thereby memorizing knowledge and patterns that would otherwise be accessed through retrieval. Once trained on a specific domain, the memory can be reused across LLMs of different sizes. During generation, a learned router dynamically fuses the output distributions of the memory and backbone at each decoding step, allowing domain expertise to be invoked selectively. Across biology, geoscience, and law, evaluations with models ranging from Qwen3-8B to Qwen3-235B-A22B show that MemSFT consistently improves domain performance with negligible degradation in general performance, whereas full SFT suffers severe forgetting on general tasks. Overall, our results demonstrate a practical path to decoupling general model capabilities from domain-specific knowledge at the parameter level, thereby equipping LLMs with new specialized capabilities without compromising their general capabilities.
展开完整摘要收起摘要↓
Adapting Large Language Models (LLMs) to specialized domains often incurs an alignment tax, as fine-tuning on domain-specific tasks can cause catastrophic forgetting and substantially degrade performance on general tasks. We propose MemSFT, which mitigates the alignment tax by decoupling domain specialization from backbone parameter updates through a plug-and-play parametric memory. The memory is trained to imitate the behavior of a non-parametric retriever operating over domain data, thereby memorizing knowledge and patterns that would otherwise be accessed through retrieval. Once trained on a specific domain, the memory can be reused across LLMs of different sizes. During generation, a learned router dynamically fuses the output distributions of the memory and backbone at each decoding step, allowing domain expertise to be invoked selectively. Across biology, geoscience, and law, evaluations with models ranging from Qwen3-8B to Qwen3-235B-A22B show that MemSFT consistently improves domain performance with negligible degradation in general performance, whereas full SFT suffers severe forgetting on general tasks. Overall, our results demonstrate a practical path to decoupling general model capabilities from domain-specific knowledge at the parameter level, thereby equipping LLMs with new specialized capabilities without compromising their general capabilities.
作者Shuyue Wei, Chang Liu, Zimu Zhou, Yongxin Tong, Lizhen Cui
Recently, memory management has become a key infrastructure for LLM-based agents, as it directly affects long-horizon reasoning, personalized responses, and knowledge reuse. However, existing LLM memory systems typically adopt a coarse-grained (utility-agnostic) manner that treats heterogeneous user-LLM interaction records uniformly, leading to redundant and low-impact records persisting in the memory repository. To address this challenge, we present MemLens, a value-aware memory management system that takes memory records as first-class data objects. MemLens provides an end-to-end interactive analytics dashboard that exposes the complete memory lifecycle, including Shapley-style memory evaluation, value-aware storage, and memory-assisted response. Through a study-copilot application, the system enables users to inspect memory values, visualize hierarchical memory structures, and compare various memory management strategies in terms of response quality, retrieval latency, and token consumption. Therefore, our MemLens can serve as an efficient, interpretable, and personalized long-term memory management system for LLM-based agents.
展开完整摘要收起摘要↓
Recently, memory management has become a key infrastructure for LLM-based agents, as it directly affects long-horizon reasoning, personalized responses, and knowledge reuse. However, existing LLM memory systems typically adopt a coarse-grained (utility-agnostic) manner that treats heterogeneous user-LLM interaction records uniformly, leading to redundant and low-impact records persisting in the memory repository. To address this challenge, we present MemLens, a value-aware memory management system that takes memory records as first-class data objects. MemLens provides an end-to-end interactive analytics dashboard that exposes the complete memory lifecycle, including Shapley-style memory evaluation, value-aware storage, and memory-assisted response. Through a study-copilot application, the system enables users to inspect memory values, visualize hierarchical memory structures, and compare various memory management strategies in terms of response quality, retrieval latency, and token consumption. Therefore, our MemLens can serve as an efficient, interpretable, and personalized long-term memory management system for LLM-based agents.
Syntactic convergence (the tendency of speakers to adapt in language towards the grammatical profiles of their interlocutors) is a well-documented feature of human dialogue widely considered to operate below conscious awareness. Whether large language models exhibit analogous syntactic convergence toward human users relative to human baselines and across a broad range of syntactic constructions remains an open question. Using substitution-paradigm data in which model generations replace one speaker's turns in pre-existing human dialogues, this study measures turn-adjacent reuse of context-free grammar (CFG) rules across sixteen open-weight Llama and Gemma models (1B-70B, pretrained and instruction-tuned) at 1,901 matched positions per model. Every model showed greater CFG-rule overlap with the preceding human turn than with a sampled unrelated human prime, and in every model this actual-versus-random difference was larger for lower-frequency rules. Each instruction-tuned model also showed greater natural-output overlap with the actual prime than the human response it replaced, and all eight matched architecture pairs exhibited greater actual-prime overlap after instruction tuning. However, relative to pretrained variants, instruction-tuned outputs overlapped more with unrelated primes, showed a smaller actual-versus-random increment, and had lower conditional rule-reuse odds once target rule-set size was held constant. In exploratory analyses, each model exhibited greater mean lexical and semantic similarity to the preceding turn than the matched human responses did. Instruction-tuned models additionally produced responses with greater mean semantic similarity than their pretrained counterparts in all eight architecture pairs, whereas the lexical similarity results were more heterogeneous.
