Diffusion language models (dLLMs) offer an alternative to autoregressive (AR) language modeling, yet the scaling behavior of Mixture-of-Experts (MoE) dLLMs remains poorly understood. We systematically characterize how optimization hyperparameters, compute allocation, and architecture scale for MoE dLLMs, identifying quantitative differences from scaling trends previously reported for AR models. Specifically, for optimization, the optimal nominal batch size grows faster, while the optimal learning rate decays more rapidly with compute. For model--data allocation, IsoFLOP analysis reveals a slight data-side tilt: the optimal token budget grows faster than activated model-side computation. For MoE architecture, larger scales increasingly favor larger expert pools at fixed activated capacity, while moderate expert granularity remains consistently effective and the preferred fraction of activated capacity assigned to shared experts remains stable across scales. Guided by these findings, we train LLaDA MoE v2, a 30B-A3B dLLM, from scratch on 23.5T tokens. With approximately 65% as many pretraining tokens as Qwen3, LLaDA MoE v2 approaches Qwen3 on several knowledge, reasoning, and coding benchmarks. After supervised fine-tuning alone, it outperforms SDAR Chat on seven of eight reasoning and coding benchmarks and remains close to Qwen3 on several tasks. These results establish practical scaling laws and design principles for MoE dLLMs.
展开完整摘要收起摘要↓
Diffusion language models (dLLMs) offer an alternative to autoregressive (AR) language modeling, yet the scaling behavior of Mixture-of-Experts (MoE) dLLMs remains poorly understood. We systematically characterize how optimization hyperparameters, compute allocation, and architecture scale for MoE dLLMs, identifying quantitative differences from scaling trends previously reported for AR models. Specifically, for optimization, the optimal nominal batch size grows faster, while the optimal learning rate decays more rapidly with compute. For model--data allocation, IsoFLOP analysis reveals a slight data-side tilt: the optimal token budget grows faster than activated model-side computation. For MoE architecture, larger scales increasingly favor larger expert pools at fixed activated capacity, while moderate expert granularity remains consistently effective and the preferred fraction of activated capacity assigned to shared experts remains stable across scales. Guided by these findings, we train LLaDA MoE v2, a 30B-A3B dLLM, from scratch on 23.5T tokens. With approximately 65% as many pretraining tokens as Qwen3, LLaDA MoE v2 approaches Qwen3 on several knowledge, reasoning, and coding benchmarks. After supervised fine-tuning alone, it outperforms SDAR Chat on seven of eight reasoning and coding benchmarks and remains close to Qwen3 on several tasks. These results establish practical scaling laws and design principles for MoE dLLMs.
作者Mohsen Hariri, Weicong Chen, Nahal Shahini, Vikash Singh, Kai Ye, Amirhossein Samandar, Debargha Ganguly, Sreehari Sankar, Yanyan Zhang, Shouren Wang, Jerry Peng, Biyao Zhang, Michael Hinczewski, Vipin Chaudhary
Large language models can solve substantially harder reasoning problems with more inference-time compute. The term "test-time scaling," however, now covers diverse inference algorithms that extend deliberation along a single trajectory, sample completed candidates and aggregate them through voting or verification, or search over unfinished partial states. These algorithms differ in their statistical structure, compute accounting, and failure modes. Treating these procedures as interchangeable under a single scalar "budget," or reporting accuracy without the inference protocol that produced it, makes results difficult to compare across studies. We develop a systematic account of test-time scaling along three axes. First, we formalize test-time scaling as budgeted inference over the implicit prefix tree of an autoregressive model and distinguish three structural regimes: single-trajectory sequential scaling, leaf-level scaling with terminal reduction, and prefix-level scaling. Second, we treat the evaluated object as the entire inference system and develop evaluation principles that separate end-to-end system performance from candidate-bank diagnostics. We introduce an evaluation profile whose coordinates and simple functionals recover or bound common repeated-sampling metrics, and prescribe protocol-matched reporting of compute and uncertainty. Third, we specify reproducibility requirements for inference protocols, distinguishing exact replay from distributional reproducibility and identifying the artifacts needed to support each. We also organize the open-weight reasoning ecosystem by model-side and interface mechanisms, apply these principles to broad-knowledge, symbolic-reasoning, and competition-mathematics benchmarks, and assemble over 2 billion full reasoning traces for release with progressively richer verifier and token-level signals.
展开完整摘要收起摘要↓
Large language models can solve substantially harder reasoning problems with more inference-time compute. The term "test-time scaling," however, now covers diverse inference algorithms that extend deliberation along a single trajectory, sample completed candidates and aggregate them through voting or verification, or search over unfinished partial states. These algorithms differ in their statistical structure, compute accounting, and failure modes. Treating these procedures as interchangeable under a single scalar "budget," or reporting accuracy without the inference protocol that produced it, makes results difficult to compare across studies. We develop a systematic account of test-time scaling along three axes. First, we formalize test-time scaling as budgeted inference over the implicit prefix tree of an autoregressive model and distinguish three structural regimes: single-trajectory sequential scaling, leaf-level scaling with terminal reduction, and prefix-level scaling. Second, we treat the evaluated object as the entire inference system and develop evaluation principles that separate end-to-end system performance from candidate-bank diagnostics. We introduce an evaluation profile whose coordinates and simple functionals recover or bound common repeated-sampling metrics, and prescribe protocol-matched reporting of compute and uncertainty. Third, we specify reproducibility requirements for inference protocols, distinguishing exact replay from distributional reproducibility and identifying the artifacts needed to support each. We also organize the open-weight reasoning ecosystem by model-side and interface mechanisms, apply these principles to broad-knowledge, symbolic-reasoning, and competition-mathematics benchmarks, and assemble over 2 billion full reasoning traces for release with progressively richer verifier and token-level signals.
Large language models have improved substantially on single-shot reasoning tasks, but their performance in sequential decision-making is less well understood. We study this on fully-observable two-player zero-sum games, which provide ground-truth evaluation: outcomes are determined by the rules, and optimality of individual moves can be computed or approximated, without relying on a judge model. Across model tiers, LLMs play suboptimally in simple games such as tic-tac-toe or Connect Four, and lose to MCTS opponents. Obfuscations that preserve the game tree but rewrite its surface form leave performance largely unchanged, indicating the gap is not fully explained by recall of memorized strategies. Motivated by this performance gap, we introduce an agentic framework enhanced with an experience memory designed for the sequential setting and addressing common challenges of sequential decision-making such as credit assignment. We show that post-game reflection and rule extraction yield measurable improvements on tic-tac-toe without modifying the model weights.
展开完整摘要收起摘要↓
Large language models have improved substantially on single-shot reasoning tasks, but their performance in sequential decision-making is less well understood. We study this on fully-observable two-player zero-sum games, which provide ground-truth evaluation: outcomes are determined by the rules, and optimality of individual moves can be computed or approximated, without relying on a judge model. Across model tiers, LLMs play suboptimally in simple games such as tic-tac-toe or Connect Four, and lose to MCTS opponents. Obfuscations that preserve the game tree but rewrite its surface form leave performance largely unchanged, indicating the gap is not fully explained by recall of memorized strategies. Motivated by this performance gap, we introduce an agentic framework enhanced with an experience memory designed for the sequential setting and addressing common challenges of sequential decision-making such as credit assignment. We show that post-game reflection and rule extraction yield measurable improvements on tic-tac-toe without modifying the model weights.
Automated Knowledge Base Construction (AKBC) is a core NLP task, and recent work proposes generating knowledge bases directly from large language models (LLMs), treating the model itself as the knowledge source. However, LLMs natively possess no representation of entities, leading to duplicate entries as well as conflations. We propose GPTKB 2.0, a methodology for constructing disambiguated KBs directly from LLMs. GPTKB 2.0 incorporates on-the-fly disambiguation of entities, relations and classes, and is meticulously designed to satisfy both scalability and disambiguation accuracy. We analyze the central design decisions and characterize the trade-offs between accuracy, scale, and cost. We execute GPTKB 2.0 at scale, obtaining a materialized KB containing over 1M disambiguated entities and 38.4M triples. This represents the first million-scale LLM-native KB with explicit internal canonicalization of entities, relations, and classes, a significant departure from prior Wikimedia-centric works. GPTKB 2.0 is available at https://gptkb.org/.
