In this paper we propose NinaXander, a series of composed language models obtained by connecting layers of frozen language models from different architecture families with a single trained shared-latent adapter. A composed model runs the first layers of one model, converts the resulting intermediate representation once with the adapter, and then runs the remaining layers of the other model. Once the adapter is trained, several composed models that connect at different layers are obtained without retraining. Using the recurrent RWKV-4-Raven-7B and the Transformer-based Tulu-Pythia-6.9b, abbreviated as RWKV and Pythia, this study examines whether frozen models from different families can be recombined post hoc. The composed models answered multiple-choice questions, and those whose generations we examined produced syntactically well-formed text. The configuration that combines the first 5 layers of Pythia with the remaining 27 layers of RWKV reduced the Transformer key-value (KV) cache by 84.4% with accuracy not significantly different from that of RWKV alone. In multiple-choice accuracy, however, no composed model matched the parent model Pythia, and language-modeling performance decreased sharply on WikiText, a corpus of Wikipedia articles outside the training domain. The correspondence between intermediate representations was also obtained in one favorable case, with a shared tokenizer, the same depth, and the same hidden width, and does not show that the models share a general semantic space.
展开完整摘要收起摘要↓
In this paper we propose NinaXander, a series of composed language models obtained by connecting layers of frozen language models from different architecture families with a single trained shared-latent adapter. A composed model runs the first layers of one model, converts the resulting intermediate representation once with the adapter, and then runs the remaining layers of the other model. Once the adapter is trained, several composed models that connect at different layers are obtained without retraining. Using the recurrent RWKV-4-Raven-7B and the Transformer-based Tulu-Pythia-6.9b, abbreviated as RWKV and Pythia, this study examines whether frozen models from different families can be recombined post hoc. The composed models answered multiple-choice questions, and those whose generations we examined produced syntactically well-formed text. The configuration that combines the first 5 layers of Pythia with the remaining 27 layers of RWKV reduced the Transformer key-value (KV) cache by 84.4% with accuracy not significantly different from that of RWKV alone. In multiple-choice accuracy, however, no composed model matched the parent model Pythia, and language-modeling performance decreased sharply on WikiText, a corpus of Wikipedia articles outside the training domain. The correspondence between intermediate representations was also obtained in one favorable case, with a shared tokenizer, the same depth, and the same hidden width, and does not show that the models share a general semantic space.
Scaling laws for optimization against reward models in AI alignment have pinned down how performance depends on optimization effort---measured by a KL-divergence budget relative to a reference policy. Beyond a certain budget, over-optimization (or reward hacking) can arise: because we optimize against a proxy reward model (distinct from true rewards), performance can plateau or degrade. Naturally, the proxy reward's accuracy depends on how much preference data (often in the form of pairwise comparisons) was used to train it. However, existing research does not cleanly identify how performance jointly scales with the amount of training data and the divergence budget. Our main contribution is to provide an empirically accurate and theoretically grounded scaling law in such context. Performance roughly scales as $Θ(\sqrt{\min\{\log(M),K\}})$, where $M$ is the number of comparisons in training data and $K$ is the policy's divergence budget. We develop an information-theoretic model to establish this upper bound and prove it is tightly achievable through a constructive procedure. Informed by this, we conduct extensive empirical evaluations using a real-world annotation setup, whereby a large 70B gold reward model generates feedback data and proxy reward models are trained from less capable models (0.6B to 4B). Our scaling law provides an excellent fit (R2 from 97% to 99%), outperforms alternative specifications, and remains robust across model sizes, noise, and optimization procedures (best-of-$N$ or policy tilting). Our evidence suggests that reward optimization is analogous to a surprisingly simple selection task: choosing from a sequence of IID Gaussian random variables using noisy preference feedback.
展开完整摘要收起摘要↓
Scaling laws for optimization against reward models in AI alignment have pinned down how performance depends on optimization effort---measured by a KL-divergence budget relative to a reference policy. Beyond a certain budget, over-optimization (or reward hacking) can arise: because we optimize against a proxy reward model (distinct from true rewards), performance can plateau or degrade. Naturally, the proxy reward's accuracy depends on how much preference data (often in the form of pairwise comparisons) was used to train it. However, existing research does not cleanly identify how performance jointly scales with the amount of training data and the divergence budget. Our main contribution is to provide an empirically accurate and theoretically grounded scaling law in such context. Performance roughly scales as $Θ(\sqrt{\min\{\log(M),K\}})$, where $M$ is the number of comparisons in training data and $K$ is the policy's divergence budget. We develop an information-theoretic model to establish this upper bound and prove it is tightly achievable through a constructive procedure. Informed by this, we conduct extensive empirical evaluations using a real-world annotation setup, whereby a large 70B gold reward model generates feedback data and proxy reward models are trained from less capable models (0.6B to 4B). Our scaling law provides an excellent fit (R2 from 97% to 99%), outperforms alternative specifications, and remains robust across model sizes, noise, and optimization procedures (best-of-$N$ or policy tilting). Our evidence suggests that reward optimization is analogous to a surprisingly simple selection task: choosing from a sequence of IID Gaussian random variables using noisy preference feedback.
Values such as honesty, autonomy, and confidentiality are often regarded as general principles underpinning AI alignment. However, what it means to act in accordance with these values can depend on the context in which a decision is made. In this paper, we ask whether large language models (LLMs) appropriately adapt the application of a value across professional settings, while remaining consistent when contextual changes do not alter the relevant professional norm. To study this, we introduce ContextAdapt, an evaluation framework covering honesty, autonomy, and confidentiality across medicine, law, finance, and national security. Drawing on primary-source professional and regulatory documents, we construct a value x domain framework and use this to develop scenarios testing both default professional rules and recognised exceptions. We evaluate 12 LLMs on both the actions they recommend and the justifications they provide. In our main experiment, models achieve 95.6% mean appropriateness, although the use of the correct domain-specific justification varies substantially across models, from 25.6% to 76.9%. In a separate factorial experiment, explicitly naming the professional domain and changing the role of the model have limited effect on behaviour. Varying stakes, however, reveals severe but localised failures: in some cases, models alter their responses even though the underlying professional obligation remains unchanged. In particular, perceived severity appears to act as a cue for disclosure across both honesty and confidentiality scenarios. These results show that evaluating value alignment requires us to consider not only whether models follow abstract principles, but whether they apply them appropriately across different contexts.