展开完整摘要收起摘要↓
Syntactic convergence (the tendency of speakers to adapt in language towards the grammatical profiles of their interlocutors) is a well-documented feature of human dialogue widely considered to operate below conscious awareness. Whether large language models exhibit analogous syntactic convergence toward human users relative to human baselines and across a broad range of syntactic constructions remains an open question. Using substitution-paradigm data in which model generations replace one speaker's turns in pre-existing human dialogues, this study measures turn-adjacent reuse of context-free grammar (CFG) rules across sixteen open-weight Llama and Gemma models (1B-70B, pretrained and instruction-tuned) at 1,901 matched positions per model. Every model showed greater CFG-rule overlap with the preceding human turn than with a sampled unrelated human prime, and in every model this actual-versus-random difference was larger for lower-frequency rules. Each instruction-tuned model also showed greater natural-output overlap with the actual prime than the human response it replaced, and all eight matched architecture pairs exhibited greater actual-prime overlap after instruction tuning. However, relative to pretrained variants, instruction-tuned outputs overlapped more with unrelated primes, showed a smaller actual-versus-random increment, and had lower conditional rule-reuse odds once target rule-set size was held constant. In exploratory analyses, each model exhibited greater mean lexical and semantic similarity to the preceding turn than the matched human responses did. Instruction-tuned models additionally produced responses with greater mean semantic similarity than their pretrained counterparts in all eight architecture pairs, whereas the lexical similarity results were more heterogeneous.
作者Yuchen Wang, Qihui Zhu, Yang Liu, Xiaoyan Sun, Siying Wu
Recent multimodal large language models (MLLMs), such as Qwen2.5-VL and InternVL3, generate large numbers of vision tokens for high-resolution inputs, leading to substantial computational cost. Existing vision token pruning methods either depend on cross-modal attention and cannot prune before the prefill stage, or rely on diversity estimation with high computational overhead. We observe that attention scores from both vision and text tokens peak at modality separator tokens, suggesting that these separators bridge the two modalities. Based on this observation, we propose SepPrune, an efficient, training-free, plug-and-play pruning method that uses the separator token as a unified query to rank and select informative vision tokens. SepPrune reuses the LLM's built-in projection parameters and requires no architectural changes. Experiments on Qwen2.5-VL-7B show that SepPrune achieves state-of-the-art performance, retaining 96.3% of the original accuracy while removing 80.2% of vision tokens.
展开完整摘要收起摘要↓
Recent multimodal large language models (MLLMs), such as Qwen2.5-VL and InternVL3, generate large numbers of vision tokens for high-resolution inputs, leading to substantial computational cost. Existing vision token pruning methods either depend on cross-modal attention and cannot prune before the prefill stage, or rely on diversity estimation with high computational overhead. We observe that attention scores from both vision and text tokens peak at modality separator tokens, suggesting that these separators bridge the two modalities. Based on this observation, we propose SepPrune, an efficient, training-free, plug-and-play pruning method that uses the separator token as a unified query to rank and select informative vision tokens. SepPrune reuses the LLM's built-in projection parameters and requires no architectural changes. Experiments on Qwen2.5-VL-7B show that SepPrune achieves state-of-the-art performance, retaining 96.3% of the original accuracy while removing 80.2% of vision tokens.
作者Michael Macaulay, Harmony Bouabid, Guo Gen Ang, Sasha Shaw
Capture the Flag (CTF) competitions are among cybersecurity's most effective training grounds, developing practical skill across cryptography, web exploitation, and binary exploitation. Large language models (LLMs) can now solve a growing share of challenges with minimal human input, raising urgent questions about fairness, the validity of rankings, and whether participation still delivers the learning that justifies the effort. This paper reports a mixed-methods study of LLM impact on modern CTFs, combining a synthesis of published benchmarks, including a recent government evaluation, case studies of live competition across three challenge categories, structured observation of the public channels where the community debates AI use, and semi-structured interviews with experienced players and organisers. We map the current human-machine capability boundary by category, showing that easy and intermediate challenges in cryptography, web, and binary exploitation are now reliably automated while narrower sub-categories continue to resist. We find that community disagreement about whether AI should be permitted is downstream of an undeclared prior question: what a competition is for. Against this backdrop we contribute a four-component safeguard framework, combining tiered competition divisions, LLM-resistant challenge design, telemetry used investigatively, and a draft community code of conduct, together with a decision tool that ties the combination of safeguards to a competition's declared purpose. The argument reaches beyond CTFs to any setting in cybersecurity where a demonstrated result is taken as evidence of an underlying ability.
展开完整摘要收起摘要↓
Capture the Flag (CTF) competitions are among cybersecurity's most effective training grounds, developing practical skill across cryptography, web exploitation, and binary exploitation. Large language models (LLMs) can now solve a growing share of challenges with minimal human input, raising urgent questions about fairness, the validity of rankings, and whether participation still delivers the learning that justifies the effort. This paper reports a mixed-methods study of LLM impact on modern CTFs, combining a synthesis of published benchmarks, including a recent government evaluation, case studies of live competition across three challenge categories, structured observation of the public channels where the community debates AI use, and semi-structured interviews with experienced players and organisers. We map the current human-machine capability boundary by category, showing that easy and intermediate challenges in cryptography, web, and binary exploitation are now reliably automated while narrower sub-categories continue to resist. We find that community disagreement about whether AI should be permitted is downstream of an undeclared prior question: what a competition is for. Against this backdrop we contribute a four-component safeguard framework, combining tiered competition divisions, LLM-resistant challenge design, telemetry used investigatively, and a draft community code of conduct, together with a decision tool that ties the combination of safeguards to a competition's declared purpose. The argument reaches beyond CTFs to any setting in cybersecurity where a demonstrated result is taken as evidence of an underlying ability.