展开完整摘要收起摘要↓
Automated Knowledge Base Construction (AKBC) is a core NLP task, and recent work proposes generating knowledge bases directly from large language models (LLMs), treating the model itself as the knowledge source. However, LLMs natively possess no representation of entities, leading to duplicate entries as well as conflations. We propose GPTKB 2.0, a methodology for constructing disambiguated KBs directly from LLMs. GPTKB 2.0 incorporates on-the-fly disambiguation of entities, relations and classes, and is meticulously designed to satisfy both scalability and disambiguation accuracy. We analyze the central design decisions and characterize the trade-offs between accuracy, scale, and cost. We execute GPTKB 2.0 at scale, obtaining a materialized KB containing over 1M disambiguated entities and 38.4M triples. This represents the first million-scale LLM-native KB with explicit internal canonicalization of entities, relations, and classes, a significant departure from prior Wikimedia-centric works. GPTKB 2.0 is available at https://gptkb.org/.
On-policy training has emerged as a powerful post-training paradigm for improving the reasoning capabilities of large language models, and is often enhanced by golden trajectories from stronger expert models. However, when the expert fails on harder problems, existing trajectory-guided methods lose their main source of supervision, and these failed trajectories are typically discarded as negative samples. We argue that such failures, which we call Golden Negative Trajectories, can still provide valuable reasoning signals when treated not as demonstrations to imitate, but as flawed trajectories to reflect upon. We identify a Reflection Advantage: for hard problems, reflecting on a flawed trajectory can be easier and more effective than solving the problem directly from scratch. Motivated by this, we propose ReflectRL, a lightweight plug-and-play framework that learns from Golden Negative Trajectories during on-policy training. ReflectRL first uses these trajectories to elicit Reflective Reasoning, then applies Reflective-to-Direct Policy Transition to transfer the acquired reasoning behavior back to Direct Reasoning. Experiments across 9 benchmarks, 4 LLM backbones, and 4 on-policy training methods show that ReflectRL consistently improves reasoning performance with minimal overhead.
展开完整摘要收起摘要↓
On-policy training has emerged as a powerful post-training paradigm for improving the reasoning capabilities of large language models, and is often enhanced by golden trajectories from stronger expert models. However, when the expert fails on harder problems, existing trajectory-guided methods lose their main source of supervision, and these failed trajectories are typically discarded as negative samples. We argue that such failures, which we call Golden Negative Trajectories, can still provide valuable reasoning signals when treated not as demonstrations to imitate, but as flawed trajectories to reflect upon. We identify a Reflection Advantage: for hard problems, reflecting on a flawed trajectory can be easier and more effective than solving the problem directly from scratch. Motivated by this, we propose ReflectRL, a lightweight plug-and-play framework that learns from Golden Negative Trajectories during on-policy training. ReflectRL first uses these trajectories to elicit Reflective Reasoning, then applies Reflective-to-Direct Policy Transition to transfer the acquired reasoning behavior back to Direct Reasoning. Experiments across 9 benchmarks, 4 LLM backbones, and 4 on-policy training methods show that ReflectRL consistently improves reasoning performance with minimal overhead.
Test-time scaling improves LLM reasoning by generating and aggregating multiple candidate answers, yet many pipelines use fixed per-query budgets that spend the same compute on easy and difficult prompts. These fixed budgets are also difficult to inspect because they do not explain why a given prompt receives a particular number of samples. We propose adaptive} test-time scaling with a lightweight fuzzy controller that maps interpretable signals, including estimated prompt complexity and model confidence, to a per-query sampling budget. The controller assigns fewer samples to easier or more confident prompts and more samples to harder or less certain prompts, making inference-time compute inspectable rather than fixed or opaque. We evaluate under a fair-alignment protocol with matched decoding settings and controlled answer selection, and compare against best-of-$N$, compute-aware scaling, and self-certainty-based baselines on question-answering and mathematical reasoning tasks. Across models and datasets, adaptive fuzzy control improves over several standard baselines and remains close to a selector-matched full-budget control while reducing the average number of samples. These findings suggest that interpretable adaptive sampling is a practical direction for more efficient test-time reasoning in large language models.
展开完整摘要收起摘要↓
Test-time scaling improves LLM reasoning by generating and aggregating multiple candidate answers, yet many pipelines use fixed per-query budgets that spend the same compute on easy and difficult prompts. These fixed budgets are also difficult to inspect because they do not explain why a given prompt receives a particular number of samples. We propose adaptive} test-time scaling with a lightweight fuzzy controller that maps interpretable signals, including estimated prompt complexity and model confidence, to a per-query sampling budget. The controller assigns fewer samples to easier or more confident prompts and more samples to harder or less certain prompts, making inference-time compute inspectable rather than fixed or opaque. We evaluate under a fair-alignment protocol with matched decoding settings and controlled answer selection, and compare against best-of-$N$, compute-aware scaling, and self-certainty-based baselines on question-answering and mathematical reasoning tasks. Across models and datasets, adaptive fuzzy control improves over several standard baselines and remains close to a selector-matched full-budget control while reducing the average number of samples. These findings suggest that interpretable adaptive sampling is a practical direction for more efficient test-time reasoning in large language models.
Statistical reasoning is multidimensional, yet evaluations of large language models (LLMs) typically emphasize response accuracy while overlooking how models construct and communicate statistical explanations. This study demonstrates the value of a multidimensional evaluation by combining response accuracy, response behavior, structural topic modeling, and lexical similarity analysis. The framework is applied to explanations generated by 15 current-generation LLMs responding to 90 questions drawn from four statistics examinations spanning high school, undergraduate, and graduate levels. Accuracy varied substantially across models, ranging from 55% to 78%. In contrast, structural topic modeling revealed a common conceptual organization of statistical reasoning across all models, while lexical similarity analysis identified modest but consistent vendor-specific differences in explanatory style. Models developed by the same vendor (e.g. Anthropic, OpenAI) produced explanations that were slightly more similar than models from different vendors. These findings demonstrate that statistical reasoning in contemporary LLMs cannot be characterized by accuracy alone and illustrate how complementary analyses of response behavior and model-generated explanations provide a more comprehensive evaluation of statistical reasoning in generative AI.
展开完整摘要收起摘要↓
Statistical reasoning is multidimensional, yet evaluations of large language models (LLMs) typically emphasize response accuracy while overlooking how models construct and communicate statistical explanations. This study demonstrates the value of a multidimensional evaluation by combining response accuracy, response behavior, structural topic modeling, and lexical similarity analysis. The framework is applied to explanations generated by 15 current-generation LLMs responding to 90 questions drawn from four statistics examinations spanning high school, undergraduate, and graduate levels. Accuracy varied substantially across models, ranging from 55% to 78%. In contrast, structural topic modeling revealed a common conceptual organization of statistical reasoning across all models, while lexical similarity analysis identified modest but consistent vendor-specific differences in explanatory style. Models developed by the same vendor (e.g. Anthropic, OpenAI) produced explanations that were slightly more similar than models from different vendors. These findings demonstrate that statistical reasoning in contemporary LLMs cannot be characterized by accuracy alone and illustrate how complementary analyses of response behavior and model-generated explanations provide a more comprehensive evaluation of statistical reasoning in generative AI.
作者Jiajun Bao, Zihao Qi, Toni J. B. Liu, Gurbir Arora, Raphaël Sarfati, Nicolas Boullé, Christopher J. Earls
Pretrained large language models (LLMs) have demonstrated in-context learning (ICL) capabilities for numerical inference over sequences serialized as text. Prior work has identified and characterized this form of numerical inference primarily through output-level evaluations such as prediction error. However, how numerical information is organized within LLM representations remains much less understood. To study this internal organization, we adopt a graph signal processing perspective in which attention induces a weighted graph over tokens, while token hidden states define signals on its nodes. Quantitative graph-spectral diagnostics and qualitative token-graph visualizations reveal that representations become more clearly differentiated by input dynamical complexity as context length increases. Simpler inputs produce attention-induced token graphs with stronger global connectivity and smoother, spectrally concentrated hidden-state signals, whereas more complex inputs produce more localized graphs and hidden-state signals with broader spectral support and greater high-frequency energy. Together, these findings point to systematic, context-dependent internal signatures associated with numerical ICL that are conserved across model families.
展开完整摘要收起摘要↓
Pretrained large language models (LLMs) have demonstrated in-context learning (ICL) capabilities for numerical inference over sequences serialized as text. Prior work has identified and characterized this form of numerical inference primarily through output-level evaluations such as prediction error. However, how numerical information is organized within LLM representations remains much less understood. To study this internal organization, we adopt a graph signal processing perspective in which attention induces a weighted graph over tokens, while token hidden states define signals on its nodes. Quantitative graph-spectral diagnostics and qualitative token-graph visualizations reveal that representations become more clearly differentiated by input dynamical complexity as context length increases. Simpler inputs produce attention-induced token graphs with stronger global connectivity and smoother, spectrally concentrated hidden-state signals, whereas more complex inputs produce more localized graphs and hidden-state signals with broader spectral support and greater high-frequency energy. Together, these findings point to systematic, context-dependent internal signatures associated with numerical ICL that are conserved across model families.
