Large Language Models (LLMs) are being increasingly used in everyday applications. A major challenge in the context of LLMs or Artificial Intelligence (AI) in general is to ensure privacy when using them, meaning that personally identifiable information (PII) is removed from any text that enters an LLM. These challenges have become more urgent with novel EU legislation. Uncertainty around LLM usage with respect to privacy concerns in EU countries can be a major blocker for the speed of innovation and transfer from research to applications. Here we present Redakto, a tool that can be used for anonymizing text prior to feeding it to an LLM or other downstream text processing. We provide state-of-the-art functionalities for both redaction of PII but also when used for pseudonymization. These functionalities are exposed such that they can easily be used by end-users, through the Redakto web application, and by developers and researchers, via REST APIs and model context protocol (MCP) hooks. The implementation is fully open source, requires modest compute resources, and can be readily deployed on local hardware. In contrast to prior work and in order to better assess the quality of the anonymized texts, we conduct extensive empirical evaluations on textual data from legal and medical domain with respect to both privacy and utility of the redacted texts. Our empirical results demonstrate that the texts anonymized with different redaction strategies achieve utility scores on par with the original texts, suggesting that anonymization with Redakto can be used for LLM tasks without substantial negative impact for the tasks we explored.
展开完整摘要收起摘要↓
Large Language Models (LLMs) are being increasingly used in everyday applications. A major challenge in the context of LLMs or Artificial Intelligence (AI) in general is to ensure privacy when using them, meaning that personally identifiable information (PII) is removed from any text that enters an LLM. These challenges have become more urgent with novel EU legislation. Uncertainty around LLM usage with respect to privacy concerns in EU countries can be a major blocker for the speed of innovation and transfer from research to applications. Here we present Redakto, a tool that can be used for anonymizing text prior to feeding it to an LLM or other downstream text processing. We provide state-of-the-art functionalities for both redaction of PII but also when used for pseudonymization. These functionalities are exposed such that they can easily be used by end-users, through the Redakto web application, and by developers and researchers, via REST APIs and model context protocol (MCP) hooks. The implementation is fully open source, requires modest compute resources, and can be readily deployed on local hardware. In contrast to prior work and in order to better assess the quality of the anonymized texts, we conduct extensive empirical evaluations on textual data from legal and medical domain with respect to both privacy and utility of the redacted texts. Our empirical results demonstrate that the texts anonymized with different redaction strategies achieve utility scores on par with the original texts, suggesting that anonymization with Redakto can be used for LLM tasks without substantial negative impact for the tasks we explored.
作者Ruhai Lin, Yiyang Guo, Rui-Jie Zhu, Hao Ye, Jason K. Eshraghian
Looped language models improve reasoning and knowledge manipulation by applying shared computation repeatedly. Existing systems usually repeat an entire layer stack, although a mixer and a dense feed-forward network (FFN) perform different operations and have different costs. We ask a narrower question: what should loop? We view recurrence as repeated composition of a state update and argue that an application is valuable when it exposes a new cross-position influence direction that remains observable at the task readout. Iterative Transport Rank (ITR) describes the cumulative influence trajectory; marginal ITR describes the nonredundant influence contributed by successive applications. This view motivates MixerLoop, which repeats each Gated DeltaNet mixer while applying its dense FFN once. We compare MixerLoop with no recurrence and full-block recurrence at 15M and 110M parameters under the same data, initialization, and architecture. A finite context-off intervention tests whether later mixer applications produce distinct, non-negligible, and beneficial changes at the final language-model readout. MixerLoop surpasses FullLoop on aggregate CORE at 15M and retains 41.5% of its CORE improvement at 110M while reducing recurrent-backbone projection FLOPs by 45.9%. These results show that the benefits of recurrent depth can be retained without repeatedly executing the dense FFN.
展开完整摘要收起摘要↓
Looped language models improve reasoning and knowledge manipulation by applying shared computation repeatedly. Existing systems usually repeat an entire layer stack, although a mixer and a dense feed-forward network (FFN) perform different operations and have different costs. We ask a narrower question: what should loop? We view recurrence as repeated composition of a state update and argue that an application is valuable when it exposes a new cross-position influence direction that remains observable at the task readout. Iterative Transport Rank (ITR) describes the cumulative influence trajectory; marginal ITR describes the nonredundant influence contributed by successive applications. This view motivates MixerLoop, which repeats each Gated DeltaNet mixer while applying its dense FFN once. We compare MixerLoop with no recurrence and full-block recurrence at 15M and 110M parameters under the same data, initialization, and architecture. A finite context-off intervention tests whether later mixer applications produce distinct, non-negligible, and beneficial changes at the final language-model readout. MixerLoop surpasses FullLoop on aggregate CORE at 15M and retains 41.5% of its CORE improvement at 110M while reducing recurrent-backbone projection FLOPs by 45.9%. These results show that the benefits of recurrent depth can be retained without repeatedly executing the dense FFN.
Small language models can grade open-ended examination answers as reliably as substantially more expensive models when they grade against an explicit rubric. We test this claim as the design principle behind any-to-bench: a frontier model reads source documents once, at ingestion, to extract each question and its rubric; lower-cost models then perform all repeated grading work. We evaluate six cost-efficient model configurations from two model families at three reasoning-effort levels. Each configuration answers 24 open-ended examination questions, and each also grades every answer sheet three times, yielding 3,456 per-question grades. Scores depend overwhelmingly on the answer being graded: answer identity explains 95.6% of score variance, whereas judge identity explains only 0.2%. Raising a writer's reasoning effort moves earned scores by as much as 0.143 of full marks, while raising a judge's reasoning effort moves assigned scores by at most 0.006. Six frontier-tier judges, added as a check, reproduce these scores and are no more reliable as a panel. Two ablations then decompose the rubric on the same questions and answers. Removing its criteria and levels while keeping the official answer changes nothing measurable. Removing the official answer as well collapses reliability (ICC 0.888 to 0.628), inflates scores, and makes judge reasoning effort matter again. The rubric is what decouples grading from judge intelligence, and within the rubric the official answer does nearly all the work. We find no evidence of length preference or same-family preference under rubric-anchored grading.
展开完整摘要收起摘要↓
Small language models can grade open-ended examination answers as reliably as substantially more expensive models when they grade against an explicit rubric. We test this claim as the design principle behind any-to-bench: a frontier model reads source documents once, at ingestion, to extract each question and its rubric; lower-cost models then perform all repeated grading work. We evaluate six cost-efficient model configurations from two model families at three reasoning-effort levels. Each configuration answers 24 open-ended examination questions, and each also grades every answer sheet three times, yielding 3,456 per-question grades. Scores depend overwhelmingly on the answer being graded: answer identity explains 95.6% of score variance, whereas judge identity explains only 0.2%. Raising a writer's reasoning effort moves earned scores by as much as 0.143 of full marks, while raising a judge's reasoning effort moves assigned scores by at most 0.006. Six frontier-tier judges, added as a check, reproduce these scores and are no more reliable as a panel. Two ablations then decompose the rubric on the same questions and answers. Removing its criteria and levels while keeping the official answer changes nothing measurable. Removing the official answer as well collapses reliability (ICC 0.888 to 0.628), inflates scores, and makes judge reasoning effort matter again. The rubric is what decouples grading from judge intelligence, and within the rubric the official answer does nearly all the work. We find no evidence of length preference or same-family preference under rubric-anchored grading.
作者Kyle Chickering, Wei-An Lin, Swayam Bhanded, Dan Saunders, Akshat Tripathi, Jiaming Song, Shyamal Buch, Xinchen Yan
Compute-optimal scaling laws guide the training of frontier language models yet remain largely unexplored for visual generation. We present a systematic scaling law study for text-to-image diffusion models using Abra, a controlled family of flow-matching transformers trained across three orders of magnitude worth of compute ($10^{19}$ to $10^{22}$ FLOPs), reaching significantly larger compute budgets than previous works. We demonstrate that diffusion models scale just as predictably as language models but require far more data to train optimally: compute optimality occurs at approximately $200$ image tokens per parameter, ten times the Chinchilla compute-optimal prescription for LLMs. We show that unlike language models, diffusion models are robust to overtraining and that practitioners should err on the side of more data rather than a larger model. Finally, we show that this predictability extends beyond training loss to generative quality metrics, optimal CFG settings, representation quality, and even the shape of the training curves, which collapse onto a universal form.