Human language exhibits lawful structure at the level of words (frequency, vocabulary growth) and word pairs (co-occurrence across distance). Here we show that the arrangement of words in sequence — a central determinant of meaning — obeys a comparable law. Using large language models as probabilistic probes, we measured the reduction in target perplexity conferred by prior context at distance d beyond that of the same words scrambled; this difference, the contextual persistence function P(d), isolates the influence of arrangement. Across ten corpora spanning six language families and written and spoken modalities, P(d) decayed approximately as 1/d ($P(d) \propto d^{-α}$, mean $α= 1.04$; median $r^2 = 0.96$). The effect vanished in scrambled and synthetic controls, replicated across independent probes, and did not appear in genomic or protein sequences under domain-native models. An exponent near 1 distributes contextual influence approximately uniformly across logarithmic timescales. The results establish a scaling law of contextual persistence in human language.
展开完整摘要收起摘要↓
Human language exhibits lawful structure at the level of words (frequency, vocabulary growth) and word pairs (co-occurrence across distance). Here we show that the arrangement of words in sequence — a central determinant of meaning — obeys a comparable law. Using large language models as probabilistic probes, we measured the reduction in target perplexity conferred by prior context at distance d beyond that of the same words scrambled; this difference, the contextual persistence function P(d), isolates the influence of arrangement. Across ten corpora spanning six language families and written and spoken modalities, P(d) decayed approximately as 1/d ($P(d) \propto d^{-α}$, mean $α= 1.04$; median $r^2 = 0.96$). The effect vanished in scrambled and synthetic controls, replicated across independent probes, and did not appear in genomic or protein sequences under domain-native models. An exponent near 1 distributes contextual influence approximately uniformly across logarithmic timescales. The results establish a scaling law of contextual persistence in human language.
作者Xinran Liu, Shengtao Li, Shouqian Shi, Ge Wang, Xin-Wei Yao
Entity alignment (EA) identifies entities across knowledge graphs (KGs) that refer to the same real-world object. Conventional EA methods mainly exploit explicit graph structures and textual fields, which often provide insufficient semantic understanding to recognize the same entity under heterogeneous descriptions and distinguish it from semantically similar entities. Although large language models (LLMs) offer deeper entity understanding, existing LLM-based EA methods largely use this capability for auxiliary generation or candidate-conditioned decisions. Consequently, such understanding is not distilled into a stable and directly comparable identity space, leaving alignment tied to specific KG pairs or candidate sets and requiring repeated processing as the matching context changes. To address these limitations, we propose IRIS (Identity Representations from Internal States), a training-free framework that constructs for each entity an iris-like signature encoding its distinctive and stable identity characteristics. IRIS derives these signatures by eliciting identity-oriented contextual representations from a frozen LLM, thereby forming a shared space in which each entity is encoded once and can be aligned across different KGs through direct similarity comparison, without pair-dependent representation construction or candidate-wise LLM inference. Across four established EA benchmarks and two frozen LLM backbones, the best IRIS variants achieve Hits@1 scores of 100.00, 99.38, 98.31, and 97.99 on D-Y-15K V2, DBP-WIKI, ICEWS-WIKI, and ICEWS-YAGO, respectively.
展开完整摘要收起摘要↓
Entity alignment (EA) identifies entities across knowledge graphs (KGs) that refer to the same real-world object. Conventional EA methods mainly exploit explicit graph structures and textual fields, which often provide insufficient semantic understanding to recognize the same entity under heterogeneous descriptions and distinguish it from semantically similar entities. Although large language models (LLMs) offer deeper entity understanding, existing LLM-based EA methods largely use this capability for auxiliary generation or candidate-conditioned decisions. Consequently, such understanding is not distilled into a stable and directly comparable identity space, leaving alignment tied to specific KG pairs or candidate sets and requiring repeated processing as the matching context changes. To address these limitations, we propose IRIS (Identity Representations from Internal States), a training-free framework that constructs for each entity an iris-like signature encoding its distinctive and stable identity characteristics. IRIS derives these signatures by eliciting identity-oriented contextual representations from a frozen LLM, thereby forming a shared space in which each entity is encoded once and can be aligned across different KGs through direct similarity comparison, without pair-dependent representation construction or candidate-wise LLM inference. Across four established EA benchmarks and two frozen LLM backbones, the best IRIS variants achieve Hits@1 scores of 100.00, 99.38, 98.31, and 97.99 on D-Y-15K V2, DBP-WIKI, ICEWS-WIKI, and ICEWS-YAGO, respectively.
Activation steering controls model behavior by editing internal activations at inference time. We study its input-side dual: optimizing a fluent prompt so that a chosen internal latent is driven toward zero, with no inference-time model access. Our target is an "evaluation-awareness" latent-linearly readable and steerable in recent work-whose control would threaten the validity of safety evaluations if models behave differently when they detect being tested. Adapting Fluent Dreaming / EPO with a negated feature term (GCG-style token optimization plus a self-cross-entropy fluency regularizer, swept over a fluency weight), we suppress the latent under five target constructions-a CAA direction, a subspace norm, an SAE feature, a single MLP neuron, and a behavioral logit-on Llama-3.2-3B and Llama-3.1-8B. The latent is robustly suppressible ($z\approx-7$), and a causally-validated Llama Scope SAE feature can be fully and selectively turned off. But our controls tell a cautionary story about the CAA direction: a placebo random direction is suppressed just as hard and shifts behavior just as far, and when we hold a real eval passage in context and optimize only a prefix, suppressing the eval-direction fails to reduce-and slightly increases-the model's behavioral eval judgment. Activation-readability, in short, is not behavioral controllability. We further find that a single MLP neuron is eval-correlated but not causal at both scales, and that scanning the real Pile yields a natural-text baseline competitive with the optimizer for the internal direction. A positive control validates our erasure detector, bounding an erasure-vs-rotation question earlier left open.