How can we better represent the impact of sociocultural structures on decision making in computational cognitive models? Modeling this impact requires traversing multiple levels of semantic representation, however it is not immediately clear to a modeler which levels of representation are most salient to a given situation. Though large language models and cognitively grounded corpus models can represent broad semantic associations through co-occurences, the role of self representations in memory should be accounted for to determine how cultural associations shape decision making. We propose a declarative memory system to be used in the ACT-R cognitive architecture that represents semantic associations at multiple levels via a vector-symbolic autoencoder. We use a simple HRR operation to encode episodic memories differently from semantic memory vectors extracted from text to produce a final chunk activation for a memory request. We use ACT-R cognitive models of a racially contextualized implicit association test (IAT) to test this new declarative memory system.
展开完整摘要收起摘要↓
How can we better represent the impact of sociocultural structures on decision making in computational cognitive models? Modeling this impact requires traversing multiple levels of semantic representation, however it is not immediately clear to a modeler which levels of representation are most salient to a given situation. Though large language models and cognitively grounded corpus models can represent broad semantic associations through co-occurences, the role of self representations in memory should be accounted for to determine how cultural associations shape decision making. We propose a declarative memory system to be used in the ACT-R cognitive architecture that represents semantic associations at multiple levels via a vector-symbolic autoencoder. We use a simple HRR operation to encode episodic memories differently from semantic memory vectors extracted from text to produce a final chunk activation for a memory request. We use ACT-R cognitive models of a racially contextualized implicit association test (IAT) to test this new declarative memory system.
作者Xiao Fei, Yang Zhang, Sarah Almeida Carneiro, Michalis Vazirgiannis
Most multiple-choice question (MCQ) benchmarks evaluate Large Language Models (LLMs) only by whether they select the correct answers. This binary scoring treats all incorrect responses alike, even though an LLM's preferences among incorrect options may contain systematic and useful information about its behavior and ability. We introduce the LLM Nominal Response Model (LLM-NRM), an option-aware psychometric framework that models the full distribution over answer choices to jointly estimate LLM ability and option-level item characteristics, while separating model-specific response calibration sharpness, positional preference, and difficulty-dependent fallback behavior. Across 189 LLMs and 31,554 items from 14 benchmarks, LLM-NRM predicts held-out LLM-item interactions more accurately than binary Item Response models and conventional nominal-response baselines, and its ability estimates achieve the strongest Spearman correlation of 0.920 with the external human-preference Arena.ai Elo leaderboard. Distractor identity contributes +101% additional Fisher Information per item beyond correctness, and incorrect responses alone recover full-information ability estimates with Spearman 0.943. The learned item parameters also enable efficient benchmarking, where 41 selected items preserve the full-bank ranking with Kendall's correlation 0.85, corresponding to a 770 times reduction. In conclusion, we show that incorrect answers carry distinct and useful measurement information rather than representing equivalent mistakes.
展开完整摘要收起摘要↓
Most multiple-choice question (MCQ) benchmarks evaluate Large Language Models (LLMs) only by whether they select the correct answers. This binary scoring treats all incorrect responses alike, even though an LLM's preferences among incorrect options may contain systematic and useful information about its behavior and ability. We introduce the LLM Nominal Response Model (LLM-NRM), an option-aware psychometric framework that models the full distribution over answer choices to jointly estimate LLM ability and option-level item characteristics, while separating model-specific response calibration sharpness, positional preference, and difficulty-dependent fallback behavior. Across 189 LLMs and 31,554 items from 14 benchmarks, LLM-NRM predicts held-out LLM-item interactions more accurately than binary Item Response models and conventional nominal-response baselines, and its ability estimates achieve the strongest Spearman correlation of 0.920 with the external human-preference Arena.ai Elo leaderboard. Distractor identity contributes +101% additional Fisher Information per item beyond correctness, and incorrect responses alone recover full-information ability estimates with Spearman 0.943. The learned item parameters also enable efficient benchmarking, where 41 selected items preserve the full-bank ranking with Kendall's correlation 0.85, corresponding to a 770 times reduction. In conclusion, we show that incorrect answers carry distinct and useful measurement information rather than representing equivalent mistakes.
作者Şuayp Talha Kocabay, Talha Rüzgar Akkuş, Kamer Ali Yuksel
Weight-only quantization substantially reduces the storage of large language model (LLM) transformer blocks, but practical backends often retain the final language-modeling head (LM-head) in BF16 or FP16. Quantizing this projection naively can strongly perturb the vocabulary-logit distribution. We present ARCHead, a packed LM-head compressor that combines a quantized low-rank core, group-wise INT4 residuals, and a low-rank correction fitted in an activation-derived metric. ARCHead stores no dense BF16 head and reduces persistent LM-head storage by 3.7-3.9x. On Qwen3-8B-Base, it uses 25.6% of BF16 head storage while attaining 1.007 relative perplexity; storage-matched naive INT4 yields 1.14-1.16. Replacing the BF16 head left by AWQ or bitsandbytes adds only 0.006-0.007 cross-entropy, with less than 2% throughput change in our measurements. ARCHead therefore complements block quantizers by compressing the large output projection they can leave untouched. Code is available at https://github.com/suayptalha/archead.
展开完整摘要收起摘要↓
Weight-only quantization substantially reduces the storage of large language model (LLM) transformer blocks, but practical backends often retain the final language-modeling head (LM-head) in BF16 or FP16. Quantizing this projection naively can strongly perturb the vocabulary-logit distribution. We present ARCHead, a packed LM-head compressor that combines a quantized low-rank core, group-wise INT4 residuals, and a low-rank correction fitted in an activation-derived metric. ARCHead stores no dense BF16 head and reduces persistent LM-head storage by 3.7-3.9x. On Qwen3-8B-Base, it uses 25.6% of BF16 head storage while attaining 1.007 relative perplexity; storage-matched naive INT4 yields 1.14-1.16. Replacing the BF16 head left by AWQ or bitsandbytes adds only 0.006-0.007 cross-entropy, with less than 2% throughput change in our measurements. ARCHead therefore complements block quantizers by compressing the large output projection they can leave untouched. Code is available at https://github.com/suayptalha/archead.
This research paper describes an exploratory study on the effectiveness of Chat Debugging: troubleshooting malfunctioning analog circuits on breadboards and printed circuit boards (PCB) by undergraduates through conversations with public-domain large language models (LLMs). Through thematic analysis of students' voluntarily shared chat logs when debugging pre-determined buggy circuits under exam and time pressure, we discovered multimodal usage patterns by students and considerable domain knowledge and sensible debugging suggestions offered by off-the-shelf LLMs. Meanwhile, we also identified major gaps in LLM technologies and students' skills during human-AI collaborative debugging, such as LLMs' limitations in 2D/3D image-based reasoning, unjustified tone of confidence, and students' deficits in fundamental concepts and critical thinking.
展开完整摘要收起摘要↓
This research paper describes an exploratory study on the effectiveness of Chat Debugging: troubleshooting malfunctioning analog circuits on breadboards and printed circuit boards (PCB) by undergraduates through conversations with public-domain large language models (LLMs). Through thematic analysis of students' voluntarily shared chat logs when debugging pre-determined buggy circuits under exam and time pressure, we discovered multimodal usage patterns by students and considerable domain knowledge and sensible debugging suggestions offered by off-the-shelf LLMs. Meanwhile, we also identified major gaps in LLM technologies and students' skills during human-AI collaborative debugging, such as LLMs' limitations in 2D/3D image-based reasoning, unjustified tone of confidence, and students' deficits in fundamental concepts and critical thinking.