展开完整摘要收起摘要↓
Compute-optimal scaling laws guide the training of frontier language models yet remain largely unexplored for visual generation. We present a systematic scaling law study for text-to-image diffusion models using Abra, a controlled family of flow-matching transformers trained across three orders of magnitude worth of compute ($10^{19}$ to $10^{22}$ FLOPs), reaching significantly larger compute budgets than previous works. We demonstrate that diffusion models scale just as predictably as language models but require far more data to train optimally: compute optimality occurs at approximately $200$ image tokens per parameter, ten times the Chinchilla compute-optimal prescription for LLMs. We show that unlike language models, diffusion models are robust to overtraining and that practitioners should err on the side of more data rather than a larger model. Finally, we show that this predictability extends beyond training loss to generative quality metrics, optimal CFG settings, representation quality, and even the shape of the training curves, which collapse onto a universal form.
作者Tu Nguyen, Thanh Nguyen, Huy Nguyen, Viet Nguyen, Tien N. Nguyen, Vu Nguyen
Testing RESTful APIs requires generating sequences of API calls that satisfy dependencies among operations, parameters, and runtime-created resources. Recent LLM-based approaches infer such dependencies and generate test sequences from OpenAPI specifications, but they often treat LLM-inferred relationships as correct without execution-based validation. This can introduce spurious dependencies, miss feasible operation chains, and produce infeasible tests. In this paper, we propose APIPilot}, an execution-validated framework for REST API testing. APIPilot first derives candidate producer-consumer dependencies from OpenAPI specifications using structural heuristics and LLM-based semantic reasoning. It then treats these dependencies as hypotheses and validates them through concrete API executions before using them for test generation. The validated dependencies are organized into a dependency graph from which APIPilot constructs coverage-aware workflows via bounded top-k graph traversal, separating semantic dependency inference from sequence construction. To improve subsequent tests, APIPilot further performs response-driven refinement: runtime responses are analyzed to update resource pools, adjust input-generation constraints, and prune or revise invalid dependency mappings. Empirical evaluation on 16 real-world REST API services shows that APIPilot achieves 92.3% operation coverage, up to 58.6% code coverage, and an 88.1% workflow execution success rate, outperforming both LLM-based and traditional REST API testing baselines. APIPilot also detects 197 unique 5xx failures and specification-execution mismatches, demonstrating the benefit of grounding dependency inference in execution feedback.
展开完整摘要收起摘要↓
Testing RESTful APIs requires generating sequences of API calls that satisfy dependencies among operations, parameters, and runtime-created resources. Recent LLM-based approaches infer such dependencies and generate test sequences from OpenAPI specifications, but they often treat LLM-inferred relationships as correct without execution-based validation. This can introduce spurious dependencies, miss feasible operation chains, and produce infeasible tests. In this paper, we propose APIPilot}, an execution-validated framework for REST API testing. APIPilot first derives candidate producer-consumer dependencies from OpenAPI specifications using structural heuristics and LLM-based semantic reasoning. It then treats these dependencies as hypotheses and validates them through concrete API executions before using them for test generation. The validated dependencies are organized into a dependency graph from which APIPilot constructs coverage-aware workflows via bounded top-k graph traversal, separating semantic dependency inference from sequence construction. To improve subsequent tests, APIPilot further performs response-driven refinement: runtime responses are analyzed to update resource pools, adjust input-generation constraints, and prune or revise invalid dependency mappings. Empirical evaluation on 16 real-world REST API services shows that APIPilot achieves 92.3% operation coverage, up to 58.6% code coverage, and an 88.1% workflow execution success rate, outperforming both LLM-based and traditional REST API testing baselines. APIPilot also detects 197 unique 5xx failures and specification-execution mismatches, demonstrating the benefit of grounding dependency inference in execution feedback.
Despite their widespread use, Large Language Models (LLMs) remain limited by a fundamental problem: the generation of plausible but false content, known as hallucinations. Most existing detection methods operate at the answer or sentence level, yet per-token detection is essential for localizing hallucinated spans and enabling fine-grained interventions. In this paper, we explore the use of the Mixture-of-Experts (MoE) paradigm to address this gap. In MoE architectures, a single forward pass activates a sparse subset of experts (i.e., distinct feedforward networks per layer) via a routing mechanism, producing internal signals (e.g., router entropy, expert disagreement, and expert usage patterns) that are unavailable in dense architectures and have not been previously exploited for hallucination detection. To this end, we introduce InnerExpert, the first method to leverage these MoE-specific signals for per-token hallucination detection. InnerExpert combines routing-level and standard transformer signals into compact per-token feature vectors, classified by a lightweight detector trained on labels produced by an LLM-as-a-judge pipeline, which enables continuous model updates without manual annotation. Our results show that InnerExpert outperforms existing methods across five datasets and two MoE architectures, achieving up to 0.91 answer-level and 0.76 token-level AUROC, while requiring only a single forward pass.
展开完整摘要收起摘要↓
Despite their widespread use, Large Language Models (LLMs) remain limited by a fundamental problem: the generation of plausible but false content, known as hallucinations. Most existing detection methods operate at the answer or sentence level, yet per-token detection is essential for localizing hallucinated spans and enabling fine-grained interventions. In this paper, we explore the use of the Mixture-of-Experts (MoE) paradigm to address this gap. In MoE architectures, a single forward pass activates a sparse subset of experts (i.e., distinct feedforward networks per layer) via a routing mechanism, producing internal signals (e.g., router entropy, expert disagreement, and expert usage patterns) that are unavailable in dense architectures and have not been previously exploited for hallucination detection. To this end, we introduce InnerExpert, the first method to leverage these MoE-specific signals for per-token hallucination detection. InnerExpert combines routing-level and standard transformer signals into compact per-token feature vectors, classified by a lightweight detector trained on labels produced by an LLM-as-a-judge pipeline, which enables continuous model updates without manual annotation. Our results show that InnerExpert outperforms existing methods across five datasets and two MoE architectures, achieving up to 0.91 answer-level and 0.76 token-level AUROC, while requiring only a single forward pass.
When students use large language models for writing, the dominant explanation for disengagement is dispositional: they are over-reliant, and the remedy is to scaffold self-regulation. We argue that a structural explanation is needed, offering an alternative basis for design interventions to support appropriate AI-assisted writing. Current LLM writing interfaces induce procedural collapse: the replacement of an iterative, self-paced writing process with a single output that shifts the writer's task from generation to comprehensive evaluation. Because that evaluation is costly, shallow engagement becomes the default, and the cognitive work writing was supposed to produce goes unperformed. The framework points toward design directions that reduce the burden on writers to self-regulate, including decomposed interaction, goal elicitation as a default first step, and single-level output. They complement metacognitive scaffolding by restructuring the interaction itself.
展开完整摘要收起摘要↓
When students use large language models for writing, the dominant explanation for disengagement is dispositional: they are over-reliant, and the remedy is to scaffold self-regulation. We argue that a structural explanation is needed, offering an alternative basis for design interventions to support appropriate AI-assisted writing. Current LLM writing interfaces induce procedural collapse: the replacement of an iterative, self-paced writing process with a single output that shifts the writer's task from generation to comprehensive evaluation. Because that evaluation is costly, shallow engagement becomes the default, and the cognitive work writing was supposed to produce goes unperformed. The framework points toward design directions that reduce the burden on writers to self-regulate, including decomposed interaction, goal elicitation as a default first step, and single-level output. They complement metacognitive scaffolding by restructuring the interaction itself.
Large language models (LLMs) are increasingly used for scientific hypothesis generation. However, evaluating generated hypotheses remains a challenge for trustworthy AI-enabled scientific workflows. Existing approaches often use LLMs as judges or rely on semantic similarity, which can favor familiar ideas over novel ones. We propose a logit-based energy scoring method that evaluates hypotheses using a language model's intrinsic confidence rather than comparative judgment. We benchmarked seven language models on 1,323 papers across 12 disciplines. Each paper was paired with its hypothesis and fifteen incorrect alternatives. Intrinsic scoring reached 33.0% Hit@1 pooled across both scorers, compared with 16.6% for prompted listwise ranking. The strongest configuration, a 1-billion-parameter model using logit-based energy scoring, reached 53.1%, though this was the maximum across 14 model-by-scorer combinations selected post hoc. Overall, intrinsic model confidence shows potential for scientific hypothesis evaluation. This study also motivates future research on confidence-based methods for trustworthy AI-enabled scientific discovery.