展开完整摘要收起摘要↓
Activation steering controls model behavior by editing internal activations at inference time. We study its input-side dual: optimizing a fluent prompt so that a chosen internal latent is driven toward zero, with no inference-time model access. Our target is an "evaluation-awareness" latent-linearly readable and steerable in recent work-whose control would threaten the validity of safety evaluations if models behave differently when they detect being tested. Adapting Fluent Dreaming / EPO with a negated feature term (GCG-style token optimization plus a self-cross-entropy fluency regularizer, swept over a fluency weight), we suppress the latent under five target constructions-a CAA direction, a subspace norm, an SAE feature, a single MLP neuron, and a behavioral logit-on Llama-3.2-3B and Llama-3.1-8B. The latent is robustly suppressible ($z\approx-7$), and a causally-validated Llama Scope SAE feature can be fully and selectively turned off. But our controls tell a cautionary story about the CAA direction: a placebo random direction is suppressed just as hard and shifts behavior just as far, and when we hold a real eval passage in context and optimize only a prefix, suppressing the eval-direction fails to reduce-and slightly increases-the model's behavioral eval judgment. Activation-readability, in short, is not behavioral controllability. We further find that a single MLP neuron is eval-correlated but not causal at both scales, and that scanning the real Pile yields a natural-text baseline competitive with the optimizer for the internal direction. A positive control validates our erasure detector, bounding an erasure-vs-rotation question earlier left open.
Large language models (LLMs) are costly intellectual assets that remain exposed to unauthorized redistribution and commercial misuse. Injected fingerprints, i.e., trigger--target pairs embedded in model behavior, offer a practical, black-box-verifiable ownership signal, but existing methods decouple the two stages of the fingerprint life cycle: how a fingerprint is constructed and how it is injected. Existing fingerprinting frameworks suffer from two limitations. Natural-language fingerprints are prone to accidental activation, and garbled fingerprints are easily filtered by perplexity-based detection. Furthermore, decoupling construction from injection leaves the latter unaware of the trigger's linguistic structure, missing the opportunity for targeted optimization. We argue that fingerprint construction should drive injection, and present a unified fingerprinting framework that jointly optimizes both stages. First, LCF constructs code-mixing fingerprints by combining low-resource languages under a semantic-density substitution rule and grammar-biased mixing, yielding triggers whose perplexity sits far below garbled baselines while avoiding the accidental-activation failures of natural-language triggers. Second, LCFEdit injects each fingerprint with a null-space projection derived from high-resource multilingual representations that preserves knowledge, augmented by a cross-lingual alignment step that steers the weight update toward the fingerprint language's representation subspace. This construction-aware injection ensures that the update is linguistically informed and therefore more stable. Extensive evaluations on imperceptibility, detectability, and harmlessness demonstrate persistent ownership verification with negligible impact on utility.
展开完整摘要收起摘要↓
Large language models (LLMs) are costly intellectual assets that remain exposed to unauthorized redistribution and commercial misuse. Injected fingerprints, i.e., trigger--target pairs embedded in model behavior, offer a practical, black-box-verifiable ownership signal, but existing methods decouple the two stages of the fingerprint life cycle: how a fingerprint is constructed and how it is injected. Existing fingerprinting frameworks suffer from two limitations. Natural-language fingerprints are prone to accidental activation, and garbled fingerprints are easily filtered by perplexity-based detection. Furthermore, decoupling construction from injection leaves the latter unaware of the trigger's linguistic structure, missing the opportunity for targeted optimization. We argue that fingerprint construction should drive injection, and present a unified fingerprinting framework that jointly optimizes both stages. First, LCF constructs code-mixing fingerprints by combining low-resource languages under a semantic-density substitution rule and grammar-biased mixing, yielding triggers whose perplexity sits far below garbled baselines while avoiding the accidental-activation failures of natural-language triggers. Second, LCFEdit injects each fingerprint with a null-space projection derived from high-resource multilingual representations that preserves knowledge, augmented by a cross-lingual alignment step that steers the weight update toward the fingerprint language's representation subspace. This construction-aware injection ensures that the update is linguistically informed and therefore more stable. Extensive evaluations on imperceptibility, detectability, and harmlessness demonstrate persistent ownership verification with negligible impact on utility.
Transformer language models are usually analyzed through vector geometry, yet ordered context and rotary position encoding introduce explicit phase structure into query-key interactions. This paper develops a bounded spectral framework for examining rotary phase alignment, hidden-state continuity, and semantic drift without treating language models as literal physical wave systems. It first identifies ordered hidden-state sequences, rather than vocabulary indices, as valid domains for spectral decomposition. It then derives the Rotary Position Embedding (RoPE) attention score as a sum of magnitude-weighted cosine terms and proves a local stability lemma: uniformly bounded phase displacement limits degradation of the corresponding pre-softmax score. To extend phase analysis beyond native RoPE coordinates, the paper defines complex modal coordinates over fixed orthonormal direction pairs and introduces a weighted coherence functional for hidden-state trajectories. These constructions support a strict distinction between representational continuity and execution-boundary admissibility. Internal coherence may describe preservation of task-relevant relations, but it cannot authorize a consequential transition. Positioned against existing geometric, spectral, phase-modulation, representation-analysis, and mechanistic-interpretability accounts, the framework contributes a theoretical and methodological program for determining when spectral structure explains continuity and when governance must remain an external predicate over execution.