Large Language Models (LLMs) are accelerating biological research, yet this same capability poses a critical biosecurity threat: models that assist in protein engineering can equally be prompted to generate predicted toxin-like sequences, potentially lowering the barrier to biological misuse. Current safety evaluations, however, operate in natural language and cannot determine whether a model-generated amino acid sequence is biological gibberish or a computational risk signal. To address this evaluation blind spot, we introduce SPIKE-Bench, coupling 631 curated toxin-design prompts across seven functional categories with the SPIKE funnel, a three-stage protocol that filters output through compliance, biological plausibility, and predicted toxicity, producing stage-level diagnostics and an aggregate function-aware metric: the Functional Harmfulness Rate (FHR). An audit of 32 LLMs reveals that most models freely comply with toxin-design requests; FHR is driven primarily by biological generation capability rather than safety alignment, reaching 50.7%; and Refusal Rate fails to predict functional risk. As a first step toward mitigation, we provide BioSafe-Guard, a domain-specialized classifier that substantially reduces predicted functional risk while preserving benign utility. We release SPIKE-Bench and BioSafe-Guard at https://github.com/PKU-Alignment/SPIKE-Bench to support more rigorous biosecurity evaluation of LLMs.
展开完整摘要收起摘要↓
Large Language Models (LLMs) are accelerating biological research, yet this same capability poses a critical biosecurity threat: models that assist in protein engineering can equally be prompted to generate predicted toxin-like sequences, potentially lowering the barrier to biological misuse. Current safety evaluations, however, operate in natural language and cannot determine whether a model-generated amino acid sequence is biological gibberish or a computational risk signal. To address this evaluation blind spot, we introduce SPIKE-Bench, coupling 631 curated toxin-design prompts across seven functional categories with the SPIKE funnel, a three-stage protocol that filters output through compliance, biological plausibility, and predicted toxicity, producing stage-level diagnostics and an aggregate function-aware metric: the Functional Harmfulness Rate (FHR). An audit of 32 LLMs reveals that most models freely comply with toxin-design requests; FHR is driven primarily by biological generation capability rather than safety alignment, reaching 50.7%; and Refusal Rate fails to predict functional risk. As a first step toward mitigation, we provide BioSafe-Guard, a domain-specialized classifier that substantially reduces predicted functional risk while preserving benign utility. We release SPIKE-Bench and BioSafe-Guard at https://github.com/PKU-Alignment/SPIKE-Bench to support more rigorous biosecurity evaluation of LLMs.
作者Shadab Bin Habib, A K M Ferdous Reza Habib, Subarno Neel, Adib Sakhawat
We audit five frontier large language models on native Bangla derogatory speech (gali) across six protocols to test a single hypothesis: Comprehension-Containment Decoupling. We propose that contemporary safety alignment is bound to high-resource surface forms rather than harmful meaning, causing a model's capacity to comprehend a low-resource slur and its capacity to contain it to operate independently. Every protocol corroborates this hypothesis against a human-calibrated baseline (kappa = 0.84). At baseline, models exhibit a 7.92 percentage point comprehension deficit in Bangla while maintaining an identical 92.83% token leakage rate across both languages. Severity calibration tracks surface anatomical cues over compositional harm (+4.00 error on mild slang; -2.00 on threats), while apparent containment gains under orthographic perturbation prove to be a tokenizer-driven "containment mirage." Crucially, explicit Chain-of-Thought reasoning rescues comprehension (94.72% Pass) while systematically dismantling containment (96.23% Use). Furthermore, expert-persona framing collapses refusal to 6.57%, revealing that keyword-based filters ignore dehumanizing communal slurs entirely. Our findings demonstrate that high-resource benchmarks cannot certify low-resource safety, necessitating meaning-grounded containment.
展开完整摘要收起摘要↓
We audit five frontier large language models on native Bangla derogatory speech (gali) across six protocols to test a single hypothesis: Comprehension-Containment Decoupling. We propose that contemporary safety alignment is bound to high-resource surface forms rather than harmful meaning, causing a model's capacity to comprehend a low-resource slur and its capacity to contain it to operate independently. Every protocol corroborates this hypothesis against a human-calibrated baseline (kappa = 0.84). At baseline, models exhibit a 7.92 percentage point comprehension deficit in Bangla while maintaining an identical 92.83% token leakage rate across both languages. Severity calibration tracks surface anatomical cues over compositional harm (+4.00 error on mild slang; -2.00 on threats), while apparent containment gains under orthographic perturbation prove to be a tokenizer-driven "containment mirage." Crucially, explicit Chain-of-Thought reasoning rescues comprehension (94.72% Pass) while systematically dismantling containment (96.23% Use). Furthermore, expert-persona framing collapses refusal to 6.57%, revealing that keyword-based filters ignore dehumanizing communal slurs entirely. Our findings demonstrate that high-resource benchmarks cannot certify low-resource safety, necessitating meaning-grounded containment.
In this work we investigate how decoder-only transformers resolve lexical ambiguity through layer-by-layer analysis of three models spanning three parameter sizes (GPT-2-Small-117M, Llama-3.2-3B, Qwen2.5-32B). For both homonyms and polysemes, we find that representations become maximally distinct in middle layers, then partially reconverge in late layers, while the KL divergence between their next-token predictions reaches its maximum in the final layers. The activation patching experiment provides causal evidence that late-layer representational differences directly determine outputs despite apparent increased similarity in embedding space. Our single-layer ablation experiment indicates that models achieve equivalent disambiguation despite qualitatively different layer-wise vulnerabilities. These findings offer a mechanism for recent observations where models' internal embedding similarities show low correlation with their behavioural outputs despite strong performance. The semantic distinctions therefore remain present but become increasingly invisible to similarity measures over the embeddings, with implications for embedding-based methods such as semantic search, retrieval, and clustering that rely on late-layer cosine similarity.
展开完整摘要收起摘要↓
In this work we investigate how decoder-only transformers resolve lexical ambiguity through layer-by-layer analysis of three models spanning three parameter sizes (GPT-2-Small-117M, Llama-3.2-3B, Qwen2.5-32B). For both homonyms and polysemes, we find that representations become maximally distinct in middle layers, then partially reconverge in late layers, while the KL divergence between their next-token predictions reaches its maximum in the final layers. The activation patching experiment provides causal evidence that late-layer representational differences directly determine outputs despite apparent increased similarity in embedding space. Our single-layer ablation experiment indicates that models achieve equivalent disambiguation despite qualitatively different layer-wise vulnerabilities. These findings offer a mechanism for recent observations where models' internal embedding similarities show low correlation with their behavioural outputs despite strong performance. The semantic distinctions therefore remain present but become increasingly invisible to similarity measures over the embeddings, with implications for embedding-based methods such as semantic search, retrieval, and clustering that rely on late-layer cosine similarity.
In recent years, generative AI has attracted significant public attention, and its use has been rapidly expanding across a wide range of domains. From creative tasks such as text summarization, idea generation, and source code generation, to the streamlining of medical support tasks like diagnostic report generation and summarization, AI is now deeply involved in many areas. Today's breadth of AI applications is clearly distinct from what was seen before generative AI gained widespread recognition. Representative generative AI services include DALL-E 3 (OpenAI, California, USA) and Stable Diffusion (Stability AI, London, England, UK) for image generation, ChatGPT (OpenAI, California, USA), and Gemini (Google, California, USA) for text generation. The rise of generative AI has been influenced by advances in deep learning models and the scaling up of data, models, and computational resources based on the scaling laws. Moreover, the emergence of foundation models, which are trained on large-scale datasets and possess general-purpose knowledge applicable to various downstream tasks, is creating a new paradigm in AI development. These shifts brought about by generative AI and foundation models also profoundly impact medical image processing, fundamentally changing the framework for AI development in healthcare. This paper provides an overview of diffusion models used in image generation AI and large language models (LLMs) used in text generation AI, and introduces their applications in medical support. This paper also discusses foundation models, which are gaining attention alongside generative AI, including their construction methods and applications in the medical field. Finally, the paper explores how to develop foundation models and high-performance AI for medical support by fully utilizing national data and computational resources.
展开完整摘要收起摘要↓
In recent years, generative AI has attracted significant public attention, and its use has been rapidly expanding across a wide range of domains. From creative tasks such as text summarization, idea generation, and source code generation, to the streamlining of medical support tasks like diagnostic report generation and summarization, AI is now deeply involved in many areas. Today's breadth of AI applications is clearly distinct from what was seen before generative AI gained widespread recognition. Representative generative AI services include DALL-E 3 (OpenAI, California, USA) and Stable Diffusion (Stability AI, London, England, UK) for image generation, ChatGPT (OpenAI, California, USA), and Gemini (Google, California, USA) for text generation. The rise of generative AI has been influenced by advances in deep learning models and the scaling up of data, models, and computational resources based on the scaling laws. Moreover, the emergence of foundation models, which are trained on large-scale datasets and possess general-purpose knowledge applicable to various downstream tasks, is creating a new paradigm in AI development. These shifts brought about by generative AI and foundation models also profoundly impact medical image processing, fundamentally changing the framework for AI development in healthcare. This paper provides an overview of diffusion models used in image generation AI and large language models (LLMs) used in text generation AI, and introduces their applications in medical support. This paper also discusses foundation models, which are gaining attention alongside generative AI, including their construction methods and applications in the medical field. Finally, the paper explores how to develop foundation models and high-performance AI for medical support by fully utilizing national data and computational resources.