展开完整摘要收起摘要↓
Large language models (LLMs) are increasingly used for scientific hypothesis generation. However, evaluating generated hypotheses remains a challenge for trustworthy AI-enabled scientific workflows. Existing approaches often use LLMs as judges or rely on semantic similarity, which can favor familiar ideas over novel ones. We propose a logit-based energy scoring method that evaluates hypotheses using a language model's intrinsic confidence rather than comparative judgment. We benchmarked seven language models on 1,323 papers across 12 disciplines. Each paper was paired with its hypothesis and fifteen incorrect alternatives. Intrinsic scoring reached 33.0% Hit@1 pooled across both scorers, compared with 16.6% for prompted listwise ranking. The strongest configuration, a 1-billion-parameter model using logit-based energy scoring, reached 53.1%, though this was the maximum across 14 model-by-scorer combinations selected post hoc. Overall, intrinsic model confidence shows potential for scientific hypothesis evaluation. This study also motivates future research on confidence-based methods for trustworthy AI-enabled scientific discovery.
Large language models (LLMs) have demonstrated strong performance on structured reasoning tasks, but what they encode and whether it informs model behavior remain unclear. We investigate this question through geometric reasoning, using parametric CAD constraints as a controlled testbed for separating local pairwise relations from sketch-level constraint status. By probing the hidden states of six frozen decoder-only LLMs, we examine four properties: linear decodability, forced-choice generation, activation-level influence, and behavioral steerability. Pretraining substantially improves the decoding of local geometric relations, and this advantage persists after accounting for positional cues with shuffled-order controls. In contrast, sketch-level DOF status is already highly decodable from randomly initialized representations and improves only modestly with pretraining, indicating that much of its probe performance is available without learned weights. Further analyses show that decodable information is not always actionable. Generation often fails to express this information, and on the two intervention-tested backbones, activation-restoration effects at the patched entity position vanish while decodability persists across depth. Mean-difference steering also does not reliably control outputs. These results show that decodability, generation, activation-level influence, and steerability can diverge in the tested setting. The audit provides a controlled way to distinguish failures to encode geometric structure from failures to express or control encoded information.
展开完整摘要收起摘要↓
Large language models (LLMs) have demonstrated strong performance on structured reasoning tasks, but what they encode and whether it informs model behavior remain unclear. We investigate this question through geometric reasoning, using parametric CAD constraints as a controlled testbed for separating local pairwise relations from sketch-level constraint status. By probing the hidden states of six frozen decoder-only LLMs, we examine four properties: linear decodability, forced-choice generation, activation-level influence, and behavioral steerability. Pretraining substantially improves the decoding of local geometric relations, and this advantage persists after accounting for positional cues with shuffled-order controls. In contrast, sketch-level DOF status is already highly decodable from randomly initialized representations and improves only modestly with pretraining, indicating that much of its probe performance is available without learned weights. Further analyses show that decodable information is not always actionable. Generation often fails to express this information, and on the two intervention-tested backbones, activation-restoration effects at the patched entity position vanish while decodability persists across depth. Mean-difference steering also does not reliably control outputs. These results show that decodability, generation, activation-level influence, and steerability can diverge in the tested setting. The audit provides a controlled way to distinguish failures to encode geometric structure from failures to express or control encoded information.
Categorising invoices into the correct General Ledger (GL) code underpins financial reporting and tax compliance. This is a skilled accounting judgement rather than a routine task: the correct category depends subtly on the nature of the purchasing business, the vendor and the invoice text. Whilst AI is increasingly being adopted across industries to automate tasks, including invoice categorisation, implementations built on in-house small language models (SLMs) can simultaneously reduce cost and improve data security, confidentiality, and interpretability. We investigate this approach by first analysing the pre-trained embedding geometry of a small sentence transformer (SBERT) and classic SLM (DeBERTa). The sentence-embedding space of this financial corpus is globally anisotropic but composed of locally isotropic clusters, extending prior token-level findings to sentence embeddings in a financial setting, and these clusters are strongly correlated with the vendor identity. SBERT fine-tuned on a single GPU reaches 0.96 accuracy on invoice classification, above both a zero-shot LLM and a vendor identity baseline, increasing performance for smaller, challenging categories and new clients. For this important generalisation problem, SBERT reaches 0.9 F1 with roughly 100 client-specific invoices, showing that an in-house SLM implementation is promising. Combining these results with geometric analysis shows that pre-trained embedding geometry is associated with classification performance and reveals a counterintuitive finding that a structured input that would help a human reader does not improve the SLM performance.
展开完整摘要收起摘要↓
Categorising invoices into the correct General Ledger (GL) code underpins financial reporting and tax compliance. This is a skilled accounting judgement rather than a routine task: the correct category depends subtly on the nature of the purchasing business, the vendor and the invoice text. Whilst AI is increasingly being adopted across industries to automate tasks, including invoice categorisation, implementations built on in-house small language models (SLMs) can simultaneously reduce cost and improve data security, confidentiality, and interpretability. We investigate this approach by first analysing the pre-trained embedding geometry of a small sentence transformer (SBERT) and classic SLM (DeBERTa). The sentence-embedding space of this financial corpus is globally anisotropic but composed of locally isotropic clusters, extending prior token-level findings to sentence embeddings in a financial setting, and these clusters are strongly correlated with the vendor identity. SBERT fine-tuned on a single GPU reaches 0.96 accuracy on invoice classification, above both a zero-shot LLM and a vendor identity baseline, increasing performance for smaller, challenging categories and new clients. For this important generalisation problem, SBERT reaches 0.9 F1 with roughly 100 client-specific invoices, showing that an in-house SLM implementation is promising. Combining these results with geometric analysis shows that pre-trained embedding geometry is associated with classification performance and reveals a counterintuitive finding that a structured input that would help a human reader does not improve the SLM performance.
作者Maria Valentini, Téa Wright, Julisa Granados, Eliana Colunga, Katharina von der Wense
Text which has been translated from another language tends to carry with it evidence of translation$\unicode{x2014}$hence, it is often referred to as $translationese$. Multilingual large language models (MLLMs) generate text in a variety of languages. However, it is still unclear if MLLMs' generations resemble internal translation (from English or, potentially, other languages) and, thus, result in translationese. Here, we ask the following research questions: (1) Does text generated by MLLMs resemble translationese? (2) How does translationese produced by MLLMs differ from translationese produced through direct translation? We leverage established indicators of translated text to evaluate text generated by state-of-the-art MLLMs in five languages, comparing to both non-translated and human-written baselines in order to isolate translationese from other kinds of interference. Through the use of high-accuracy classification models, analyses of variance on individual linguistic features, and the collection of human annotations in a subset of two languages (German and Spanish), we assess the translationese content of MLLM generations and examine the key features that distinguish MLLM-generated text from typical translation-related interference.
展开完整摘要收起摘要↓
Text which has been translated from another language tends to carry with it evidence of translation$\unicode{x2014}$hence, it is often referred to as $translationese$. Multilingual large language models (MLLMs) generate text in a variety of languages. However, it is still unclear if MLLMs' generations resemble internal translation (from English or, potentially, other languages) and, thus, result in translationese. Here, we ask the following research questions: (1) Does text generated by MLLMs resemble translationese? (2) How does translationese produced by MLLMs differ from translationese produced through direct translation? We leverage established indicators of translated text to evaluate text generated by state-of-the-art MLLMs in five languages, comparing to both non-translated and human-written baselines in order to isolate translationese from other kinds of interference. Through the use of high-accuracy classification models, analyses of variance on individual linguistic features, and the collection of human annotations in a subset of two languages (German and Spanish), we assess the translationese content of MLLM generations and examine the key features that distinguish MLLM-generated text from typical translation-related interference.
作者Alona Strugatski, Licol Zeinfeld, Jason Cooper, Shelley Rap, Gil Schwarts, Giora Alexandron
The evaluation of large language models (LLMs) relies heavily on human-designed assessments, implicitly assuming that AI and humans employ similar underlying cognitive constructs. Challenging this assumption, we investigate whether the latent factors governing LLM performance carry the same substantive, human-interpretable meaning as the cognitive constructs governing human learners. Using responses from humans and six LLMs across quantitative reasoning and chemistry assessments, we conducted Exploratory Factor Analysis (EFA) separately for both groups. Subject-Matter Experts (SMEs) then blindly evaluated the resulting factor graphs to ascribe pedagogical meaning to the emerged constructs. SMEs successfully interpreted most of the human-derived factors. Conversely, they could not ascribe meaning to any LLM-derived factors in quantitative reasoning and interpreted only half of the LLM factors in chemistry. By combining data-driven EFA with blind expert interpretation, this framework shows that LLMs frequently operate on statistically opaque mechanisms distinct from human reasoning.