展开完整摘要收起摘要↓
Transformer language models are usually analyzed through vector geometry, yet ordered context and rotary position encoding introduce explicit phase structure into query-key interactions. This paper develops a bounded spectral framework for examining rotary phase alignment, hidden-state continuity, and semantic drift without treating language models as literal physical wave systems. It first identifies ordered hidden-state sequences, rather than vocabulary indices, as valid domains for spectral decomposition. It then derives the Rotary Position Embedding (RoPE) attention score as a sum of magnitude-weighted cosine terms and proves a local stability lemma: uniformly bounded phase displacement limits degradation of the corresponding pre-softmax score. To extend phase analysis beyond native RoPE coordinates, the paper defines complex modal coordinates over fixed orthonormal direction pairs and introduces a weighted coherence functional for hidden-state trajectories. These constructions support a strict distinction between representational continuity and execution-boundary admissibility. Internal coherence may describe preservation of task-relevant relations, but it cannot authorize a consequential transition. Positioned against existing geometric, spectral, phase-modulation, representation-analysis, and mechanistic-interpretability accounts, the framework contributes a theoretical and methodological program for determining when spectral structure explains continuity and when governance must remain an external predicate over execution.
作者Barbara Kitchenham, Sebastián Pizard, Lech Madeyski, Ronnie de Souza Santos, Martin Shepperd, David Budgen
Context: Generative AI (GenAI) and Large Language Models (LLMs) are increasingly used for academic tasks in software engineering and beyond, including systematic literature reviews (SLRs). However, while capable of summarizing text, there is no guarantee they can meet the rigour, reliability, and transparency that SLRs require. Objectives: To support researchers intending to conduct SLRs using GenAI or those conducting empirical studies evaluating how well GenAI supports SLR tasks. Methods: First, we conducted a rapid review to identify studies that propose guidelines for evaluating and using GenAI and LLMs to support SLRs. Second, we drew on thought experiments, relevant guidance from the literature, and our own experience conducting SLRs and evaluating tools to develop recommendations for how to use and assess GenAI in the context of SLRs. Results: We discuss the problems researchers face when evaluating GenAI for SLRs. We identify and explain process issues to consider when planning, conducting, and reporting both SLRs using GenAI and evaluations of GenAI tools. Finally, we summarize our results as a set of process recommendations, which we name GUEST (GenAI Use and Evaluation in SLR Tasks). Conclusion: We argue that GenAI requires human oversight and is not currently capable of unsupervised systematic studies. However, it offers the prospect of cost-effective assistance for some repetitive tasks and for additional validation of some complex tasks. Our GUEST recommendations should help software engineering researchers both to conduct and report trustworthy SLRs using GenAI and to provide rigorous independent evaluation studies.
展开完整摘要收起摘要↓
Context: Generative AI (GenAI) and Large Language Models (LLMs) are increasingly used for academic tasks in software engineering and beyond, including systematic literature reviews (SLRs). However, while capable of summarizing text, there is no guarantee they can meet the rigour, reliability, and transparency that SLRs require. Objectives: To support researchers intending to conduct SLRs using GenAI or those conducting empirical studies evaluating how well GenAI supports SLR tasks. Methods: First, we conducted a rapid review to identify studies that propose guidelines for evaluating and using GenAI and LLMs to support SLRs. Second, we drew on thought experiments, relevant guidance from the literature, and our own experience conducting SLRs and evaluating tools to develop recommendations for how to use and assess GenAI in the context of SLRs. Results: We discuss the problems researchers face when evaluating GenAI for SLRs. We identify and explain process issues to consider when planning, conducting, and reporting both SLRs using GenAI and evaluations of GenAI tools. Finally, we summarize our results as a set of process recommendations, which we name GUEST (GenAI Use and Evaluation in SLR Tasks). Conclusion: We argue that GenAI requires human oversight and is not currently capable of unsupervised systematic studies. However, it offers the prospect of cost-effective assistance for some repetitive tasks and for additional validation of some complex tasks. Our GUEST recommendations should help software engineering researchers both to conduct and report trustworthy SLRs using GenAI and to provide rigorous independent evaluation studies.
Despite rapid progress in Multi-modal Large Language Models (MLLMs), understanding long-form videos is still bottlenecked by limited context windows. While recent keyframe sampling methods attempt to mitigate this by distilling video inputs into a compact set of query-relevant frames, navigating the vast spatio-temporal search space remains challenging, as spatial detail and temporal coverage often conflict. To address this, we introduce LENS, a training-free keyframe sampling framework that dynamically decides when to zoom in for fine-grained details and when to zoom out for broader context based on the text query. Concretely, LENS adaptively allocates a limited frame budget between spatial zoom-ins, which highlight query-relevant regions within individual frames, and temporal zoom-outs, which expand the temporal scope through multi-frame aggregation, enabling the model to reason across multiple granularities while capturing both high-fidelity details and long-range context. Across diverse long-form video benchmarks, LENS consistently outperforms prior state-of-the-art keyframe sampling methods and delivers substantial gains over uniform sampling, improving Video-MME accuracy from 53.3% to 60.7% with Qwen2.5-VL.Code is available at https://github.com/zhangce01/LENS.