展开完整摘要收起摘要↓
Values such as honesty, autonomy, and confidentiality are often regarded as general principles underpinning AI alignment. However, what it means to act in accordance with these values can depend on the context in which a decision is made. In this paper, we ask whether large language models (LLMs) appropriately adapt the application of a value across professional settings, while remaining consistent when contextual changes do not alter the relevant professional norm. To study this, we introduce ContextAdapt, an evaluation framework covering honesty, autonomy, and confidentiality across medicine, law, finance, and national security. Drawing on primary-source professional and regulatory documents, we construct a value x domain framework and use this to develop scenarios testing both default professional rules and recognised exceptions. We evaluate 12 LLMs on both the actions they recommend and the justifications they provide. In our main experiment, models achieve 95.6% mean appropriateness, although the use of the correct domain-specific justification varies substantially across models, from 25.6% to 76.9%. In a separate factorial experiment, explicitly naming the professional domain and changing the role of the model have limited effect on behaviour. Varying stakes, however, reveals severe but localised failures: in some cases, models alter their responses even though the underlying professional obligation remains unchanged. In particular, perceived severity appears to act as a cue for disclosure across both honesty and confidentiality scenarios. These results show that evaluating value alignment requires us to consider not only whether models follow abstract principles, but whether they apply them appropriately across different contexts.
Test-time scaling improves model performance by allocating additional compute during inference. Using this compute effectively across multiple context windows requires deciding how to allocate fresh contexts and what information to carry between them. We call a model's ability to make these decisions contextual reasoning. Existing approaches largely prescribe these decisions through their harness; we instead shift them to the model. We introduce 1) Hermes, a family of simple, configurable harnesses that progressively varies model control over context allocation and reuse, and 2) Hermes-Learn, a two-stage framework for learning these capabilities. We find that capable models can exploit this flexibility to scale with additional inference-time compute, while smaller open-source models initially struggle to do so. Training with Hermes-Learn closes this gap, inducing adaptive contextual reasoning strategies that vary with both the problem and the progress of reasoning. These gains generalize across benchmarks and models, extrapolate beyond the inference-time compute seen during training, and transfer to complementary test-time scaling methods beyond Hermes.
展开完整摘要收起摘要↓
Test-time scaling improves model performance by allocating additional compute during inference. Using this compute effectively across multiple context windows requires deciding how to allocate fresh contexts and what information to carry between them. We call a model's ability to make these decisions contextual reasoning. Existing approaches largely prescribe these decisions through their harness; we instead shift them to the model. We introduce 1) Hermes, a family of simple, configurable harnesses that progressively varies model control over context allocation and reuse, and 2) Hermes-Learn, a two-stage framework for learning these capabilities. We find that capable models can exploit this flexibility to scale with additional inference-time compute, while smaller open-source models initially struggle to do so. Training with Hermes-Learn closes this gap, inducing adaptive contextual reasoning strategies that vary with both the problem and the progress of reasoning. These gains generalize across benchmarks and models, extrapolate beyond the inference-time compute seen during training, and transfer to complementary test-time scaling methods beyond Hermes.
We introduce Type-6 logic, a variant of dynamic epistemic logic augmented with two operators (uncertainty and recurrence), designed to model the inferential dynamics of contemporary large language model (LLM) chain-of-thought (CoT) reasoning. Type-6 accounts for common LLM reasoning pathologies such as unlicensed revision, enthymemes, loopbacks, and unverifiable/incorrect claims. We propose a verifier based on Type-6 logic that builds a graph out the trace, and checks it against Type-6's axioms and inference rules. We evaluate our framework on LLM-generated CoTs four splits spanning formal and informal reasoning. Our verifier detects structurally unsound reasoning steps that surface-level heuristics miss, and allows for easy visualisation of the model's reasoning process. In our corpus, our verifier shows that derived contradiction is the most common hard-fail category in CoT, and that only about 3% of the propositions of a trace have impact on the final derivation. Ablation studies show that other verification methods (LLMs-as-judges, other neurosymbolic approaches, etc.) cannot be considered interchangeable: for example, agreement between LLMs-as-judges and LINC is $κ\approx 0.034$, and this persists within a method across underlying models. Type-6, however, is the most agreed-with method amongst the ones we tested. We prove our verifier runs on average-case linear time; and release our logic specification and artefacts.
展开完整摘要收起摘要↓
We introduce Type-6 logic, a variant of dynamic epistemic logic augmented with two operators (uncertainty and recurrence), designed to model the inferential dynamics of contemporary large language model (LLM) chain-of-thought (CoT) reasoning. Type-6 accounts for common LLM reasoning pathologies such as unlicensed revision, enthymemes, loopbacks, and unverifiable/incorrect claims. We propose a verifier based on Type-6 logic that builds a graph out the trace, and checks it against Type-6's axioms and inference rules. We evaluate our framework on LLM-generated CoTs four splits spanning formal and informal reasoning. Our verifier detects structurally unsound reasoning steps that surface-level heuristics miss, and allows for easy visualisation of the model's reasoning process. In our corpus, our verifier shows that derived contradiction is the most common hard-fail category in CoT, and that only about 3% of the propositions of a trace have impact on the final derivation. Ablation studies show that other verification methods (LLMs-as-judges, other neurosymbolic approaches, etc.) cannot be considered interchangeable: for example, agreement between LLMs-as-judges and LINC is $κ\approx 0.034$, and this persists within a method across underlying models. Type-6, however, is the most agreed-with method amongst the ones we tested. We prove our verifier runs on average-case linear time; and release our logic specification and artefacts.