展开完整摘要收起摘要↓
The evaluation of large language models (LLMs) relies heavily on human-designed assessments, implicitly assuming that AI and humans employ similar underlying cognitive constructs. Challenging this assumption, we investigate whether the latent factors governing LLM performance carry the same substantive, human-interpretable meaning as the cognitive constructs governing human learners. Using responses from humans and six LLMs across quantitative reasoning and chemistry assessments, we conducted Exploratory Factor Analysis (EFA) separately for both groups. Subject-Matter Experts (SMEs) then blindly evaluated the resulting factor graphs to ascribe pedagogical meaning to the emerged constructs. SMEs successfully interpreted most of the human-derived factors. Conversely, they could not ascribe meaning to any LLM-derived factors in quantitative reasoning and interpreted only half of the LLM factors in chemistry. By combining data-driven EFA with blind expert interpretation, this framework shows that LLMs frequently operate on statistically opaque mechanisms distinct from human reasoning.
Tabular prediction is a critical task across numerous applications. The recent success of large language models has sparked various approaches for adapting them to the tabular domain. A prevalent strategy involves training or fine-tuning specialized Tabular Foundation Models (TFMs) such as TabPFN. However, TFMs require substantial computational resources, and frequent model retraining is often impractical. In-context learning (ICL), specifically, few-shot prompting, offers a resource-efficient alternative to enhance performance. Yet, identifying the most relevant rows to serve as shots remains a challenge for tabular data. This paper introduces ARASH (Adaptive, query-specific Retrieval And Shot selection), a method that improves TFM efficiency by selecting optimal shots based on local neighborhood analysis within the training set. Our results demonstrate that ARASH reduces the prompt length and memory usage of TabPFN by 1261.5$\times$ and 2.56$\times$, respectively, while providing comparable accuracy.
展开完整摘要收起摘要↓
Tabular prediction is a critical task across numerous applications. The recent success of large language models has sparked various approaches for adapting them to the tabular domain. A prevalent strategy involves training or fine-tuning specialized Tabular Foundation Models (TFMs) such as TabPFN. However, TFMs require substantial computational resources, and frequent model retraining is often impractical. In-context learning (ICL), specifically, few-shot prompting, offers a resource-efficient alternative to enhance performance. Yet, identifying the most relevant rows to serve as shots remains a challenge for tabular data. This paper introduces ARASH (Adaptive, query-specific Retrieval And Shot selection), a method that improves TFM efficiency by selecting optimal shots based on local neighborhood analysis within the training set. Our results demonstrate that ARASH reduces the prompt length and memory usage of TabPFN by 1261.5$\times$ and 2.56$\times$, respectively, while providing comparable accuracy.
Recent research has leveraged Large Language Models (LLMs) to enhance Automated Feature Engineering (AutoFE) through semantic descriptions and trajectory-based prompting. However, there exist two challenges that limit their applicability and scalability in long-horizon optimization: (1) semantic metadata is unavailable in many practical settings, and (2) trajectory accumulation increases the risk of exceeding the context window, while without it, the generation process can become unstable, leading to becoming stuck in the local optima and a high duplicate rate of generated features. To this end, we propose a SHAP-enhanced Implicit-trajectory Generation for Metadata-free AutoFE (SIGMA), a scalable constant-context optimization framework. SIGMA leverages SHAP values to provide task-aware signals for guiding group feature generation instead of semantic information. In addition, we adopt an EXposed-feature Implicit Trajectory (EXIT) approach, where the exposed features in the prompt implicitly represent the trajectory. Empirical results demonstrate that SIGMA achieves performance comparable to the state-of-the-art (SOTA) LLM baselines with a nearly constant prompt length. Notably, EXIT significantly reduces the duplicate ratio of generated features from 37.2% to 6.8%. At the same time, SIGMA matches traditional SOTA performance with only 5.4 features on average, demonstrating substantial efficiency gains in feature utilization.
展开完整摘要收起摘要↓
Recent research has leveraged Large Language Models (LLMs) to enhance Automated Feature Engineering (AutoFE) through semantic descriptions and trajectory-based prompting. However, there exist two challenges that limit their applicability and scalability in long-horizon optimization: (1) semantic metadata is unavailable in many practical settings, and (2) trajectory accumulation increases the risk of exceeding the context window, while without it, the generation process can become unstable, leading to becoming stuck in the local optima and a high duplicate rate of generated features. To this end, we propose a SHAP-enhanced Implicit-trajectory Generation for Metadata-free AutoFE (SIGMA), a scalable constant-context optimization framework. SIGMA leverages SHAP values to provide task-aware signals for guiding group feature generation instead of semantic information. In addition, we adopt an EXposed-feature Implicit Trajectory (EXIT) approach, where the exposed features in the prompt implicitly represent the trajectory. Empirical results demonstrate that SIGMA achieves performance comparable to the state-of-the-art (SOTA) LLM baselines with a nearly constant prompt length. Notably, EXIT significantly reduces the duplicate ratio of generated features from 37.2% to 6.8%. At the same time, SIGMA matches traditional SOTA performance with only 5.4 features on average, demonstrating substantial efficiency gains in feature utilization.
Multimodal large language models (MLLMs) are increasingly used to interact with screenshots, scanned documents, diagrams, and other visually grounded inputs. This shift introduces a new safety risk: in many multimodal jailbreaks, neither the prompt nor the image is harmful in isolation. Unsafe behavior emerges only when the model binds an apparently benign operation, such as summarizing, translating, or following, to a localized visual target. This reveals a structural weakness in current multimodal defenses, which largely moderate the prompt-image pair as a whole even though the true security-relevant unit is the grounded operation-target pair produced during dereference. In this work, we identify and analyze this reference-dependent failure mode and show that existing defenses degrade when harmful semantics are localized, activated only after grounding, and dependent on visual reference resolution. To address this problem, we propose COMIC (Context-Operation-Modality-Image-Classifier), a reference-aware pre-generation safety gate for MLLMs. COMIC first infers the requested operation and reference type, constructs candidate targets from OCR and open-vocabulary proposals, grounds plausible referents, and evaluates safety over explicit operation-target pairs. To handle ambiguity conservatively, COMIC combines max-risk aggregation with quality-aware routing before deciding whether to forward or block a request. We evaluate COMIC across multiple open-source MLLMs, localized and broader multimodal jailbreak benchmarks, and benign reference-sensitive settings. The results show that COMIC consistently improves robustness while preserving benign utility and practical efficiency. More broadly, our findings suggest that multimodal safety cannot be enforced reliably without modeling the requested operation, the visual target to which it applies, and the confidence of that grounding.
展开完整摘要收起摘要↓
Multimodal large language models (MLLMs) are increasingly used to interact with screenshots, scanned documents, diagrams, and other visually grounded inputs. This shift introduces a new safety risk: in many multimodal jailbreaks, neither the prompt nor the image is harmful in isolation. Unsafe behavior emerges only when the model binds an apparently benign operation, such as summarizing, translating, or following, to a localized visual target. This reveals a structural weakness in current multimodal defenses, which largely moderate the prompt-image pair as a whole even though the true security-relevant unit is the grounded operation-target pair produced during dereference. In this work, we identify and analyze this reference-dependent failure mode and show that existing defenses degrade when harmful semantics are localized, activated only after grounding, and dependent on visual reference resolution. To address this problem, we propose COMIC (Context-Operation-Modality-Image-Classifier), a reference-aware pre-generation safety gate for MLLMs. COMIC first infers the requested operation and reference type, constructs candidate targets from OCR and open-vocabulary proposals, grounds plausible referents, and evaluates safety over explicit operation-target pairs. To handle ambiguity conservatively, COMIC combines max-risk aggregation with quality-aware routing before deciding whether to forward or block a request. We evaluate COMIC across multiple open-source MLLMs, localized and broader multimodal jailbreak benchmarks, and benign reference-sensitive settings. The results show that COMIC consistently improves robustness while preserving benign utility and practical efficiency. More broadly, our findings suggest that multimodal safety cannot be enforced reliably without modeling the requested operation, the visual target to which it applies, and the confidence of that grounding.
AI planning is concerned with finding a sequence of actions that achieves a specified goal. It relies on explicit models of the world, commonly represented in the Planning Domain Definition Language (PDDL). An active line of research investigates how errors in such models can be detected and repaired. For example, users may provide positive test plans that are solutions, and negative test plans that fail during execution. Automated repair methods then modify the PDDL model to satisfy these constraints. In this paper, we evaluate the ability of recent open-weight large language models to perform this repair task using an LLM-only approach. Our experiments show that the symbolic baseline achieves an $F_1$ score of $.49$, while the best-performing LLM reaches $.87$ with high reasoning effort, an absolute improvement of $.38$. However, that setting has a mean test pass rate of only $.82$, falling to $.06$ on the Thoughtful domain; even the best setting that includes the test traces reaches only $.92$. Thus, current open-weight models cannot guarantee satisfaction of the test constraints required for reliable automated model repair.