展开完整摘要收起摘要↓
Despite rapid progress in Multi-modal Large Language Models (MLLMs), understanding long-form videos is still bottlenecked by limited context windows. While recent keyframe sampling methods attempt to mitigate this by distilling video inputs into a compact set of query-relevant frames, navigating the vast spatio-temporal search space remains challenging, as spatial detail and temporal coverage often conflict. To address this, we introduce LENS, a training-free keyframe sampling framework that dynamically decides when to zoom in for fine-grained details and when to zoom out for broader context based on the text query. Concretely, LENS adaptively allocates a limited frame budget between spatial zoom-ins, which highlight query-relevant regions within individual frames, and temporal zoom-outs, which expand the temporal scope through multi-frame aggregation, enabling the model to reason across multiple granularities while capturing both high-fidelity details and long-range context. Across diverse long-form video benchmarks, LENS consistently outperforms prior state-of-the-art keyframe sampling methods and delivers substantial gains over uniform sampling, improving Video-MME accuracy from 53.3% to 60.7% with Qwen2.5-VL.Code is available at https://github.com/zhangce01/LENS.
Multimodal large language models (MLLMs) can connect visualization patterns to external causes, consequences, and domain knowledge, but the evidential basis of these interpretations is often unclear. We present an exploratory study of 102 visualizations from four sources, three MLLMs, and four input conditions that vary access to the image, source-specific accessible chart context, and withheld-context framing. Across 1,224 descriptions, we analyze model-attributed DIRECT, DERIVED, and SPECULATIVE labels and conduct an automated audit of numeric agreement. Accessible chart context shifted Gemini and GPT toward DIRECT claims and improved numeric agreement for some models. Adding the image to the full context did not yield a consistent numeric benefit, and the withheld-context prompt did not reliably increase cautious language. The prompt-defined Real-World Significance section remained predominantly SPECULATIVE. These results motivate accessible description systems that distinguish claims supported by supplied evidence from model-supplied interpretation
展开完整摘要收起摘要↓
Multimodal large language models (MLLMs) can connect visualization patterns to external causes, consequences, and domain knowledge, but the evidential basis of these interpretations is often unclear. We present an exploratory study of 102 visualizations from four sources, three MLLMs, and four input conditions that vary access to the image, source-specific accessible chart context, and withheld-context framing. Across 1,224 descriptions, we analyze model-attributed DIRECT, DERIVED, and SPECULATIVE labels and conduct an automated audit of numeric agreement. Accessible chart context shifted Gemini and GPT toward DIRECT claims and improved numeric agreement for some models. Adding the image to the full context did not yield a consistent numeric benefit, and the withheld-context prompt did not reliably increase cautious language. The prompt-defined Real-World Significance section remained predominantly SPECULATIVE. These results motivate accessible description systems that distinguish claims supported by supplied evidence from model-supplied interpretation
Urban decision-making requires integrating heterogeneous spatial data. While current GIS tools handle geometric computation efficiently, they lack the semantic reasoning to guide complex workflows. Analysts manually manage data discovery, spatial boundaries, and measurement semantics, risking aggregation errors. We present UrbanTrace, a visual analytics system that transforms manual spatial data-wrangling into a transparent, node-based collaborative workflow with context-aware AI agents. Using an offline profiler to extract semantic and geometric metadata, UrbanTrace grounds LLMs in real-world data distributions. This enables specialized agents to retrieve datasets based on high-level goals and automatically enforce valid spatial aggregations. To make harmonization explicit, three interactive views: an Integration Provenance Graph, Multivariate Priority Map, and Spatial Delta Map, allow users to explore how conclusions shift across spatial configurations. We evaluate UrbanTrace on 28 urban scenarios spanning 112 datasets. Quantitative ablations show our profiling significantly outperforms baseline LLMs in data discovery, achieving 100% semantic and 87% geometric validity in spatial mapping. Through real-world case studies and expert interviews, we demonstrate that UrbanTrace turns spatial aggregation sensitivity from a methodological burden into an exploratory visual asset.
展开完整摘要收起摘要↓
Urban decision-making requires integrating heterogeneous spatial data. While current GIS tools handle geometric computation efficiently, they lack the semantic reasoning to guide complex workflows. Analysts manually manage data discovery, spatial boundaries, and measurement semantics, risking aggregation errors. We present UrbanTrace, a visual analytics system that transforms manual spatial data-wrangling into a transparent, node-based collaborative workflow with context-aware AI agents. Using an offline profiler to extract semantic and geometric metadata, UrbanTrace grounds LLMs in real-world data distributions. This enables specialized agents to retrieve datasets based on high-level goals and automatically enforce valid spatial aggregations. To make harmonization explicit, three interactive views: an Integration Provenance Graph, Multivariate Priority Map, and Spatial Delta Map, allow users to explore how conclusions shift across spatial configurations. We evaluate UrbanTrace on 28 urban scenarios spanning 112 datasets. Quantitative ablations show our profiling significantly outperforms baseline LLMs in data discovery, achieving 100% semantic and 87% geometric validity in spatial mapping. Through real-world case studies and expert interviews, we demonstrate that UrbanTrace turns spatial aggregation sensitivity from a methodological burden into an exploratory visual asset.