作者Bo Ni, Li Li, Ryan A. Rossi, Franck Dernoncourt, Tyler Derr
Skills are external artifacts that Large Language Models (LLMs) consume at inference time to improve their performance on specialized domains by incorporating relevant procedural and domain knowledge. Expert-authored skills are expensive to produce, and the resulting artifacts are not optimized for the specific model that consumes them, whose failure modes can vary with version, scale and training. In addition, emerging tasks may fall outside the scope of existing skill libraries, creating a need to develop new skills before curated training data become available. Recent works have explored automated skill optimization through reflection, but they require a curated, in-distribution training set, which users might not always have. To address these limitations, we present Prompt2Skill, a framework that builds skills from natural-language task description alone. From the prompt, the system derives a task specification, discovers or synthesizes datasets, and refines the skill in a closed loop of reflective editing. Across four domains spanning question answering, reading comprehension, spreadsheet manipulation, and mathematical reasoning, Prompt2Skill consistently outperforms the direct prompting baseline, achieving an average improvement of 10.8 across open-source and frontier models.
展开完整摘要收起摘要↓
Skills are external artifacts that Large Language Models (LLMs) consume at inference time to improve their performance on specialized domains by incorporating relevant procedural and domain knowledge. Expert-authored skills are expensive to produce, and the resulting artifacts are not optimized for the specific model that consumes them, whose failure modes can vary with version, scale and training. In addition, emerging tasks may fall outside the scope of existing skill libraries, creating a need to develop new skills before curated training data become available. Recent works have explored automated skill optimization through reflection, but they require a curated, in-distribution training set, which users might not always have. To address these limitations, we present Prompt2Skill, a framework that builds skills from natural-language task description alone. From the prompt, the system derives a task specification, discovers or synthesizes datasets, and refines the skill in a closed loop of reflective editing. Across four domains spanning question answering, reading comprehension, spreadsheet manipulation, and mathematical reasoning, Prompt2Skill consistently outperforms the direct prompting baseline, achieving an average improvement of 10.8 across open-source and frontier models.
Large language models (LLMs) remain vulnerable to adversarial attacks that circumvent safety alignment to elicit harmful outputs. It remains unclear how harmfulness and refusal representations evolve over the course of multi-turn attacks, and why single-turn defenses are less effective in multi-turn settings. This work investigates how the geometry and temporal dynamics of harmfulness and refusal representations evolve across multi-turn attacks. We analyzed hidden-state representations from three instruction-tuned LLMs (Llama-3.1-8B-Instruct, Qwen2.5-7B-Instruct, and Gemma-2-9B-it) using three multi-turn attack frameworks (Crescendo, ActorAttack, and X-Teaming), and examined representation behavior across conversation turns, model layers, and token positions under various context configurations. Across models and frameworks, we found that (1) each attack framework traverses different geometric directions, yet each achieves comparable success in eliciting harmful outputs; (2) multi-turn harmfulness directions became increasingly linearly separable at the end-of-turn token position across turns in middle to late model layers; and (3) harmfulness representations are weakly aligned with refusal-related representations. The results indicate that multi-turn attacks do not succeed by suppressing the model's internal representation of harmfulness. Instead, harmfulness representations become increasingly separable across conversation turns, while remaining only weakly aligned with refusal-related representations. The findings are one possible explanation for why static single-turn safety probes may degrade in multi-turn settings, and suggest that robust defenses must consider temporal representation dynamics rather than identifying harmfulness with isolated or single-turn prompts.
展开完整摘要收起摘要↓
Large language models (LLMs) remain vulnerable to adversarial attacks that circumvent safety alignment to elicit harmful outputs. It remains unclear how harmfulness and refusal representations evolve over the course of multi-turn attacks, and why single-turn defenses are less effective in multi-turn settings. This work investigates how the geometry and temporal dynamics of harmfulness and refusal representations evolve across multi-turn attacks. We analyzed hidden-state representations from three instruction-tuned LLMs (Llama-3.1-8B-Instruct, Qwen2.5-7B-Instruct, and Gemma-2-9B-it) using three multi-turn attack frameworks (Crescendo, ActorAttack, and X-Teaming), and examined representation behavior across conversation turns, model layers, and token positions under various context configurations. Across models and frameworks, we found that (1) each attack framework traverses different geometric directions, yet each achieves comparable success in eliciting harmful outputs; (2) multi-turn harmfulness directions became increasingly linearly separable at the end-of-turn token position across turns in middle to late model layers; and (3) harmfulness representations are weakly aligned with refusal-related representations. The results indicate that multi-turn attacks do not succeed by suppressing the model's internal representation of harmfulness. Instead, harmfulness representations become increasingly separable across conversation turns, while remaining only weakly aligned with refusal-related representations. The findings are one possible explanation for why static single-turn safety probes may degrade in multi-turn settings, and suggest that robust defenses must consider temporal representation dynamics rather than identifying harmfulness with isolated or single-turn prompts.
Supervised fine-tuning of discrete diffusion language models masks some response tokens and trains the model to recover their original values from the visible context. The masking pattern therefore determines both the context available to the model and the tokens it learns to predict. Uniform random masking does not explicitly account for the interaction between these choices. We introduce GoldiMask, which selects tokens to reveal as context by approximately maximizing a submodular objective. This objective uses model signals to balance the benefit of revealing tokens against their value as prediction targets. GoldiMask then weights the remaining targets according to how they benefit from the selected context and their remaining learning potential. Across three backbones and three training datasets, GoldiMask achieves the highest average accuracy in most evaluated settings, demonstrating gains on both reasoning and code generation. Component ablations show that both context selection and target weighting contribute to the gains. GoldiMask also reduces decoding iterations on GSM8K and MATH-500 under confidence-threshold parallel decoding, while maintaining comparable accuracy at higher confidence thresholds.