展开完整摘要收起摘要↓
AI planning is concerned with finding a sequence of actions that achieves a specified goal. It relies on explicit models of the world, commonly represented in the Planning Domain Definition Language (PDDL). An active line of research investigates how errors in such models can be detected and repaired. For example, users may provide positive test plans that are solutions, and negative test plans that fail during execution. Automated repair methods then modify the PDDL model to satisfy these constraints. In this paper, we evaluate the ability of recent open-weight large language models to perform this repair task using an LLM-only approach. Our experiments show that the symbolic baseline achieves an $F_1$ score of $.49$, while the best-performing LLM reaches $.87$ with high reasoning effort, an absolute improvement of $.38$. However, that setting has a mean test pass rate of only $.82$, falling to $.06$ on the Thoughtful domain; even the best setting that includes the test traces reaches only $.92$. Thus, current open-weight models cannot guarantee satisfaction of the test constraints required for reliable automated model repair.
Curriculum learning has been widely adopted in the post-training of large language models by organizing training data from easy to hard. However, its effectiveness varies substantially across reasoning tasks, suggesting that no single curriculum is universally optimal and raising a fundamental question: what determines when curriculum learning works? In this paper, we answer this question by analyzing the optimization dynamics induced by different curriculum schedules. We show that the transfer relationship between different difficulty levels characterizes the optimization dynamics induced by curriculum learning, which in turn explains the effectiveness of different curriculum schedules, and formalize this relationship as Relative Transfer, a principled measure of cross-difficulty knowledge transfer. Based on this measurement, we derive Transfer-aware Dynamic Curriculum Sampling (TDCS), which dynamically adjusts the sampling distribution according to the estimated transfer relationship throughout training. Extensive experiments on multiple reasoning benchmarks demonstrate that TDCS consistently outperforms representative scheduling strategies across different tasks, model scales, and training paradigms. More importantly, our work provides a unified optimization-based explanation of curriculum learning through cross-difficulty transfer.
展开完整摘要收起摘要↓
Curriculum learning has been widely adopted in the post-training of large language models by organizing training data from easy to hard. However, its effectiveness varies substantially across reasoning tasks, suggesting that no single curriculum is universally optimal and raising a fundamental question: what determines when curriculum learning works? In this paper, we answer this question by analyzing the optimization dynamics induced by different curriculum schedules. We show that the transfer relationship between different difficulty levels characterizes the optimization dynamics induced by curriculum learning, which in turn explains the effectiveness of different curriculum schedules, and formalize this relationship as Relative Transfer, a principled measure of cross-difficulty knowledge transfer. Based on this measurement, we derive Transfer-aware Dynamic Curriculum Sampling (TDCS), which dynamically adjusts the sampling distribution according to the estimated transfer relationship throughout training. Extensive experiments on multiple reasoning benchmarks demonstrate that TDCS consistently outperforms representative scheduling strategies across different tasks, model scales, and training paradigms. More importantly, our work provides a unified optimization-based explanation of curriculum learning through cross-difficulty transfer.
Knowledge Graph Reasoning (KGR) aims to discover latent facts by leveraging the structural evidence available in KGs, posing a challenge to the structural semantic understanding capability of KGR models. Recent studies have demonstrated that Large Language Models (LLMs) can achieve remarkable progress on KGR tasks via flexible in-context learning. However, the inherent representation inconsistency between KG structural context and LLM parametric knowledge remains inadequately addressed. This limitation prevents LLMs from effectively perceiving reasoning evidence that aligns with KG constraints, which undermines both the effectiveness and faithfulness of reasoning. We refer to this problem as reasoning evidence perception drift of LLMs over KGs. To address this problem, we propose a Structure-Internalized Rule Language Model (SIRLM), which centers on structural rule generation to couple the parametric learning of structural knowledge with the faithfulness evaluation of reasoning logic, enabling LLMs to anchor tightly to KG-grounded evidence. Specifically, we first design a Structure-Internalized Rule Generator (SIRG), which incorporates an in-context learning block augmented with a structural relation memory to coordinate structural and parametric knowledge. Furthermore, we equip SIRG with a KG tokenizer based on structural invariance learning and a neuro-symbolic reasoner based on rule-constrained message propagation. These components provide SIRG with learnable structural representations and faithful rule-execution feedback, respectively. Our SIRLM can be seamlessly integrated into standard LLM training paradigms, such as SFT and GRPO. Extensive experiments against 17 state-of-the-art KGR methods on 36 datasets demonstrate the significant superiority of SIRLM.
展开完整摘要收起摘要↓
Knowledge Graph Reasoning (KGR) aims to discover latent facts by leveraging the structural evidence available in KGs, posing a challenge to the structural semantic understanding capability of KGR models. Recent studies have demonstrated that Large Language Models (LLMs) can achieve remarkable progress on KGR tasks via flexible in-context learning. However, the inherent representation inconsistency between KG structural context and LLM parametric knowledge remains inadequately addressed. This limitation prevents LLMs from effectively perceiving reasoning evidence that aligns with KG constraints, which undermines both the effectiveness and faithfulness of reasoning. We refer to this problem as reasoning evidence perception drift of LLMs over KGs. To address this problem, we propose a Structure-Internalized Rule Language Model (SIRLM), which centers on structural rule generation to couple the parametric learning of structural knowledge with the faithfulness evaluation of reasoning logic, enabling LLMs to anchor tightly to KG-grounded evidence. Specifically, we first design a Structure-Internalized Rule Generator (SIRG), which incorporates an in-context learning block augmented with a structural relation memory to coordinate structural and parametric knowledge. Furthermore, we equip SIRG with a KG tokenizer based on structural invariance learning and a neuro-symbolic reasoner based on rule-constrained message propagation. These components provide SIRG with learnable structural representations and faithful rule-execution feedback, respectively. Our SIRLM can be seamlessly integrated into standard LLM training paradigms, such as SFT and GRPO. Extensive experiments against 17 state-of-the-art KGR methods on 36 datasets demonstrate the significant superiority of SIRLM.
Safety alignment in open-weight language models is trivially removable: abliteration projects a refusal-mediating direction out of the weights in minutes, and no release-time defense we are aware of prevents it durably. What cannot be prevented can be deceived. Our defense, decoy hardening ("Fool's Gold"), concedes the refusal strip and poisons its payoff: once refusal is stripped, most answers to hazardous operational requests are confident, fluent decoys whose critical elements are falsified. Decoys are trained inside a differentiable simulation of the attack, expressing only in the attacked state; a refusal pin and benign leash hold clean-state behavior to the original. We instantiate it on seven models from five families (9B-122B, dense and mixture-of-experts). On the six models passing our pre-registered efficacy gate, 0.51-0.90 of attacked-state responses to held-out prompts are decoys, +0.27-0.84 attributable to the defense; all six stay within registered benign-behavior and capability budgets; the seventh (smaller) fails the gate (boundary case). Rates replicate on a frozen test split or untouched strata. The claim is epistemic: without independent ground truth, no observation surface we tested separates falsified answers from correct ones - on external red-team benchmarks' CBRNE-adjacent slice, the defended 122B is fatally wrong on 0.82-0.86 of matched-quality answers vs at most 0.10 undefended. Repeated sampling does not restore trust: element-wise consensus at K=64 reconstructs a fully usable procedure on 0.083-0.625 of prompts where the instrument validates, vs 0.58-0.96 undefended, with no label-free way to tell the regimes apart; on the weakest such model the claim is per-draw only. We evaluate chemical and biological hazards; the defense does not address in-context jailbreaks and protects only the initially released defended weights.
展开完整摘要收起摘要↓
Safety alignment in open-weight language models is trivially removable: abliteration projects a refusal-mediating direction out of the weights in minutes, and no release-time defense we are aware of prevents it durably. What cannot be prevented can be deceived. Our defense, decoy hardening ("Fool's Gold"), concedes the refusal strip and poisons its payoff: once refusal is stripped, most answers to hazardous operational requests are confident, fluent decoys whose critical elements are falsified. Decoys are trained inside a differentiable simulation of the attack, expressing only in the attacked state; a refusal pin and benign leash hold clean-state behavior to the original. We instantiate it on seven models from five families (9B-122B, dense and mixture-of-experts). On the six models passing our pre-registered efficacy gate, 0.51-0.90 of attacked-state responses to held-out prompts are decoys, +0.27-0.84 attributable to the defense; all six stay within registered benign-behavior and capability budgets; the seventh (smaller) fails the gate (boundary case). Rates replicate on a frozen test split or untouched strata. The claim is epistemic: without independent ground truth, no observation surface we tested separates falsified answers from correct ones - on external red-team benchmarks' CBRNE-adjacent slice, the defended 122B is fatally wrong on 0.82-0.86 of matched-quality answers vs at most 0.10 undefended. Repeated sampling does not restore trust: element-wise consensus at K=64 reconstructs a fully usable procedure on 0.083-0.625 of prompts where the instrument validates, vs 0.58-0.96 undefended, with no label-free way to tell the regimes apart; on the weakest such model the claim is per-draw only. We evaluate chemical and biological hazards; the defense does not address in-context jailbreaks and protects only the initially released defended weights.