Automated verification of numerical claims is a challenging problem, as it requires both language understanding and quantitative reasoning. This paper describes our system for CLEF 2026 CheckThat! Task 2, which focuses on ranking reasoning traces generated by large language models (LLMs) and predicting a final verdict for numerical claims in English and Arabic. We explore two approaches. The first approach fine-tunes an LLM-based verifier using LoRA to score each reasoning trace independently as a binary classification problem, and selects the final verdict using Best-of-N selection. We further experiment with adaptive sub-claim decomposition to break complex claims into simpler parts before verification. The second approach uses a lightweight TF-IDF reward model with handcrafted numeric and temporal overlap features to score traces, and aggregates scores by verdict group to determine the final prediction. For Arabic, we compare a general multilingual model against AraBERT, a language-specific model pretrained on Arabic text. Our results show that the LLM-based approach outperforms the lightweight reward model on most metrics, particularly Recall@5, while the reward-based approach shows stronger performance on the Conflicting class. Sub-claim decomposition did not improve performance, suggesting that claim splitting introduces noise rather than aiding reasoning. For Arabic, AraBERT outperforms the multilingual baseline across most metrics.
展开完整摘要收起摘要↓
Automated verification of numerical claims is a challenging problem, as it requires both language understanding and quantitative reasoning. This paper describes our system for CLEF 2026 CheckThat! Task 2, which focuses on ranking reasoning traces generated by large language models (LLMs) and predicting a final verdict for numerical claims in English and Arabic. We explore two approaches. The first approach fine-tunes an LLM-based verifier using LoRA to score each reasoning trace independently as a binary classification problem, and selects the final verdict using Best-of-N selection. We further experiment with adaptive sub-claim decomposition to break complex claims into simpler parts before verification. The second approach uses a lightweight TF-IDF reward model with handcrafted numeric and temporal overlap features to score traces, and aggregates scores by verdict group to determine the final prediction. For Arabic, we compare a general multilingual model against AraBERT, a language-specific model pretrained on Arabic text. Our results show that the LLM-based approach outperforms the lightweight reward model on most metrics, particularly Recall@5, while the reward-based approach shows stronger performance on the Conflicting class. Sub-claim decomposition did not improve performance, suggesting that claim splitting introduces noise rather than aiding reasoning. For Arabic, AraBERT outperforms the multilingual baseline across most metrics.
Pretraining data processing is critical to the downstream performance of Large Language Models (LLMs). However, many existing approaches define a fixed processing strategy at the corpus or domain level and apply it uniformly to many examples, without adapting to the needs of each example. We propose DataOrchestra, a framework that unifies different processing operations and orchestrates an example-specific pipeline for each example. Given a chunk of pretraining data, an orchestrator decides whether to drop, untouch, or clean it. For a chunk to be cleaned, it selects one or more downstream operations, ranging from programmatic editing to different forms of LLM-based rewriting. For each rewriting step, it further generates a concrete instruction, which is executed by the corresponding downstream tool model. We pretrain models from 0.5B to 7B from scratch on web data processed by DataOrchestra and observe stable average gains over individual data-processing methods across 11 benchmarks. DataOrchestra is also effective for math continued pretraining and outperforms stronger processing baselines, while reducing processing compute by skipping unnecessary downstream operations.
展开完整摘要收起摘要↓
Pretraining data processing is critical to the downstream performance of Large Language Models (LLMs). However, many existing approaches define a fixed processing strategy at the corpus or domain level and apply it uniformly to many examples, without adapting to the needs of each example. We propose DataOrchestra, a framework that unifies different processing operations and orchestrates an example-specific pipeline for each example. Given a chunk of pretraining data, an orchestrator decides whether to drop, untouch, or clean it. For a chunk to be cleaned, it selects one or more downstream operations, ranging from programmatic editing to different forms of LLM-based rewriting. For each rewriting step, it further generates a concrete instruction, which is executed by the corresponding downstream tool model. We pretrain models from 0.5B to 7B from scratch on web data processed by DataOrchestra and observe stable average gains over individual data-processing methods across 11 benchmarks. DataOrchestra is also effective for math continued pretraining and outperforms stronger processing baselines, while reducing processing compute by skipping unnecessary downstream operations.
Advanced Persistent Threats (APTs) are difficult to detect and interpret due to their multi-stage and stealthy nature. While recent autonomous defense systems leverage provenance graphs and learning-based models for detection and mitigation, their outputs remain largely machine-oriented and difficult for analysts to interpret. Large language models (LLMs) offer a promising interface for report generation, but often produce hallucinated or weakly grounded content. In this paper, we propose DeepFaith, an evidence-grounded framework for faithful incident reporting in multi-stage APT defense. DeepFaith transforms structured outputs from autonomous defense and explainability modules into natural-language reports that are explicitly aligned with underlying system evidence. The framework integrates a unified evidence representation, evidence-grounded prompting, faithfulness-aware generation, and post-generation verification to ensure that all generated statements are supported. Experiments in a realistic enterprise testbed demonstrate that DeepFaith improves faithfulness from 0.68 to 0.92, reduces unsupported claims from 0.32 to 0.08, and increases temporal consistency from 0.6 to 0.88, while maintaining concise reports and lower error rates than existing template-based and LLM-based solutions. These results show that evidence-grounded generation enables reliable, interpretable, and actionable reporting for security operations centers.