展开完整摘要收起摘要↓
Supervised fine-tuning of discrete diffusion language models masks some response tokens and trains the model to recover their original values from the visible context. The masking pattern therefore determines both the context available to the model and the tokens it learns to predict. Uniform random masking does not explicitly account for the interaction between these choices. We introduce GoldiMask, which selects tokens to reveal as context by approximately maximizing a submodular objective. This objective uses model signals to balance the benefit of revealing tokens against their value as prediction targets. GoldiMask then weights the remaining targets according to how they benefit from the selected context and their remaining learning potential. Across three backbones and three training datasets, GoldiMask achieves the highest average accuracy in most evaluated settings, demonstrating gains on both reasoning and code generation. Component ablations show that both context selection and target weighting contribute to the gains. GoldiMask also reduces decoding iterations on GSM8K and MATH-500 under confidence-threshold parallel decoding, while maintaining comparable accuracy at higher confidence thresholds.
作者Ismael T. Freire, Marceau Nahon, Maud van Lier, Katie Evans, Hélie Bazin, Michele Farisco, Kathinka Evers, Raja Chatila, Mehdi Khamassi
Large Language Models (LLMs) and more broadly Artificial Intelligence (AI) systems are often described and understood in human-like terms, a phenomenon known as anthropomorphism. This paper provides a synthesis of recent literature on anthropomorphism in AI, covering theoretical frameworks, the role of language in framing AI as human-like, the various risks of anthropomorphizing machines, and strategies to mitigate these issues. After examining why we tend to anthropomorphize AI systems and whether we are right to do so, we highlight the impact of linguistic framing on anthropomorphism. Then, we introduce a conceptual taxonomy of risks associated with AI anthropomorphism. This taxonomy groups twenty-one concerns within five analytical categories: epistemic, affective, human agency, normative, and societal and institutional risks. Finally, we relate these concerns to proposed interventions in design, communication, education, and governance. We argue that a better understanding of AI systems requires concepts and theories grounded in their organization and demonstrated capacities. The linguistic shaping of anthropomorphic perceptions should form part of this scientific effort, since our descriptions influence both how these systems are understood and the roles we allow them to occupy in society.
展开完整摘要收起摘要↓
Large Language Models (LLMs) and more broadly Artificial Intelligence (AI) systems are often described and understood in human-like terms, a phenomenon known as anthropomorphism. This paper provides a synthesis of recent literature on anthropomorphism in AI, covering theoretical frameworks, the role of language in framing AI as human-like, the various risks of anthropomorphizing machines, and strategies to mitigate these issues. After examining why we tend to anthropomorphize AI systems and whether we are right to do so, we highlight the impact of linguistic framing on anthropomorphism. Then, we introduce a conceptual taxonomy of risks associated with AI anthropomorphism. This taxonomy groups twenty-one concerns within five analytical categories: epistemic, affective, human agency, normative, and societal and institutional risks. Finally, we relate these concerns to proposed interventions in design, communication, education, and governance. We argue that a better understanding of AI systems requires concepts and theories grounded in their organization and demonstrated capacities. The linguistic shaping of anthropomorphic perceptions should form part of this scientific effort, since our descriptions influence both how these systems are understood and the roles we allow them to occupy in society.
作者Haocheng Xu, Ye Qiao, Phyo Pyae Moe Aung, Alok Mishra, Pavana Prakash, Rolando Pablo Hong Enriquez, Adam Han Wu, Zhiheng Chen, Dejan Milojicic, Sitao Huang
Pragma insertion drives the quality of high-level synthesis (HLS) designs. Choosing the right directives demands expert knowledge and reasoning about loop nesting, data dependences, and memory layout. While existing large language models (LLMs) show promise in code generation, they lack explicit program-structure awareness, limiting their ability to suggest effective pragmas. We present PRISM, a novel structure-augmented LLM that closes this gap by adding compiler-grade structural reasoning to a pretrained, frozen code LLM. It combines three hierarchical program representations, Abstract Syntax Tree (AST), Control-Flow Graph (CFG), and Data-Flow Graph (DFG), injecting them into a specific transformer layer while keeping original code tokens in a separate stream. The cross-attention gate at the injection point allows falling back to the pretrained representation when its structural signal is unhelpful. On zero-shot evaluation in HLS-Eval, PRISM synthesizes 3.5\times as many kernels as Llama3-8B (26.9% vs. 7.7%), and on the kernels where it does succeed, it produces designs that are 2.31\times faster (geomean) than GPT-5-mini's. In the agentic flow, the PRISM codegen outperforms other baselines when optimizing complex code and drives the average normalized improvement across the HLS-Eval suite to 26.4%.
展开完整摘要收起摘要↓
Pragma insertion drives the quality of high-level synthesis (HLS) designs. Choosing the right directives demands expert knowledge and reasoning about loop nesting, data dependences, and memory layout. While existing large language models (LLMs) show promise in code generation, they lack explicit program-structure awareness, limiting their ability to suggest effective pragmas. We present PRISM, a novel structure-augmented LLM that closes this gap by adding compiler-grade structural reasoning to a pretrained, frozen code LLM. It combines three hierarchical program representations, Abstract Syntax Tree (AST), Control-Flow Graph (CFG), and Data-Flow Graph (DFG), injecting them into a specific transformer layer while keeping original code tokens in a separate stream. The cross-attention gate at the injection point allows falling back to the pretrained representation when its structural signal is unhelpful. On zero-shot evaluation in HLS-Eval, PRISM synthesizes 3.5\times as many kernels as Llama3-8B (26.9% vs. 7.7%), and on the kernels where it does succeed, it produces designs that are 2.31\times faster (geomean) than GPT-5-mini's. In the agentic flow, the PRISM codegen outperforms other baselines when optimizing complex code and drives the average normalized improvement across the HLS-Eval suite to 26.4%.