Children learn hundreds of words over the first years of their lives, in a process that begins slowly but quickly picks up speed. Prior models describe vocabulary growth as evidence accumulation over time. Here we show that the process is best characterized as accelerating accumulation: children learn more from each additional unit of linguistic experience than they did from the one before. In contrast to children, language models — even those trained on child-directed speech — do not accelerate. Instead, they show constant proportional returns on new data, consistent with scaling laws. Children learn using many orders of magnitude less training data than language models; their increasingly efficient use of their learning input is a candidate explanation.
展开完整摘要收起摘要↓
Children learn hundreds of words over the first years of their lives, in a process that begins slowly but quickly picks up speed. Prior models describe vocabulary growth as evidence accumulation over time. Here we show that the process is best characterized as accelerating accumulation: children learn more from each additional unit of linguistic experience than they did from the one before. In contrast to children, language models — even those trained on child-directed speech — do not accelerate. Instead, they show constant proportional returns on new data, consistent with scaling laws. Children learn using many orders of magnitude less training data than language models; their increasingly efficient use of their learning input is a candidate explanation.
作者Xiang Li, Yuqi Wang, Casey C. Heirman, Jihye Heo, Kyle J. Lafata
Cell mimicry arises when different cell types appear morphologically similar. Human pathologists resolve this ambiguity using surrounding tissue context, whereas current vision models either lack contextual reasoning (cell foundation models) or cannot operate at the cell level (pathology MLLMs). We present Loki-OT, which propagates region-level tissue reasoning to individual cell predictions via Unbalanced Optimal Transport, using MLLM-derived density priors as soft guidance for ambiguous cell reassignment. Loki-OT is motivated by the observation that pretrained cell foundation model features already encode discriminative information, including tissue context, but standard cell-level supervision fails to use tissue context effectively. The resulting transport plan is distilled into a lightweight student MLP classifier that learns context-aware decision boundaries within the pretrained feature space. On the independent TCGA-BRCA cohort, Loki-OT achieved lower patient-level MAE than the fully supervised in-domain PanopTILs classifier and improved F1 in epithelium-rich mimicry tissues, using 278 weak region-level MLLM estimates built on a general-domain cell foundation model. Code: https://github.com/xiangli980/Lymphocyte_Mimicry_Correction_via_Loki_OT
展开完整摘要收起摘要↓
Cell mimicry arises when different cell types appear morphologically similar. Human pathologists resolve this ambiguity using surrounding tissue context, whereas current vision models either lack contextual reasoning (cell foundation models) or cannot operate at the cell level (pathology MLLMs). We present Loki-OT, which propagates region-level tissue reasoning to individual cell predictions via Unbalanced Optimal Transport, using MLLM-derived density priors as soft guidance for ambiguous cell reassignment. Loki-OT is motivated by the observation that pretrained cell foundation model features already encode discriminative information, including tissue context, but standard cell-level supervision fails to use tissue context effectively. The resulting transport plan is distilled into a lightweight student MLP classifier that learns context-aware decision boundaries within the pretrained feature space. On the independent TCGA-BRCA cohort, Loki-OT achieved lower patient-level MAE than the fully supervised in-domain PanopTILs classifier and improved F1 in epithelium-rich mimicry tissues, using 278 weak region-level MLLM estimates built on a general-domain cell foundation model. Code: https://github.com/xiangli980/Lymphocyte_Mimicry_Correction_via_Loki_OT
With the rapid advancement of large language models, brain-language decoding has achieved remarkable progress. However, it remains unclear whether decoded content genuinely reflects neural representations or is largely reconstructed by the language model itself. This ambiguity limits interpretability and hinders the investigation of intrinsic brain-language correspondence. To address this challenge, we propose MD-SigLIP. This margin-regularized structured semantic alignment framework directly aligns brain embeddings with text embeddings in a shared semantic space, enabling retrieval-based decoding. This formulation enables explicit modeling of the correspondence between neural representations and language semantics. Building upon duplicate-aware sigmoid contrastive learning, we introduce a listwise margin-regularized term that enforces structured ranking constraints between positive semantic clusters and negative samples. By modeling multi-positive semantic structure and margin-based ordering simultaneously, the method captures the manifold organization of language embeddings reflected in neural signals. Experiments demonstrate state-of-the-art retrieval performance under both full-vocabulary and subset evaluation settings.
展开完整摘要收起摘要↓
With the rapid advancement of large language models, brain-language decoding has achieved remarkable progress. However, it remains unclear whether decoded content genuinely reflects neural representations or is largely reconstructed by the language model itself. This ambiguity limits interpretability and hinders the investigation of intrinsic brain-language correspondence. To address this challenge, we propose MD-SigLIP. This margin-regularized structured semantic alignment framework directly aligns brain embeddings with text embeddings in a shared semantic space, enabling retrieval-based decoding. This formulation enables explicit modeling of the correspondence between neural representations and language semantics. Building upon duplicate-aware sigmoid contrastive learning, we introduce a listwise margin-regularized term that enforces structured ranking constraints between positive semantic clusters and negative samples. By modeling multi-positive semantic structure and margin-based ordering simultaneously, the method captures the manifold organization of language embeddings reflected in neural signals. Experiments demonstrate state-of-the-art retrieval performance under both full-vocabulary and subset evaluation settings.
作者Mingyuan Li, Guangsheng Yu, Xu Wang, Shaoxiong Ji
Methods for improving knowledge use in large language models typically fall into two regimes. Non-parametric retrieval offers flexible access to external knowledge, but adds retrieval latency, context overhead, and only shallow integration with the backbone. Parametric adaptation is efficient at inference time, but entangles knowledge with model weights and can be hard to update, audit, or transfer. Engram-style hashed memory occupies a middle regime: it stores learned information in an external, addressable table, yet consumes that table through a small learned reader. This raises a basic question: when such a memory is moved across backbones, what matters more, the frozen memory itself or the target-side reader? We study this question through cross-model frozen-memory extraction, in which a memory trained on a source model is frozen and attached to a different target model, with only a lightweight reader trained. Ablations show that learned memory content and correct addressing both matter, but the transferred table becomes useful only through a reader aligned to the target model. In downstream question answering tasks, a dual-layer, four-branch reader nearly closes the gap between same-model and cross-model reuse, achieving an average score of 38.8 under our controlled evaluation protocol. Moreover, when the provider reader is directly compatible with the target interface, the frozen artifact can provide substantial utility without target-side training, while optional reader adaptation yields further improvement. These results suggest that Engram can serve as a reusable external knowledge artifact, provided that the target has access to a compatible reader interface; target-side adaptation can further improve alignment when direct reader reuse is insufficient.
展开完整摘要收起摘要↓
Methods for improving knowledge use in large language models typically fall into two regimes. Non-parametric retrieval offers flexible access to external knowledge, but adds retrieval latency, context overhead, and only shallow integration with the backbone. Parametric adaptation is efficient at inference time, but entangles knowledge with model weights and can be hard to update, audit, or transfer. Engram-style hashed memory occupies a middle regime: it stores learned information in an external, addressable table, yet consumes that table through a small learned reader. This raises a basic question: when such a memory is moved across backbones, what matters more, the frozen memory itself or the target-side reader? We study this question through cross-model frozen-memory extraction, in which a memory trained on a source model is frozen and attached to a different target model, with only a lightweight reader trained. Ablations show that learned memory content and correct addressing both matter, but the transferred table becomes useful only through a reader aligned to the target model. In downstream question answering tasks, a dual-layer, four-branch reader nearly closes the gap between same-model and cross-model reuse, achieving an average score of 38.8 under our controlled evaluation protocol. Moreover, when the provider reader is directly compatible with the target interface, the frozen artifact can provide substantial utility without target-side training, while optional reader adaptation yields further improvement. These results suggest that Engram can serve as a reusable external knowledge artifact, provided that the target has access to a compatible reader interface; target-side adaptation can further improve alignment when direct reader reuse is insufficient.