Post-training can substantially alter language-model behavior, yet aggregate behavior rates do not reveal whether training removes an existing mechanism, creates a new one, or changes how an inherited mechanism is used. We study this question through two mechanistically distinct failures, repetition as a decoding-attractor pathology and sycophancy as a preference-related alignment failure. We introduce behavioral manifold analysis, which isolates behavior-specific geometry by selecting sparse behavior-associated coordinates and lifting them into low-dimensional local charts. We construct these charts in two complementary spaces. ACT captures runtime activation states, while NOC quantifies how strongly the model routes functional information flow through the shared behavior-associated subspace. Across multiple model families, the resulting charts are highly compressed and partially alignable across architectures. Contribution-space charts expose a more architecture-robust shared core, whereas activation-space charts retain stronger family-specific structure. Tracking these charts through controlled post-training reveals a consistent asymmetry. Supervised fine-tuning substantially alters the inherited behavioral geometry, whereas reward optimization changes behavior while largely preserving the underlying chart. This geometric perspective provides a unified framework for understanding the mechanistic distinction between the two objectives. SFT tends to rewrite behavioral geometry, whereas reward optimization primarily reweights it. Code is available at https://github.com/ronglingze/Manifold-Analysis
展开完整摘要收起摘要↓
Post-training can substantially alter language-model behavior, yet aggregate behavior rates do not reveal whether training removes an existing mechanism, creates a new one, or changes how an inherited mechanism is used. We study this question through two mechanistically distinct failures, repetition as a decoding-attractor pathology and sycophancy as a preference-related alignment failure. We introduce behavioral manifold analysis, which isolates behavior-specific geometry by selecting sparse behavior-associated coordinates and lifting them into low-dimensional local charts. We construct these charts in two complementary spaces. ACT captures runtime activation states, while NOC quantifies how strongly the model routes functional information flow through the shared behavior-associated subspace. Across multiple model families, the resulting charts are highly compressed and partially alignable across architectures. Contribution-space charts expose a more architecture-robust shared core, whereas activation-space charts retain stronger family-specific structure. Tracking these charts through controlled post-training reveals a consistent asymmetry. Supervised fine-tuning substantially alters the inherited behavioral geometry, whereas reward optimization changes behavior while largely preserving the underlying chart. This geometric perspective provides a unified framework for understanding the mechanistic distinction between the two objectives. SFT tends to rewrite behavioral geometry, whereas reward optimization primarily reweights it. Code is available at https://github.com/ronglingze/Manifold-Analysis
Large language models (LLMs) achieve remarkable performance but are expensive to deploy due to their enormous size. FP4 quantization, with formats such as MXFP4 and NVFP4, offers an appealing solution with native hardware support on modern accelerators. However, maintaining accuracy under FP4 precision remains difficult. A key bottleneck lies in scale optimization: existing methods tightly couple the quantization and dequantization scales, forcing both to conform to the discrete low-precision format required by hardware, such as E8M0 in MXFP4. Yet the quantization scale is never stored and need not obey this constraint, suggesting a significant untapped optimization space. In this work, we propose FOCUS, a post-training quantization framework with end-to-end scale learning for FP4 Optimization via Coupled-Relaxation and Dual-Granularity Scaling. Coupled-Relaxation Scaling (CRS) relaxes the tight coupling between quantization and dequantization scales with a learnable full-precision coefficient, enabling more effective optimization without breaking hardware compliance. Dual-Granularity Scaling (DGS) further refines the quantization scale at a finer sub-block granularity, allowing more precise adaptation to local weight distributions. Experiments across multiple LLM families and benchmarks show that FOCUS achieves state-of-the-art FP4 accuracy under both MXFP4 and NVFP4 formats, while introducing no additional inference overhead. Code and quantized models will be released at https://github.com/tencent/AngelSlim.
展开完整摘要收起摘要↓
Large language models (LLMs) achieve remarkable performance but are expensive to deploy due to their enormous size. FP4 quantization, with formats such as MXFP4 and NVFP4, offers an appealing solution with native hardware support on modern accelerators. However, maintaining accuracy under FP4 precision remains difficult. A key bottleneck lies in scale optimization: existing methods tightly couple the quantization and dequantization scales, forcing both to conform to the discrete low-precision format required by hardware, such as E8M0 in MXFP4. Yet the quantization scale is never stored and need not obey this constraint, suggesting a significant untapped optimization space. In this work, we propose FOCUS, a post-training quantization framework with end-to-end scale learning for FP4 Optimization via Coupled-Relaxation and Dual-Granularity Scaling. Coupled-Relaxation Scaling (CRS) relaxes the tight coupling between quantization and dequantization scales with a learnable full-precision coefficient, enabling more effective optimization without breaking hardware compliance. Dual-Granularity Scaling (DGS) further refines the quantization scale at a finer sub-block granularity, allowing more precise adaptation to local weight distributions. Experiments across multiple LLM families and benchmarks show that FOCUS achieves state-of-the-art FP4 accuracy under both MXFP4 and NVFP4 formats, while introducing no additional inference overhead. Code and quantized models will be released at https://github.com/tencent/AngelSlim.
Large language model agents increasingly solve complex tasks by composing reusable skills from a library. To address this, the key challenge is not merely to retrieve individually relevant skills, but to identify a complete and executable skill composition. In this paper, we argue that this problem can be solved in a graph with three levels: compositional relations among skill queries, similarity between queries and candidates in the skill library, and the dependencies among the selected candidates. We introduce SkillTrace, which organizes the user query into a semantic hierarchy, matches skill queries and candidates, and propagates over the skill dependencies. Experiments on SkillsBench and ALFWorld demonstrate that SkillTrace achieves state-of-the-art performance, reaching a success rate of 53.17% on SkillsBench and 91.43% on ALFWorld. SkillTrace also delivers consistent improvements across different backbone language models, demonstrating the generality and robustness of graph-based skill retrieval.
展开完整摘要收起摘要↓
Large language model agents increasingly solve complex tasks by composing reusable skills from a library. To address this, the key challenge is not merely to retrieve individually relevant skills, but to identify a complete and executable skill composition. In this paper, we argue that this problem can be solved in a graph with three levels: compositional relations among skill queries, similarity between queries and candidates in the skill library, and the dependencies among the selected candidates. We introduce SkillTrace, which organizes the user query into a semantic hierarchy, matches skill queries and candidates, and propagates over the skill dependencies. Experiments on SkillsBench and ALFWorld demonstrate that SkillTrace achieves state-of-the-art performance, reaching a success rate of 53.17% on SkillsBench and 91.43% on ALFWorld. SkillTrace also delivers consistent improvements across different backbone language models, demonstrating the generality and robustness of graph-based skill retrieval.
Predicting transcriptomic responses to small-molecule perturbations across cell lines is central to drug discovery, but exhaustive profiling of drug-cell combinations is infeasible. We frame molecular perturbation prediction as retrieve-and-aggregate: approximate an unmeasured drug's response in a cell line by aggregating measured responses of a small set of biologically related compounds. We propose LLM-Guided Retrieval (LGR), where a large language model (LLM) ranks candidate neighbor drugs (restricted to those profiled in the target cell line); after which a fixed mean aggregator combines their observed expression deltas to form the prediction. We evaluate on the Tahoe-100M single-cell perturbation atlas under unseen-drug, unseen-cell-line, and open-world regimes. LGR consistently improves over drug mean, ChemCPA, and chemistry-based kNN baselines, with the strongest gains for unseen cell-line generalization, where it achieves higher correlation and lower error than mean baselines. Across settings, LGR improves directional (sign) accuracy of gene regulation, indicating better recovery of biologically meaningful perturbation effects even when magnitude-based metrics are similar. These results suggest that retrieval quality, rather than predictor complexity, is a key driver of zero-shot molecular perturbation prediction, and that LLMs can provide a useful biological prior when used as constrained retrieval modules.