Beam-search-based test-time methods provide an effective way to improve large language model (LLM) performance on long-horizon generation by pruning invalid reasoning paths early, leading to significantly improved reasoning efficiency and more favorable test-time cost scaling. Despite strong empirical success, the theoretical understanding of beam search remains limited. In this paper, we study the test-time compute guarantee of the commonly used beam search framework that uses the model's internal log-likelihood for intermediate scoring, while relying on an external reward model only after a complete response is generated. We first establish a lower bound for vanilla beam search, showing that at least $Ω(C^\star(x)^2)$ samples are required for the optimal response to survive, where $C^\star(x)$ is the token-level coverage coefficient for prompt $x$. This motivates our modified confidence-filtered beam search (CF-Beam), which reduces the sufficient coverage dependence from quadratic to nearly linear under prefix competitiveness, for fixed horizon, gap, and target accuracy. We then show that the regret of CF-Beam is upper-bounded by the probability of rare failure events and the reward estimation error scaled by a path-level coverage coefficient, where the rare-failure term vanishes as per-step sampling increases. Our results highlight a fundamental advantage of beam search over sequence-level inference methods such as Best-of-N and Best-of-Majority. While the guarantees of these approaches typically involve coverage coefficients that grow exponentially with the horizon $L$, CF-Beam controls the dominant search-induced term through a token-level coverage coefficient that scales polynomially with $L$. Our numerical experiments further confirm that beam search is more robust on hard instances and under increasing reasoning horizons.
展开完整摘要收起摘要↓
Beam-search-based test-time methods provide an effective way to improve large language model (LLM) performance on long-horizon generation by pruning invalid reasoning paths early, leading to significantly improved reasoning efficiency and more favorable test-time cost scaling. Despite strong empirical success, the theoretical understanding of beam search remains limited. In this paper, we study the test-time compute guarantee of the commonly used beam search framework that uses the model's internal log-likelihood for intermediate scoring, while relying on an external reward model only after a complete response is generated. We first establish a lower bound for vanilla beam search, showing that at least $Ω(C^\star(x)^2)$ samples are required for the optimal response to survive, where $C^\star(x)$ is the token-level coverage coefficient for prompt $x$. This motivates our modified confidence-filtered beam search (CF-Beam), which reduces the sufficient coverage dependence from quadratic to nearly linear under prefix competitiveness, for fixed horizon, gap, and target accuracy. We then show that the regret of CF-Beam is upper-bounded by the probability of rare failure events and the reward estimation error scaled by a path-level coverage coefficient, where the rare-failure term vanishes as per-step sampling increases. Our results highlight a fundamental advantage of beam search over sequence-level inference methods such as Best-of-N and Best-of-Majority. While the guarantees of these approaches typically involve coverage coefficients that grow exponentially with the horizon $L$, CF-Beam controls the dominant search-induced term through a token-level coverage coefficient that scales polynomially with $L$. Our numerical experiments further confirm that beam search is more robust on hard instances and under increasing reasoning horizons.
Long-horizon LLM agents require reinforcement learning methods that can assign credit to intermediate decisions under sparse and delayed rewards. Existing group-based methods such as GRPO and GiGPO alleviate this issue by comparing rollout returns or repeated anchor states, but they still fail when the compared returns have no variation. We identify this failure mode as zero-credit failure: during early training, many failed rollouts contain useful prefixes, yet existing methods assign them no task-discriminative advantage. To address this issue, we propose Milestone Viability Potential Policy Optimization (MVPO), a potential-routed policy optimization algorithm that learns from viable failure prefixes. MVPO estimates prefix potential over Union-Find viability regions, repairs zero-credit groups with potential-difference advantages, and attenuates the potential branch according to relative performance progress. Experiments with Qwen2.5-1.5B-Instruct show that MVPO outperforms eight strong baselines, including GRPO and GiGPO. Under the same training length, MVPO improves over the GiGPO baseline by +4.4 success points on ALFWorld and +5.3 on WebShop, while adding only 0.16%-0.20% advantage-construction overhead.
展开完整摘要收起摘要↓
Long-horizon LLM agents require reinforcement learning methods that can assign credit to intermediate decisions under sparse and delayed rewards. Existing group-based methods such as GRPO and GiGPO alleviate this issue by comparing rollout returns or repeated anchor states, but they still fail when the compared returns have no variation. We identify this failure mode as zero-credit failure: during early training, many failed rollouts contain useful prefixes, yet existing methods assign them no task-discriminative advantage. To address this issue, we propose Milestone Viability Potential Policy Optimization (MVPO), a potential-routed policy optimization algorithm that learns from viable failure prefixes. MVPO estimates prefix potential over Union-Find viability regions, repairs zero-credit groups with potential-difference advantages, and attenuates the potential branch according to relative performance progress. Experiments with Qwen2.5-1.5B-Instruct show that MVPO outperforms eight strong baselines, including GRPO and GiGPO. Under the same training length, MVPO improves over the GiGPO baseline by +4.4 success points on ALFWorld and +5.3 on WebShop, while adding only 0.16%-0.20% advantage-construction overhead.