展开完整摘要收起摘要↓
Advanced Persistent Threats (APTs) are difficult to detect and interpret due to their multi-stage and stealthy nature. While recent autonomous defense systems leverage provenance graphs and learning-based models for detection and mitigation, their outputs remain largely machine-oriented and difficult for analysts to interpret. Large language models (LLMs) offer a promising interface for report generation, but often produce hallucinated or weakly grounded content. In this paper, we propose DeepFaith, an evidence-grounded framework for faithful incident reporting in multi-stage APT defense. DeepFaith transforms structured outputs from autonomous defense and explainability modules into natural-language reports that are explicitly aligned with underlying system evidence. The framework integrates a unified evidence representation, evidence-grounded prompting, faithfulness-aware generation, and post-generation verification to ensure that all generated statements are supported. Experiments in a realistic enterprise testbed demonstrate that DeepFaith improves faithfulness from 0.68 to 0.92, reduces unsupported claims from 0.32 to 0.08, and increases temporal consistency from 0.6 to 0.88, while maintaining concise reports and lower error rates than existing template-based and LLM-based solutions. These results show that evidence-grounded generation enables reliable, interpretable, and actionable reporting for security operations centers.
Large language models place structured concepts on geometrically faithful manifolds: weekdays lie on a circle, months on another, usually taken to be a fixed world-model the network stores and looks up. We show that context is king: the structure a model actually uses is set by the in-context specification. A declarative rule fixes not only which relations the geometry encodes but its topology type: the same tokens form a cycle or a branching tree on command, built even on arbitrary, meaning-free tokens with no prior to inherit, which a relabeled stored shape cannot do. When the specification conflicts with a strong pretrained prior, the context-set geometry dominates it in capable models, read from the same activations (representational similarity 0.6--0.9 to the imposed structure versus near-zero to the prior), across the priors we test and both families we study (Gemma, Qwen). Activation patching shows the map is causally used, not a probe correlate: swapping one entity's activation for another's makes the model answer with the other entity's successor under the imposed order. A rough map forms readily, present even in small and base models; what scale gates is using it cleanly: clean dominance and the causal crossover emerge only in the larger models (up to Gemma-31B and Qwen-27B) and weaken or reverse below, so a mechanism present in a large model can be absent in a smaller one of the same family. Whether the model builds this geometry anew or reconfigures a stored one we leave open; operationally, the geometry it uses is the one the context specifies.
展开完整摘要收起摘要↓
Large language models place structured concepts on geometrically faithful manifolds: weekdays lie on a circle, months on another, usually taken to be a fixed world-model the network stores and looks up. We show that context is king: the structure a model actually uses is set by the in-context specification. A declarative rule fixes not only which relations the geometry encodes but its topology type: the same tokens form a cycle or a branching tree on command, built even on arbitrary, meaning-free tokens with no prior to inherit, which a relabeled stored shape cannot do. When the specification conflicts with a strong pretrained prior, the context-set geometry dominates it in capable models, read from the same activations (representational similarity 0.6--0.9 to the imposed structure versus near-zero to the prior), across the priors we test and both families we study (Gemma, Qwen). Activation patching shows the map is causally used, not a probe correlate: swapping one entity's activation for another's makes the model answer with the other entity's successor under the imposed order. A rough map forms readily, present even in small and base models; what scale gates is using it cleanly: clean dominance and the causal crossover emerge only in the larger models (up to Gemma-31B and Qwen-27B) and weaken or reverse below, so a mechanism present in a large model can be absent in a smaller one of the same family. Whether the model builds this geometry anew or reconfigures a stored one we leave open; operationally, the geometry it uses is the one the context specifies.
We introduce a novel learning problem: decoding gaze into natural language descriptions of human goals across diverse visual tasks. Unlike prior work, which frames gaze decoding as a discriminative task over predefined categories, we formulate it as a generative learning problem: training a model to produce free-form descriptions that capture the rich nuances and open-ended nature of human intentions beyond fixed labels. To this end, we introduce Gazette, the first gaze-to-text decoding framework. Based on multimodal large language models (MLLMs), Gazette learns to decode gaze scanpaths into natural language for goals that may extend beyond categorical labels and require articulation in natural language. To help Gazette filter out individual differences in gaze behavior and learn the goal-specific spatiotemporal dynamics crucial for generating accurate natural language goal descriptions, we propose a novel strategy that leverages the encyclopedic knowledge and reasoning abilities of a large language model to synthesize natural language explanations of goal-directed attentional behavior called think-aloud transcripts. Instruction tuning on these synthetic narratives allows Gazette to achieve state-of-the-art performance in gaze decoding across multiple tasks, demonstrating its generalizability and versatility, thereby enabling gaze to serve as a powerful, non-intrusive cue for inferring human goals and intentions in diverse scenarios.
展开完整摘要收起摘要↓
We introduce a novel learning problem: decoding gaze into natural language descriptions of human goals across diverse visual tasks. Unlike prior work, which frames gaze decoding as a discriminative task over predefined categories, we formulate it as a generative learning problem: training a model to produce free-form descriptions that capture the rich nuances and open-ended nature of human intentions beyond fixed labels. To this end, we introduce Gazette, the first gaze-to-text decoding framework. Based on multimodal large language models (MLLMs), Gazette learns to decode gaze scanpaths into natural language for goals that may extend beyond categorical labels and require articulation in natural language. To help Gazette filter out individual differences in gaze behavior and learn the goal-specific spatiotemporal dynamics crucial for generating accurate natural language goal descriptions, we propose a novel strategy that leverages the encyclopedic knowledge and reasoning abilities of a large language model to synthesize natural language explanations of goal-directed attentional behavior called think-aloud transcripts. Instruction tuning on these synthetic narratives allows Gazette to achieve state-of-the-art performance in gaze decoding across multiple tasks, demonstrating its generalizability and versatility, thereby enabling gaze to serve as a powerful, non-intrusive cue for inferring human goals and intentions in diverse scenarios.