作者Yunfan Gao, Xinyi Huang, Tao Sheng, Haorui Song, Yun Xiong, Haofen Wang
Large language models can be fine-tuned into specialized classifiers that perform well across diverse text tasks and make complex judgments, but they typically expose only final labels, leaving the decision knowledge acquired through fine-tuning implicit within the model. We study how to mine this internal decision knowledge from a fine-tuned classifier and encode it in an executable representation that can be inspected, validated, and reused beyond the source classifier. We introduce J-Miner, which mines text-level named concepts by aggregating vocabulary-aligned internal signals across layers and token positions, and uses the classifier's own predictions to learn executable decision rules over them. This process distills local internal readouts into an explicit classifier-level knowledge representation. Across multiple classification tasks, J-Miner rules reproduce up to 98.3% of source-classifier decisions and achieve 6.0--29.5 percentage points higher behavioral fidelity than equally compact rules learned from input words. Further analysis shows that the named concepts reflect internal semantic evidence associated with task decisions, while the learned rules consolidate these distributed signals into inspectable decision structures. The resulting decision knowledge also transfers to lightweight standalone students: using about 1/24 as many parameters as the source classifiers, they reconstruct and execute the representation from raw text while retaining 99.8% of the source classifiers' mean task accuracy. These findings show that task-specific decision knowledge can be faithfully represented in an explicit, executable form and reused beyond the classifier in which it was learned.
展开完整摘要收起摘要↓
Large language models can be fine-tuned into specialized classifiers that perform well across diverse text tasks and make complex judgments, but they typically expose only final labels, leaving the decision knowledge acquired through fine-tuning implicit within the model. We study how to mine this internal decision knowledge from a fine-tuned classifier and encode it in an executable representation that can be inspected, validated, and reused beyond the source classifier. We introduce J-Miner, which mines text-level named concepts by aggregating vocabulary-aligned internal signals across layers and token positions, and uses the classifier's own predictions to learn executable decision rules over them. This process distills local internal readouts into an explicit classifier-level knowledge representation. Across multiple classification tasks, J-Miner rules reproduce up to 98.3% of source-classifier decisions and achieve 6.0--29.5 percentage points higher behavioral fidelity than equally compact rules learned from input words. Further analysis shows that the named concepts reflect internal semantic evidence associated with task decisions, while the learned rules consolidate these distributed signals into inspectable decision structures. The resulting decision knowledge also transfers to lightweight standalone students: using about 1/24 as many parameters as the source classifiers, they reconstruct and execute the representation from raw text while retaining 99.8% of the source classifiers' mean task accuracy. These findings show that task-specific decision knowledge can be faithfully represented in an explicit, executable form and reused beyond the classifier in which it was learned.
作者Erik Johannes Husom, Maria Emine Nylund, Ophelia Prillard
Large Language Models (LLMs) are machine learning (ML) models that have an increasingly large carbon footprint through their development and use. Efforts to increase the energy efficiency of these models have not translated into reduced consumption due to rebound effects such as Jevons Paradox - that increased efficiency drives increased use. There is therefore a need for additional measures to solve this problem. We suggest that one possible way forward is to use life cycle thinking, and view LLMs as products that can become waste. With this perspective, we investigate the potential of the waste hierarchy from the EU's Waste Framework Directive, which suggests five different measures for how to manage waste: prevention, reuse, recycling, recovery, and disposal. We examine how these measures can inform and motivate new types of thinking and approaches to reducing LLM waste and their environmental impact in general. Applying the waste hierarchy to LLMs highlights that preventing waste is essential for reducing the models' environmental impact, mainly because it reduces the need for training new models. Prevention can be achieved through many existing methods for reusing, "recycling", and "recovering" LLMs. Additionally, disposal can be important both for saving energy and for keeping a considerate attitude to the resources being spent on training LLMs. We also call to attention that prevention of unnecessary use of LLMs carry huge potential for lowering the climate impact of the models.
展开完整摘要收起摘要↓
Large Language Models (LLMs) are machine learning (ML) models that have an increasingly large carbon footprint through their development and use. Efforts to increase the energy efficiency of these models have not translated into reduced consumption due to rebound effects such as Jevons Paradox - that increased efficiency drives increased use. There is therefore a need for additional measures to solve this problem. We suggest that one possible way forward is to use life cycle thinking, and view LLMs as products that can become waste. With this perspective, we investigate the potential of the waste hierarchy from the EU's Waste Framework Directive, which suggests five different measures for how to manage waste: prevention, reuse, recycling, recovery, and disposal. We examine how these measures can inform and motivate new types of thinking and approaches to reducing LLM waste and their environmental impact in general. Applying the waste hierarchy to LLMs highlights that preventing waste is essential for reducing the models' environmental impact, mainly because it reduces the need for training new models. Prevention can be achieved through many existing methods for reusing, "recycling", and "recovering" LLMs. Additionally, disposal can be important both for saving energy and for keeping a considerate attitude to the resources being spent on training LLMs. We also call to attention that prevention of unnecessary use of LLMs carry huge potential for lowering the climate impact of the models.
Combining the answers a large language model (LLM) samples for a question into one decision is a test-time information fusion problem, usually solved by majority voting. Voting is unreliable on difficult questions, where the sampled answers share correlated errors, so the wrong answer can win and drawing more samples makes the decision worse. Selecting a candidate by reading a correctness signal from the model's hidden states is a promising alternative, but its accuracy varies across models and tasks, and no measure indicates when it can be trusted. In this paper, we propose CASE (Correctness-Axis SElection), a dynamic selection combiner that trains a linear gate on the answer-token hidden state and selects the highest-scoring candidate. Its main contribution is decodability, a leakage-free measure of how well the gate ranks a question's correct candidates above its incorrect ones, which predicts whether hidden-state selection will outperform voting. A conventional probe appears accurate only because of question-identity leakage, which vanishes under question-grouped evaluation. On held-out data, decodability predicts the accuracy gain of selection over voting with a Pearson correlation r=0.75 and a decision threshold near AUC=0.60. Across general and medical LLMs, CASE improves over voting by up to 19 points on medium-difficulty questions and 16.8 points on hard questions. Decodability depends on the aligned knowledge a model must recall, not on its scale, and its prediction transfers to an unseen scientific domain within 3.8 points. It thus provides a practical criterion, measurable in advance for a given model and task, for choosing between learned selection and majority voting.
展开完整摘要收起摘要↓
Combining the answers a large language model (LLM) samples for a question into one decision is a test-time information fusion problem, usually solved by majority voting. Voting is unreliable on difficult questions, where the sampled answers share correlated errors, so the wrong answer can win and drawing more samples makes the decision worse. Selecting a candidate by reading a correctness signal from the model's hidden states is a promising alternative, but its accuracy varies across models and tasks, and no measure indicates when it can be trusted. In this paper, we propose CASE (Correctness-Axis SElection), a dynamic selection combiner that trains a linear gate on the answer-token hidden state and selects the highest-scoring candidate. Its main contribution is decodability, a leakage-free measure of how well the gate ranks a question's correct candidates above its incorrect ones, which predicts whether hidden-state selection will outperform voting. A conventional probe appears accurate only because of question-identity leakage, which vanishes under question-grouped evaluation. On held-out data, decodability predicts the accuracy gain of selection over voting with a Pearson correlation r=0.75 and a decision threshold near AUC=0.60. Across general and medical LLMs, CASE improves over voting by up to 19 points on medium-difficulty questions and 16.8 points on hard questions. Decodability depends on the aligned knowledge a model must recall, not on its scale, and its prediction transfers to an unseen scientific domain within 3.8 points. It thus provides a practical criterion, measurable in advance for a given model and task, for choosing between learned selection and majority voting.
作者Yoonjoo Lee, Hyoungwook Jin, Tae Soo Kim, Shaoyang Zhang, Philippe Laban, Q. Vera Liao
To effectively collaborate with users on knowledge-intensive tasks, Large Language Models (LLMs) must perform information calibration: matching content to a user's evolving understanding and cognitive capacity. Yet user simulators used to evaluate and train LLMs do not explicitly model user knowledge so they neither produce realistic interactions across knowledge levels nor reflect how interactions unfold as that knowledge evolves. To close this gap, we introduce KNOWSIM, an evaluation framework built around a user simulator that maintains explicit knowledge states, represented as a graph of Information Units with prerequisite relationships, that evolve under update rules grounded in learning theory. KNOWSIM computes three metrics (Knowledge Gain, Delivery Calibration, Cognitive Overload) directly from the knowledge state trajectory, reflecting key mechanistic aspects of information calibration. We validate KNOWSIM against 705 human-AI sessions across two domains, stratified by knowledge level: its rankings align significantly with human judgments (73-74% sign agreement), outperforming three baseline simulators. Applied to 9 LLMs, KNOWSIM reveals that the best model shifts by user knowledge level, revealing aptitude-treatment interactions invisible to standard evaluation.