Group-relative methods for reinforcement learning with verifiable rewards (RLVR) learn from differences in rollout outcomes. Independently sampling complete trajectories is costly and does not explicitly explore the decision space at critical positions. Feedback on a completed trajectory can reveal which earlier choices the policy reconsiders, suggesting where to sample alternative continuations. We introduce Hindsight-Divergence Localization (HDL), which uses hindsight-induced changes in token log-likelihoods to select branch points. HDL generates a small number of complete root trajectories and fills each training group with continuations from the selected positions under the original task context. Each continuation reuses its root prefix and contributes policy updates only through its newly generated suffix, reducing generation cost while focusing additional exploration and learning on decisions after branching. Experiments with three models across math, code, and agent tasks show gains in both rollout efficiency and task performance. Compared with GRPO at matched group sizes and training steps, HDL yields up to a 2.5$\times$ reduction in generated tokens and a 1.8$\times$ speedup in rollout wall-clock time. Despite this reduced generation budget, HDL improves performance across all three domains, with gains of up to 12.5 points on agent tasks.
展开完整摘要收起摘要↓
Group-relative methods for reinforcement learning with verifiable rewards (RLVR) learn from differences in rollout outcomes. Independently sampling complete trajectories is costly and does not explicitly explore the decision space at critical positions. Feedback on a completed trajectory can reveal which earlier choices the policy reconsiders, suggesting where to sample alternative continuations. We introduce Hindsight-Divergence Localization (HDL), which uses hindsight-induced changes in token log-likelihoods to select branch points. HDL generates a small number of complete root trajectories and fills each training group with continuations from the selected positions under the original task context. Each continuation reuses its root prefix and contributes policy updates only through its newly generated suffix, reducing generation cost while focusing additional exploration and learning on decisions after branching. Experiments with three models across math, code, and agent tasks show gains in both rollout efficiency and task performance. Compared with GRPO at matched group sizes and training steps, HDL yields up to a 2.5$\times$ reduction in generated tokens and a 1.8$\times$ speedup in rollout wall-clock time. Despite this reduced generation budget, HDL improves performance across all three domains, with gains of up to 12.5 points on agent tasks.
作者Shengtian Yang, Ziyu Xiong, Kaibing Yang, Guangfeng Cai, Yewen Li, Peng Jiang, Gai Kun, Qingpeng Cai, Lei Feng
GUI agents automate tasks on digital devices by grounding language instructions in visual interfaces. Existing group-relative reinforcement learning improves GUI action prediction by comparing the rewards of multiple responses sampled from the same GUI state. However, binary evaluation treats spatially different failed clicks as identical and provides no relative signal when all sampled clicks fail. To address these limitations, we propose Spatial Credit Assignment (SCA), which uses the screen coordinates of sampled clicks to refine group-relative credit. Specifically, SCA predicts each held-out response's reward from the other responses in groups containing both successes and failures, then uses the prediction residual to adjust credit. When all sampled clicks fail, SCA instead orders them by distance to the annotated target. These spatial references are used only to construct the training update; the deployed policy remains unchanged. We evaluate whether this correction improves the policy update itself by comparing its error and directional alignment with the exact return gradient in a controlled synthetic study. Across GUI grounding and offline action-prediction benchmarks, SCA improves grounding across professional domains and achieves the strongest results among reinforcement-fine-tuned models on most action-prediction metrics, with consistent gains across the reported GUI suites.
展开完整摘要收起摘要↓
GUI agents automate tasks on digital devices by grounding language instructions in visual interfaces. Existing group-relative reinforcement learning improves GUI action prediction by comparing the rewards of multiple responses sampled from the same GUI state. However, binary evaluation treats spatially different failed clicks as identical and provides no relative signal when all sampled clicks fail. To address these limitations, we propose Spatial Credit Assignment (SCA), which uses the screen coordinates of sampled clicks to refine group-relative credit. Specifically, SCA predicts each held-out response's reward from the other responses in groups containing both successes and failures, then uses the prediction residual to adjust credit. When all sampled clicks fail, SCA instead orders them by distance to the annotated target. These spatial references are used only to construct the training update; the deployed policy remains unchanged. We evaluate whether this correction improves the policy update itself by comparing its error and directional alignment with the exact return gradient in a controlled synthetic study. Across GUI grounding and offline action-prediction benchmarks, SCA improves grounding across professional domains and achieves the strongest results among reinforcement-fine-tuned models on most action-prediction metrics, with consistent gains across the reported GUI suites.
作者Sabrina Kaniewski, Tim Krämer, Julius Bächle, Markus Enzweiler, Michael Menth, Tobias Heer
Retrieval-Augmented Generation (RAG) is increasingly used to enhance Large Language Model (LLM)-based software vulnerability detection by grounding predictions in retrieved vulnerability knowledge, such as vulnerability reports. However, existing RAG-based software vulnerability detection (RAG4SVD) systems are often evaluated using proprietary models, which challenges open science and reproducibility. Further, studies use different datasets, custom knowledge bases, different backbone models, and diverse metrics, which hinders meaningful cross-system comparison. In this work, we study six open-source RAG4SVD systems and address these reproducibility and comparability challenges through (i) reproduction of their experimental settings under an open-weight setting, and (ii) a unified benchmark using a common dataset, metric suite, and pool of open-weight models. Further, RAG4SVD systems typically consist of multiple components, yet are often evaluated only as a whole system, i.e., end-to-end. Therefore, we perform (iii) a component-level analysis that decomposes representative RAG4SVD pipelines into input abstraction, knowledge retrieval, and detection. Our results demonstrate that reproducibility varies substantially across systems. Under the presented unified benchmark, published RAG4SVD performance does not transfer under a controlled open-weight evaluation and depends strongly on the used model. The component analysis shows that effective RAG4SVD depends on the alignment between pipeline stages. For example, oracle knowledge raises retrieval to near-optimal, yet performance remains low (0.51 pairwise accuracy), demonstrating that retrieval effectiveness alone is insufficient for reliable detection. These findings motivate evaluating RAG4SVD not only end-to-end, but at the level of pipeline components, and provide a basis for more standardized, RAG-aware evaluation practices.