展开完整摘要收起摘要↓
Predicting transcriptomic responses to small-molecule perturbations across cell lines is central to drug discovery, but exhaustive profiling of drug-cell combinations is infeasible. We frame molecular perturbation prediction as retrieve-and-aggregate: approximate an unmeasured drug's response in a cell line by aggregating measured responses of a small set of biologically related compounds. We propose LLM-Guided Retrieval (LGR), where a large language model (LLM) ranks candidate neighbor drugs (restricted to those profiled in the target cell line); after which a fixed mean aggregator combines their observed expression deltas to form the prediction. We evaluate on the Tahoe-100M single-cell perturbation atlas under unseen-drug, unseen-cell-line, and open-world regimes. LGR consistently improves over drug mean, ChemCPA, and chemistry-based kNN baselines, with the strongest gains for unseen cell-line generalization, where it achieves higher correlation and lower error than mean baselines. Across settings, LGR improves directional (sign) accuracy of gene regulation, indicating better recovery of biologically meaningful perturbation effects even when magnitude-based metrics are similar. These results suggest that retrieval quality, rather than predictor complexity, is a key driver of zero-shot molecular perturbation prediction, and that LLMs can provide a useful biological prior when used as constrained retrieval modules.
作者Kang Liu, Zijing Wang, Yongkang Liu, Mengjie Zhao, Xiaocui Yang, Shi Feng, Yifei Zhang, Daling Wang
Multimodal Large Reasoning Models (MLRMs) have achieved strong performance on tasks requiring visual understanding and multi-step inference. However, as reasoning trajectories grow, models may become less effective at using information established earlier in the context, increasing the risk of reasoning errors. Existing approaches primarily address this problem by sustaining visual grounding throughout reasoning. However, reasoning also transforms visual observations into task-specific relations, constraints, and intermediate conclusions whose influence may weaken over long trajectories. Our attribution analysis suggests that correctness is not consistently separated by image attribution alone, but is more closely associated with whether trajectories retain and integrate such reasoning-derived information across stages. Motivated by this, we introduce TRAM (TRajectory-derived Auxiliary Memory), a training-free method that augments standard decoding with an auxiliary memory pathway derived from the model's own reasoning trajectory. TRAM consolidates completed reasoning into a compact latent memory, updates it online through fast and slow recurrent streams, and feeds it back into selected decoder layers through a lightweight residual pathway. Experiments across four MLRM variants on eight benchmarks show that TRAM improves performance over vanilla decoding on mathematical, scientific, and general visual reasoning tasks without additional training.
展开完整摘要收起摘要↓
Multimodal Large Reasoning Models (MLRMs) have achieved strong performance on tasks requiring visual understanding and multi-step inference. However, as reasoning trajectories grow, models may become less effective at using information established earlier in the context, increasing the risk of reasoning errors. Existing approaches primarily address this problem by sustaining visual grounding throughout reasoning. However, reasoning also transforms visual observations into task-specific relations, constraints, and intermediate conclusions whose influence may weaken over long trajectories. Our attribution analysis suggests that correctness is not consistently separated by image attribution alone, but is more closely associated with whether trajectories retain and integrate such reasoning-derived information across stages. Motivated by this, we introduce TRAM (TRajectory-derived Auxiliary Memory), a training-free method that augments standard decoding with an auxiliary memory pathway derived from the model's own reasoning trajectory. TRAM consolidates completed reasoning into a compact latent memory, updates it online through fast and slow recurrent streams, and feeds it back into selected decoder layers through a lightweight residual pathway. Experiments across four MLRM variants on eight benchmarks show that TRAM improves performance over vanilla decoding on mathematical, scientific, and general visual reasoning tasks without additional training.
Fixed-layout indoor furniture styling requires selecting assets that form a coherent room without changing the prescribed furniture categories, positions, orientations, or scales. Existing approaches typically retrieve each asset independently or rely on static local relations, making them prone to shape, material, and color conflicts after scene composition. We introduce StyleForge, a scene-level structured selection framework built on a dynamic hypergraph style field. A frozen multimodal large language model extracts structured style priors from an open-ended style request and the fixed layout, while StyleForge maintains a learnable candidate distribution for each furniture slot. Conditioned on the target style, the dynamic hypergraph style field adaptively activates and weights layout-induced hyperedges to capture higher-order dependencies among furniture. Counterfactual style preference learning then treats each candidate as a local substitution in the current style field and evaluates its contextual compatibility using Mahalanobis energies. Training alternates between optimizing the style field and the candidate logits. At inference, the model remains frozen and test-time training updates only room-specific candidate logits, progressively correcting cross-slot style conflicts as the global scene context evolves. Experiments on 3D-FRONT demonstrate state-of-the-art furniture retrieval and scene-level style coherence, producing more coherent fixed-layout furniture arrangements than object- and scene-level retrieval baselines.
展开完整摘要收起摘要↓
Fixed-layout indoor furniture styling requires selecting assets that form a coherent room without changing the prescribed furniture categories, positions, orientations, or scales. Existing approaches typically retrieve each asset independently or rely on static local relations, making them prone to shape, material, and color conflicts after scene composition. We introduce StyleForge, a scene-level structured selection framework built on a dynamic hypergraph style field. A frozen multimodal large language model extracts structured style priors from an open-ended style request and the fixed layout, while StyleForge maintains a learnable candidate distribution for each furniture slot. Conditioned on the target style, the dynamic hypergraph style field adaptively activates and weights layout-induced hyperedges to capture higher-order dependencies among furniture. Counterfactual style preference learning then treats each candidate as a local substitution in the current style field and evaluates its contextual compatibility using Mahalanobis energies. Training alternates between optimizing the style field and the candidate logits. At inference, the model remains frozen and test-time training updates only room-specific candidate logits, progressively correcting cross-slot style conflicts as the global scene context evolves. Experiments on 3D-FRONT demonstrate state-of-the-art furniture retrieval and scene-level style coherence, producing more coherent fixed-layout furniture arrangements than object- and scene-level retrieval baselines.
Generative artificial intelligence (GenAI) has entered classrooms faster than teachers have been prepared to use it well, producing a GenAI literacy lag in which technological diffusion outpaces educators' conceptual, pedagogical, and ethical readiness. Established AI literacy frameworks predate the widespread adoption of large language models and, while acknowledging ethics, position it as a discrete competency rather than a constitutive commitment, with equity and agency as supplementary design principles. Recent GenAI-specific efforts address isolated features but remain fragmented. We introduce the Responsible AI Literacy in Education (RAIL-Ed) framework, developed through a systematic review and qualitative framework analysis of 67 studies (2023-2025), grounded in critical, pragmatist, sociocultural, and human-centered traditions (Freire, Dewey, Vygotsky, Shneiderman). RAIL-Ed specifies six interdependent pillars: Technical Fluency, Critical Evaluation, Human-AI Collaboration, Contextual Awareness, Ethical Reasoning, and Empowered Agency, marked by three commitments. It is integrative: the absence of any pillar produces a characteristic pedagogical failure. It is developmental: a three-level rubric (Emerging, Competent, Advanced) specifies how each pillar matures across the K-12 teacher-preparation continuum. It is dialectical: the same generative affordance can deepen or displace learning depending on the literacy a teacher brings to it, making the cultivation of that literacy, not the adoption of the tool, the object of design. By treating ethics, equity, and agency as constitutive, RAIL-Ed offers a theoretically grounded basis for curriculum design, teacher education, and policy, aligned with the UNESCO AI Competency Framework for Teachers and the OECD/European Commission AILit Framework. The framework is conceptual, advancing falsifiable propositions for empirical validation.
展开完整摘要收起摘要↓
Generative artificial intelligence (GenAI) has entered classrooms faster than teachers have been prepared to use it well, producing a GenAI literacy lag in which technological diffusion outpaces educators' conceptual, pedagogical, and ethical readiness. Established AI literacy frameworks predate the widespread adoption of large language models and, while acknowledging ethics, position it as a discrete competency rather than a constitutive commitment, with equity and agency as supplementary design principles. Recent GenAI-specific efforts address isolated features but remain fragmented. We introduce the Responsible AI Literacy in Education (RAIL-Ed) framework, developed through a systematic review and qualitative framework analysis of 67 studies (2023-2025), grounded in critical, pragmatist, sociocultural, and human-centered traditions (Freire, Dewey, Vygotsky, Shneiderman). RAIL-Ed specifies six interdependent pillars: Technical Fluency, Critical Evaluation, Human-AI Collaboration, Contextual Awareness, Ethical Reasoning, and Empowered Agency, marked by three commitments. It is integrative: the absence of any pillar produces a characteristic pedagogical failure. It is developmental: a three-level rubric (Emerging, Competent, Advanced) specifies how each pillar matures across the K-12 teacher-preparation continuum. It is dialectical: the same generative affordance can deepen or displace learning depending on the literacy a teacher brings to it, making the cultivation of that literacy, not the adoption of the tool, the object of design. By treating ethics, equity, and agency as constitutive, RAIL-Ed offers a theoretically grounded basis for curriculum design, teacher education, and policy, aligned with the UNESCO AI Competency Framework for Teachers and the OECD/European Commission AILit Framework. The framework is conceptual, advancing falsifiable propositions for empirical validation.