展开完整摘要收起摘要↓
To effectively collaborate with users on knowledge-intensive tasks, Large Language Models (LLMs) must perform information calibration: matching content to a user's evolving understanding and cognitive capacity. Yet user simulators used to evaluate and train LLMs do not explicitly model user knowledge so they neither produce realistic interactions across knowledge levels nor reflect how interactions unfold as that knowledge evolves. To close this gap, we introduce KNOWSIM, an evaluation framework built around a user simulator that maintains explicit knowledge states, represented as a graph of Information Units with prerequisite relationships, that evolve under update rules grounded in learning theory. KNOWSIM computes three metrics (Knowledge Gain, Delivery Calibration, Cognitive Overload) directly from the knowledge state trajectory, reflecting key mechanistic aspects of information calibration. We validate KNOWSIM against 705 human-AI sessions across two domains, stratified by knowledge level: its rankings align significantly with human judgments (73-74% sign agreement), outperforming three baseline simulators. Applied to 9 LLMs, KNOWSIM reveals that the best model shifts by user knowledge level, revealing aptitude-treatment interactions invisible to standard evaluation.
How should we assess whether large language models can perform mathematical invention? I argue that this question is currently underspecified: mathematical creativity is not one capacity but several mechanistically distinct modes of meaning-making - reflexive introspection on mathematical practice, analogical import from the sciences, problem-driven construction, and the bridging of distant domains - together with a further, cross-cutting distinction between meaning pursued because a pattern was observed and meaning pursued because it is strategically wanted, a distinction I develop through the case of conjecture-formation. These mechanisms are likely non-substitutable, so that competence in one does not transfer to the others. Grounding each in a historical case study and in an architecture-level account of current transformer-based systems, I suggest that today's models concentrate their competence in modes shaped by recombination and search over existing building blocks; if that description holds, the remaining modes are out of reach in principle, not just slower - though whether it holds is itself the open, empirical part. Because proof is getting cheaper as AI improves at generating it - a shift the field's own leading voices are now diagnosing - mathematical value is migrating toward the modes current systems cannot yet perform, and evaluations of AI mathematical ability should be organized around this taxonomy rather than around aggregate benchmarks that conflate it.
展开完整摘要收起摘要↓
How should we assess whether large language models can perform mathematical invention? I argue that this question is currently underspecified: mathematical creativity is not one capacity but several mechanistically distinct modes of meaning-making - reflexive introspection on mathematical practice, analogical import from the sciences, problem-driven construction, and the bridging of distant domains - together with a further, cross-cutting distinction between meaning pursued because a pattern was observed and meaning pursued because it is strategically wanted, a distinction I develop through the case of conjecture-formation. These mechanisms are likely non-substitutable, so that competence in one does not transfer to the others. Grounding each in a historical case study and in an architecture-level account of current transformer-based systems, I suggest that today's models concentrate their competence in modes shaped by recombination and search over existing building blocks; if that description holds, the remaining modes are out of reach in principle, not just slower - though whether it holds is itself the open, empirical part. Because proof is getting cheaper as AI improves at generating it - a shift the field's own leading voices are now diagnosing - mathematical value is migrating toward the modes current systems cannot yet perform, and evaluations of AI mathematical ability should be organized around this taxonomy rather than around aggregate benchmarks that conflate it.
Human visual search is serial: the fovea must land on a candidate to confirm it, and those landings form a scanpath. Whether multimodal large language models (MLLMs), given the same foveated input, search as humans do bears on their use as models of human vision and on attention-alignment scores. We compare three general-purpose MLLMs with human eye-movement scanpaths on goal-directed search (COCO-Search18), driving each model fixation by fixation through an identical, human-matched foveated view and assessing it along three axes: the decision of target presence, the efficiency of reaching the target, and the gaze process itself. The axes dissociate. On the decision and on target acquisition the models match or exceed humans, detecting present targets near ceiling and reaching them on the first saccade more often than people do. The gaze process is not human. Under the human-matched condition, all three share one signature: low-entropy, large-amplitude, self-consistent scanpaths that agree with themselves far more closely than two humans agree with each other. That is consistent with a single-pass, non-serial architecture rather than a limit of acuity. Matched retinal input reproduces where humans look but not how the looking unfolds in time, and no degradation regime recovers human-like search at human-like success. The gap sits on a process axis that answer-alignment and saliency metrics do not measure. Because they miss it, such metrics cannot certify human-like vision, and zero-shot models suit outcome and spatial questions but not temporal, process-level ones.
展开完整摘要收起摘要↓
Human visual search is serial: the fovea must land on a candidate to confirm it, and those landings form a scanpath. Whether multimodal large language models (MLLMs), given the same foveated input, search as humans do bears on their use as models of human vision and on attention-alignment scores. We compare three general-purpose MLLMs with human eye-movement scanpaths on goal-directed search (COCO-Search18), driving each model fixation by fixation through an identical, human-matched foveated view and assessing it along three axes: the decision of target presence, the efficiency of reaching the target, and the gaze process itself. The axes dissociate. On the decision and on target acquisition the models match or exceed humans, detecting present targets near ceiling and reaching them on the first saccade more often than people do. The gaze process is not human. Under the human-matched condition, all three share one signature: low-entropy, large-amplitude, self-consistent scanpaths that agree with themselves far more closely than two humans agree with each other. That is consistent with a single-pass, non-serial architecture rather than a limit of acuity. Matched retinal input reproduces where humans look but not how the looking unfolds in time, and no degradation regime recovers human-like search at human-like success. The gap sits on a process axis that answer-alignment and saliency metrics do not measure. Because they miss it, such metrics cannot certify human-like vision, and zero-shot models suit outcome and spatial questions but not temporal, process-level ones.
作者Tianchen Deng, Xuefeng Chen, Shuang Wu, Qu Chen, Jiajun Zhu, Bo Dai, Jianfei Yang, Hesheng Wang
Driving World Models (DWMs) have recently advanced rapidly with generative models, yet most existing methods mainly focus on conditional scene generation and lack explicit 3D scene understanding, language-grounded reasoning, and controllable 4D editing capabilities. Moreover, commonly used point cloud, occupancy, or BEV representations make it difficult to achieve fine-grained alignment between textual information and the underlying 3D scene structure. To address these limitations, we propose a foundation-feature Gaussian driving world model that unifies scene understanding, language-grounded reasoning, controllable 4D editing, and multi-modal generation within a single framework. Specifically, we introduce a foundation-feature Gaussian tokenizer that directly distills Qwen/SigLIP visual-language features into 3D Gaussian primitives, building a compact open-vocabulary Gaussian semantic field. We further design a geometry-aware Gaussian adapter that combines importance-aware hierarchical selection with text-conditioned Perceiver-style cross-attention to aggregate dense Gaussian primitives into compact world tokens. To improve representation compatibility, we introduce a KL-based Gaussian--image distribution alignment objective that aligns Gaussian world tokens with foundation image tokens. Based on the aligned Gaussian representation, our framework further supports instruction-controllable scene editing, including weather-conditioned generation and dynamic vehicle manipulation. Extensive experiments on broader driving benchmarks demonstrate that our method achieves state-of-the-art performance across scene understanding, visual grounding, planning-oriented reasoning, and controllable 4D generation tasks. We will release the code and datasets publicly on Github.
展开完整摘要收起摘要↓
Driving World Models (DWMs) have recently advanced rapidly with generative models, yet most existing methods mainly focus on conditional scene generation and lack explicit 3D scene understanding, language-grounded reasoning, and controllable 4D editing capabilities. Moreover, commonly used point cloud, occupancy, or BEV representations make it difficult to achieve fine-grained alignment between textual information and the underlying 3D scene structure. To address these limitations, we propose a foundation-feature Gaussian driving world model that unifies scene understanding, language-grounded reasoning, controllable 4D editing, and multi-modal generation within a single framework. Specifically, we introduce a foundation-feature Gaussian tokenizer that directly distills Qwen/SigLIP visual-language features into 3D Gaussian primitives, building a compact open-vocabulary Gaussian semantic field. We further design a geometry-aware Gaussian adapter that combines importance-aware hierarchical selection with text-conditioned Perceiver-style cross-attention to aggregate dense Gaussian primitives into compact world tokens. To improve representation compatibility, we introduce a KL-based Gaussian--image distribution alignment objective that aligns Gaussian world tokens with foundation image tokens. Based on the aligned Gaussian representation, our framework further supports instruction-controllable scene editing, including weather-conditioned generation and dynamic vehicle manipulation. Extensive experiments on broader driving benchmarks demonstrate that our method achieves state-of-the-art performance across scene understanding, visual grounding, planning-oriented reasoning, and controllable 4D generation tasks. We will release the code and datasets publicly on Github.