展开完整摘要收起摘要↓
Retrieval-Augmented Generation (RAG) is increasingly used to enhance Large Language Model (LLM)-based software vulnerability detection by grounding predictions in retrieved vulnerability knowledge, such as vulnerability reports. However, existing RAG-based software vulnerability detection (RAG4SVD) systems are often evaluated using proprietary models, which challenges open science and reproducibility. Further, studies use different datasets, custom knowledge bases, different backbone models, and diverse metrics, which hinders meaningful cross-system comparison. In this work, we study six open-source RAG4SVD systems and address these reproducibility and comparability challenges through (i) reproduction of their experimental settings under an open-weight setting, and (ii) a unified benchmark using a common dataset, metric suite, and pool of open-weight models. Further, RAG4SVD systems typically consist of multiple components, yet are often evaluated only as a whole system, i.e., end-to-end. Therefore, we perform (iii) a component-level analysis that decomposes representative RAG4SVD pipelines into input abstraction, knowledge retrieval, and detection. Our results demonstrate that reproducibility varies substantially across systems. Under the presented unified benchmark, published RAG4SVD performance does not transfer under a controlled open-weight evaluation and depends strongly on the used model. The component analysis shows that effective RAG4SVD depends on the alignment between pipeline stages. For example, oracle knowledge raises retrieval to near-optimal, yet performance remains low (0.51 pairwise accuracy), demonstrating that retrieval effectiveness alone is insufficient for reliable detection. These findings motivate evaluating RAG4SVD not only end-to-end, but at the level of pipeline components, and provide a basis for more standardized, RAG-aware evaluation practices.
A key strength of Proximal Policy Optimization (PPO) is its learned critic, which uses historical trajectories collected during reinforcement learning to estimate expected returns and reduce policy-gradient variance. However, we find that the critic is also a major source of instability in reinforcement learning for large language models (LLMs). We identify two critic failure modes that destabilize PPO. First, filtering truncated rollouts from both actor and critic shifts the policy objective to reward conditioned on completion, allowing truncation to increase even as conditional reward improves. Second, heterogeneous return noise can cause high-variance prompts to dominate critic updates in finite batches. We introduce EasyPPO to address these failures. Actor-only overlong filtering trains the critic on returns from both completed and truncated rollouts. Noise-normalized critic regression weights each prompt's critic loss by the inverse standard deviation of its sampled returns, balancing noise contributions across prompts. Moderately smaller critic mini-batches confine outlier influence to fewer rollouts during gradient clipping. Across continuous-reward coding on FrontierCS, binary-reward mathematical reasoning on AIME24, and multi-turn search on Search-R1, EasyPPO remains stable throughout the full training horizon and consistently outperforms vanilla PPO, VAPO, and HL-Gauss PPO. Its best validation scores show relative gains of 14.89%, 2.28%, and 9.47% over PPO, respectively.
展开完整摘要收起摘要↓
A key strength of Proximal Policy Optimization (PPO) is its learned critic, which uses historical trajectories collected during reinforcement learning to estimate expected returns and reduce policy-gradient variance. However, we find that the critic is also a major source of instability in reinforcement learning for large language models (LLMs). We identify two critic failure modes that destabilize PPO. First, filtering truncated rollouts from both actor and critic shifts the policy objective to reward conditioned on completion, allowing truncation to increase even as conditional reward improves. Second, heterogeneous return noise can cause high-variance prompts to dominate critic updates in finite batches. We introduce EasyPPO to address these failures. Actor-only overlong filtering trains the critic on returns from both completed and truncated rollouts. Noise-normalized critic regression weights each prompt's critic loss by the inverse standard deviation of its sampled returns, balancing noise contributions across prompts. Moderately smaller critic mini-batches confine outlier influence to fewer rollouts during gradient clipping. Across continuous-reward coding on FrontierCS, binary-reward mathematical reasoning on AIME24, and multi-turn search on Search-R1, EasyPPO remains stable throughout the full training horizon and consistently outperforms vanilla PPO, VAPO, and HL-Gauss PPO. Its best validation scores show relative gains of 14.89%, 2.28%, and 9.47% over PPO, respectively.
Reward hacking occurs when policy optimization exploits a brittle reward interface or an overly permissive proxy objective, improving the training score without improving the underlying response quality. This phenomenon is amplified in group-relative policy optimization: an unsupported reward can shift the group baseline and alter the updates of other rollouts, while post-hoc or purely relative weighting cannot represent group-wide uncertainty. We propose Self-Tuned Anchored Reliability Group-Relative Policy Optimization (STAR-GRPO), a reliability-first advantage estimator based on paired assessments of the same rollout. STAR separates the quality signal from its learning influence: score disagreement determines rollout reliability, relative reliability enters a self-tuned robust location--scale fit before group normalization, and absolute group reliability attenuates the resulting bounded advantage. The analysis establishes coordinate and second-moment bounds, characterizes exact centering through the weighted location equation, and gives reliability-dependent attenuation guarantees for outlying rewards. We evaluate STAR-GRPO in two complementary reward-hacking regimes. In token-interface exploitation, STAR prevents runaway optimization of the deployed-interface score while improving the canonical quality signal. In rubric-proxy overoptimization for medical reasoning, STAR improves independent semantic evaluation, narrows the proxy--judge discrepancy, and reduces overclaim while optimizing the same task proxy. Together, these results show that reliability-first normalization offers a principled way to limit unsupported reward influence on both group baselines and policy updates, while retaining the task reward as the optimization target.