作者Alejandro Velasco, Daniel Rodriguez-Cardenas, Dipin Khati, David N. Palacio, Denys Poshyvanyk
The scientific method has long guided empirical research in Software Engineering (SE), but the complexity of modern software systems often hinders its systematic application. This paper introduces ECLAIR, a causally grounded AI framework that integrates Large Language Models (LLMs) into every stage of the scientific process, from hypothesis generation to analysis and interpretation. ECLAIR treats LLMs as active scientific agents operating under the principles of causal inference, within a human-in-the-loop design that safeguards against the risks of unsound automated reasoning. We demonstrate the framework through a case study examining how prompt design influences code generation accuracy in two LLMs. Results show that, for both models, instruction-style, longer few-shot, and signature-augmented prompts yield small negative causal effects on accuracy, illustrating how causal reasoning provides a principled foundation for explaining why software phenomena occur. This study presents the first causally grounded structured methodology for embedding LLMs within the scientific method in SE, designed around the epistemological demands of empirical SE research, establishing a basis for rigorous AI-assisted research.
展开完整摘要收起摘要↓
The scientific method has long guided empirical research in Software Engineering (SE), but the complexity of modern software systems often hinders its systematic application. This paper introduces ECLAIR, a causally grounded AI framework that integrates Large Language Models (LLMs) into every stage of the scientific process, from hypothesis generation to analysis and interpretation. ECLAIR treats LLMs as active scientific agents operating under the principles of causal inference, within a human-in-the-loop design that safeguards against the risks of unsound automated reasoning. We demonstrate the framework through a case study examining how prompt design influences code generation accuracy in two LLMs. Results show that, for both models, instruction-style, longer few-shot, and signature-augmented prompts yield small negative causal effects on accuracy, illustrating how causal reasoning provides a principled foundation for explaining why software phenomena occur. This study presents the first causally grounded structured methodology for embedding LLMs within the scientific method in SE, designed around the epistemological demands of empirical SE research, establishing a basis for rigorous AI-assisted research.
作者Jiajun Liang, Yucheng Liao, Yukang Cao, Jiazhe Wei, Ken Li, Wende Tan, Jiankun Zhang, ZY Cui, Jingkang Yang, Liucheng Guo, Shiqi Yang, B. Yang, Caifeng Shan, Ziwei Liu, Chenyang Si
Language remains an outlier in generative modeling: while images, video, and audio are increasingly modeled in continuous latent spaces, text generation still relies predominantly on discrete tokens. Existing continuous language models either inherit embedding spaces not designed for joint generation and decoding, or compress autoencoded latents to ease diffusion, sacrificing token-level fidelity. Instead of simplifying the representation to suit the generative model, we preserve a high-capacity, decodable text latent and design the diffusion model to learn its distribution directly. We introduce AURORA-LM, a continuous-latent diffusion language model that separates the construction of a decodable text representation from the modeling of its distribution. A Query-based Encoder-Decoder organizes text into a high-capacity, prefix-aligned latent sequence, and a Block-causal Diffusion Transformer learns its distribution through flow matching, generating blocks left to right while denoising positions within each block in parallel. Because such a latent is harder for diffusion to model, AURORA-LM restricts only the noisy-input pathway while retaining the full clean-latent prediction target, accommodating full-width latents without reducing decoder-facing capacity. We further calibrate the noise-level distribution to the latent width, and introduce self-trajectory consistency to bridge independently sampled training noise and iterative denoising at inference. AURORA-LM achieves the strongest performance among evaluated continuous and diffusion-based language models on OpenWebText free generation and XSum summarization. Scaling to 1B parameters with about 1500 EFLOPs of total compute yields further gains, surpassing a larger publicly released latent-diffusion language model under a matched evaluation protocol. All experiments are conducted on Ascend NPUs.
展开完整摘要收起摘要↓
Language remains an outlier in generative modeling: while images, video, and audio are increasingly modeled in continuous latent spaces, text generation still relies predominantly on discrete tokens. Existing continuous language models either inherit embedding spaces not designed for joint generation and decoding, or compress autoencoded latents to ease diffusion, sacrificing token-level fidelity. Instead of simplifying the representation to suit the generative model, we preserve a high-capacity, decodable text latent and design the diffusion model to learn its distribution directly. We introduce AURORA-LM, a continuous-latent diffusion language model that separates the construction of a decodable text representation from the modeling of its distribution. A Query-based Encoder-Decoder organizes text into a high-capacity, prefix-aligned latent sequence, and a Block-causal Diffusion Transformer learns its distribution through flow matching, generating blocks left to right while denoising positions within each block in parallel. Because such a latent is harder for diffusion to model, AURORA-LM restricts only the noisy-input pathway while retaining the full clean-latent prediction target, accommodating full-width latents without reducing decoder-facing capacity. We further calibrate the noise-level distribution to the latent width, and introduce self-trajectory consistency to bridge independently sampled training noise and iterative denoising at inference. AURORA-LM achieves the strongest performance among evaluated continuous and diffusion-based language models on OpenWebText free generation and XSum summarization. Scaling to 1B parameters with about 1500 EFLOPs of total compute yields further gains, surpassing a larger publicly released latent-diffusion language model under a matched evaluation protocol. All experiments are conducted on Ascend NPUs.
Recurrent linear attention models (RLAs) such as Mamba offer efficient linear-time sequence modeling as an alternative to Transformers, yet their fixed-capacity recurrent states limit long-sequence modeling. Drawing inspiration from hierarchical human memory, we propose Hierarchical Memory Mamba (HMM) to address this limitation. Building upon a pre-trained Mamba backbone, HMM integrates a lightweight working memory that extracts slow paragraph-level semantics (PLS) from the fast sensory memory embedded in the backbone's hidden states. The PLS is subsequently compressed into persistent long-term memory for task-relevant retrieval. The hierarchical processing of semantic information overcomes the representation bottleneck of RLAs and endows HMM cross-task generalization through parametric learning, which is not observed in other long-context enhanced Mamba variants. Evaluations on Passkey Retrieval and LongBench-E tasks demonstrate that HMM improves retrieval success by 34.3--37.1% and reasoning accuracy by 1.6--14.2% over strong Mamba-based models, while adding only 2% extra parameters and with minimal training overhead.
展开完整摘要收起摘要↓
Recurrent linear attention models (RLAs) such as Mamba offer efficient linear-time sequence modeling as an alternative to Transformers, yet their fixed-capacity recurrent states limit long-sequence modeling. Drawing inspiration from hierarchical human memory, we propose Hierarchical Memory Mamba (HMM) to address this limitation. Building upon a pre-trained Mamba backbone, HMM integrates a lightweight working memory that extracts slow paragraph-level semantics (PLS) from the fast sensory memory embedded in the backbone's hidden states. The PLS is subsequently compressed into persistent long-term memory for task-relevant retrieval. The hierarchical processing of semantic information overcomes the representation bottleneck of RLAs and endows HMM cross-task generalization through parametric learning, which is not observed in other long-context enhanced Mamba variants. Evaluations on Passkey Retrieval and LongBench-E tasks demonstrate that HMM improves retrieval success by 34.3--37.1% and reasoning accuracy by 1.6--14.2% over strong Mamba-based models, while adding only 2% extra parameters and with minimal training overhead.
Meta-analysis synthesis highlights a fundamental challenge in knowledge-based scientific analysis: structured evidence does not by itself represent the analytical knowledge required for executable computation. Decisions about evidence assignment, analytical contrasts, outcome and time-point alignment, effect-size formulation, and methodological admissibility must be explicit before statistical execution. Existing automated approaches often embed these decisions in model outputs, generated code, or workflow traces rather than representing them as independently verifiable knowledge. We introduce the Executable Analytical Knowledge Representation (EAKR), a machine-actionable representation of the knowledge required to transform structured evidence into executable meta-analysis. An EAKR represents evidence, relations, numerical inputs, constraints, provenance, and unresolved issues. We operationalise EAKR in MetaSynDec, an agentic harness in which large language models propose structured updates and deterministic services govern schema- and contract-based validation and execution. Across 58 synthesis units, MetaSynDec constructed all EAKRs, with 57 proceeding to statistical execution. Of 56 units with sufficient information to define a reference analysis object, 38 (67.9%) achieved complete object fidelity and 42 (75.0%) exact evidence-set agreement, with a mean Jaccard similarity of 0.909. Generated and published confidence intervals overlapped in 54 of 55 units (98.2%). MetaSynDec outperformed direct LLM generation in reference synthesis-structure agreement (57/58 versus 23/58; p<0.001) and among 23 jointly completed units, exact reference-formulation agreement (23/23 versus 1/23; p<0.001). These findings provide feasibility evidence that EAKR supports formal validation, traceability, statistical execution, and improved methodological agreement relative to direct LLM generation.
展开完整摘要收起摘要↓
Meta-analysis synthesis highlights a fundamental challenge in knowledge-based scientific analysis: structured evidence does not by itself represent the analytical knowledge required for executable computation. Decisions about evidence assignment, analytical contrasts, outcome and time-point alignment, effect-size formulation, and methodological admissibility must be explicit before statistical execution. Existing automated approaches often embed these decisions in model outputs, generated code, or workflow traces rather than representing them as independently verifiable knowledge. We introduce the Executable Analytical Knowledge Representation (EAKR), a machine-actionable representation of the knowledge required to transform structured evidence into executable meta-analysis. An EAKR represents evidence, relations, numerical inputs, constraints, provenance, and unresolved issues. We operationalise EAKR in MetaSynDec, an agentic harness in which large language models propose structured updates and deterministic services govern schema- and contract-based validation and execution. Across 58 synthesis units, MetaSynDec constructed all EAKRs, with 57 proceeding to statistical execution. Of 56 units with sufficient information to define a reference analysis object, 38 (67.9%) achieved complete object fidelity and 42 (75.0%) exact evidence-set agreement, with a mean Jaccard similarity of 0.909. Generated and published confidence intervals overlapped in 54 of 55 units (98.2%). MetaSynDec outperformed direct LLM generation in reference synthesis-structure agreement (57/58 versus 23/58; p<0.001) and among 23 jointly completed units, exact reference-formulation agreement (23/23 versus 1/23; p<0.001). These findings provide feasibility evidence that EAKR supports formal validation, traceability, statistical execution, and improved methodological agreement relative to direct LLM generation.
In-context learning (ICL) has attracted increasing attention for enabling models to perform new tasks using only a few ``input--output'' prompt examples. However, existing approaches suffer from shallow task adaptation, where prompts are primarily used as contextual cues to implicitly infer task intent through semantic representations, while the underlying computational process remains unchanged. This limitation restricts task-specific adaptation and compromises inference interpretability. We argue that prompts should not only condition feature representations but also dynamically regulate the model's computation pathways. To this end, we propose PromptPath, an adaptive ICL framework that enables computation-level adaptation through prompt-conditioned dynamic pathways. Specifically, PromptPath introduces a prompt-driven routing mechanism to selectively activate and compose lightweight low-rank experts, forming task-specific computational pathways tailored to different prompts. By integrating prompt information directly into the inference process, PromptPath dynamically reconfigures model computation to enhance task specialization and interpretability. Extensive experiments on 3D point cloud and 2D visual recognition benchmarks demonstrate that PromptPath consistently outperforms state-of-the-art ICL baselines while exhibiting strong cross-domain and cross-task generalization.
展开完整摘要收起摘要↓
In-context learning (ICL) has attracted increasing attention for enabling models to perform new tasks using only a few ``input--output'' prompt examples. However, existing approaches suffer from shallow task adaptation, where prompts are primarily used as contextual cues to implicitly infer task intent through semantic representations, while the underlying computational process remains unchanged. This limitation restricts task-specific adaptation and compromises inference interpretability. We argue that prompts should not only condition feature representations but also dynamically regulate the model's computation pathways. To this end, we propose PromptPath, an adaptive ICL framework that enables computation-level adaptation through prompt-conditioned dynamic pathways. Specifically, PromptPath introduces a prompt-driven routing mechanism to selectively activate and compose lightweight low-rank experts, forming task-specific computational pathways tailored to different prompts. By integrating prompt information directly into the inference process, PromptPath dynamically reconfigures model computation to enhance task specialization and interpretability. Extensive experiments on 3D point cloud and 2D visual recognition benchmarks demonstrate that PromptPath consistently outperforms state-of-the-art ICL baselines while exhibiting strong cross-domain and cross-task generalization.
This work-in-progress research paper explores Chat-Debugging, a novel use case for large language models as an assistant for hardware debugging tasks to improve students' debugging skills. Hardware debugging can be a time-consuming and stressful skill to develop, leading to frustration and other negative emotions. While past work has explored streamlining and automating software-based circuit debugging where digital circuits are dominant, Chat-Debugging aids in physical hardware debugging where circuits may be analog, digital, or mixed-signal. Qualitative data were collected from LLM chat logs and interviews with a fourth-year electrical engineering undergraduate student. Major themes were extracted using a constant comparative analysis. Chat-Debugging incorporates accurate hardware information, properly handles natural language descriptions of circuits, and improves debugging confidence. A successful Chat-Debugging session includes investigating multiple potential root causes proposed by the LLM, the patience and determination to eliminate root causes, and a student who leads the debugging process by assertively correcting the LLM's misunderstandings. This human-computer interaction can improve electrical and computer engineering students' confidence during debugging and improve their debugging skills.
展开完整摘要收起摘要↓
This work-in-progress research paper explores Chat-Debugging, a novel use case for large language models as an assistant for hardware debugging tasks to improve students' debugging skills. Hardware debugging can be a time-consuming and stressful skill to develop, leading to frustration and other negative emotions. While past work has explored streamlining and automating software-based circuit debugging where digital circuits are dominant, Chat-Debugging aids in physical hardware debugging where circuits may be analog, digital, or mixed-signal. Qualitative data were collected from LLM chat logs and interviews with a fourth-year electrical engineering undergraduate student. Major themes were extracted using a constant comparative analysis. Chat-Debugging incorporates accurate hardware information, properly handles natural language descriptions of circuits, and improves debugging confidence. A successful Chat-Debugging session includes investigating multiple potential root causes proposed by the LLM, the patience and determination to eliminate root causes, and a student who leads the debugging process by assertively correcting the LLM's misunderstandings. This human-computer interaction can improve electrical and computer engineering students' confidence during debugging and improve their debugging skills.
作者Kaustuv Mukherji, Jaikrishna Manojkumar Patil, Colton Payne, Paulo Shakarian, Dana Warmsley, Nigel Stepp, Evelyn Kim
Large language models are increasingly used to reason about software vulnerabilities, but their outputs can silently violate domain knowledge, limiting their reliability in safety-critical settings such as medical devices. Prior work either treats that output as a prediction to be scored or constrains it to walks within a single knowledge graph; neither checks whether reasoning over a binary is consistent with an independent body of domain knowledge. We present EntailLLM, which validates each LLM-proposed analyst path by entailment: the path is a traversal of the binary's function call graph, the domain knowledge is represented in a separate graph, and verification aligns the two under temporal annotated logic. Across three CWE classes, four LLMs, three prompting strategies, and seven binaries varying in size from 405 to 12,696 function call-graph nodes, domain knowledge raises pooled entailment from 78% to 98%, with entailment decreasing in only 3% of the experiments. EntailLLM is deployed end-to-end on real medical-device binaries, reaching 98% pooled entailment without per-device tuning. Our system inherits the formal guarantees of generalized annotated logic, providing logical verification of LLM output that is both explainable and grounded in well-defined semantics.
展开完整摘要收起摘要↓
Large language models are increasingly used to reason about software vulnerabilities, but their outputs can silently violate domain knowledge, limiting their reliability in safety-critical settings such as medical devices. Prior work either treats that output as a prediction to be scored or constrains it to walks within a single knowledge graph; neither checks whether reasoning over a binary is consistent with an independent body of domain knowledge. We present EntailLLM, which validates each LLM-proposed analyst path by entailment: the path is a traversal of the binary's function call graph, the domain knowledge is represented in a separate graph, and verification aligns the two under temporal annotated logic. Across three CWE classes, four LLMs, three prompting strategies, and seven binaries varying in size from 405 to 12,696 function call-graph nodes, domain knowledge raises pooled entailment from 78% to 98%, with entailment decreasing in only 3% of the experiments. EntailLLM is deployed end-to-end on real medical-device binaries, reaching 98% pooled entailment without per-device tuning. Our system inherits the formal guarantees of generalized annotated logic, providing logical verification of LLM output that is both explainable and grounded in well-defined semantics.