展开完整摘要收起摘要↓
Reward hacking occurs when policy optimization exploits a brittle reward interface or an overly permissive proxy objective, improving the training score without improving the underlying response quality. This phenomenon is amplified in group-relative policy optimization: an unsupported reward can shift the group baseline and alter the updates of other rollouts, while post-hoc or purely relative weighting cannot represent group-wide uncertainty. We propose Self-Tuned Anchored Reliability Group-Relative Policy Optimization (STAR-GRPO), a reliability-first advantage estimator based on paired assessments of the same rollout. STAR separates the quality signal from its learning influence: score disagreement determines rollout reliability, relative reliability enters a self-tuned robust location--scale fit before group normalization, and absolute group reliability attenuates the resulting bounded advantage. The analysis establishes coordinate and second-moment bounds, characterizes exact centering through the weighted location equation, and gives reliability-dependent attenuation guarantees for outlying rewards. We evaluate STAR-GRPO in two complementary reward-hacking regimes. In token-interface exploitation, STAR prevents runaway optimization of the deployed-interface score while improving the canonical quality signal. In rubric-proxy overoptimization for medical reasoning, STAR improves independent semantic evaluation, narrows the proxy--judge discrepancy, and reduces overclaim while optimizing the same task proxy. Together, these results show that reliability-first normalization offers a principled way to limit unsupported reward influence on both group baselines and policy updates, while retaining the task reward as the optimization target.
Part-level 3D assets are essential for editing, reassembly, and interaction, yet recovering such structure from a single image remains challenging due to occlusion, ambiguous boundaries, and the need for coherent multi-part reasoning. Existing approaches struggle to achieve both controllable part-level generation and coherent multi-part structure, as part identity and spatial allocation are typically inferred implicitly. We present Seg3DParts, a segmentation-grounded framework for controllable part-level 3D generation from a single image. By treating segmentation as an explicit grounding signal, our method defines part identity during generation, enabling each component to be anchored to a corresponding image region. To ensure coherent assemblies, we introduce structured cross-part interaction that allows components to exchange global context throughout the generative process. As a result, Seg3DParts directly generates well-aligned part meshes in a shared canonical space without post-hoc alignment, supporting flexible and controllable decomposition. We further introduce PartObjectNet, a large-scale dataset with over 200K objects and 1M annotated parts. Experiments demonstrate that Seg3DParts achieves superior geometry quality, cross-part coherence, and part-level controllability over existing methods.
展开完整摘要收起摘要↓
Part-level 3D assets are essential for editing, reassembly, and interaction, yet recovering such structure from a single image remains challenging due to occlusion, ambiguous boundaries, and the need for coherent multi-part reasoning. Existing approaches struggle to achieve both controllable part-level generation and coherent multi-part structure, as part identity and spatial allocation are typically inferred implicitly. We present Seg3DParts, a segmentation-grounded framework for controllable part-level 3D generation from a single image. By treating segmentation as an explicit grounding signal, our method defines part identity during generation, enabling each component to be anchored to a corresponding image region. To ensure coherent assemblies, we introduce structured cross-part interaction that allows components to exchange global context throughout the generative process. As a result, Seg3DParts directly generates well-aligned part meshes in a shared canonical space without post-hoc alignment, supporting flexible and controllable decomposition. We further introduce PartObjectNet, a large-scale dataset with over 200K objects and 1M annotated parts. Experiments demonstrate that Seg3DParts achieves superior geometry quality, cross-part coherence, and part-level controllability over existing methods.
作者Xinkai Du, Chao Lv, Yalin Sun, Quanjie Han, Lei Yao, Maosong Sun
Retrieval-Augmented Generation has established itself as a fundamental framework in natural language processing, seamlessly integrating information retrieval with the generative capabilities of large language models. However, this process is fundamentally constrained by a critical challenge: semantic space mismatch between queries and retrieved contexts. We propose Knowledge-Aware Semantic Bridging (KASB), a novel framework that improves passage selection quality through semantic space alignment between queries and retrieved documents through intelligent knowledge fusion. Our approach leverages the complementary strengths of generative and retrieval-based knowledge through a multistage process that enhances both relevance and accuracy. We evaluate KASB on three popular open-domain Question Answering datasets to demonstrate the effectiveness of our approach.
展开完整摘要收起摘要↓
Retrieval-Augmented Generation has established itself as a fundamental framework in natural language processing, seamlessly integrating information retrieval with the generative capabilities of large language models. However, this process is fundamentally constrained by a critical challenge: semantic space mismatch between queries and retrieved contexts. We propose Knowledge-Aware Semantic Bridging (KASB), a novel framework that improves passage selection quality through semantic space alignment between queries and retrieved documents through intelligent knowledge fusion. Our approach leverages the complementary strengths of generative and retrieval-based knowledge through a multistage process that enhances both relevance and accuracy. We evaluate KASB on three popular open-domain Question Answering datasets to demonstrate the effectiveness of our approach.
Experienced professionals know more than just facts and conclusions. They know which cues matter, why a judgment is reasonable, and which action to take. Routine work records often leave out this tacit knowledge, making it difficult for Large Language Model (LLM) agents to use professional experience effectively. We introduce KUPAS MASTER, an experience engineering platform built around nine-layer cognitive corpus construction. It turns heterogeneous work records and practitioner interviews into traceable, reusable experience corpora for agents. Six case elements preserve the task process: context, cues, judgment, action, boundaries, and outcomes. Nine-layer cognitive corpus construction organizes tacit experience along nine extraction dimensions and stores the resulting assets in six libraries: rules, constraints, best practices, negative examples, corner cases, and skills. Semantic alignment, individual experience distillation, organizational consolidation, and cross-review preserve source evidence, conditions of use, and unresolved disagreements. The platform packages these assets into callable skills with explicit inputs, steps, dependencies, and stopping conditions, connecting experience collection to task execution and evaluation feedback. Using authorized samples from 20 randomly selected practitioners, the platform processed 1,576 source files into 23,024 individual experience records and 13,113 organizational assets. The evaluation spans multiple professional domains. Under common task inputs and scoring criteria, the base model, raw corpus retrieval-augmented generation (RAG), and KUPAS MASTER agent scored 70.63, 79.75, and 89.58, respectively. The KUPAS MASTER agent improved on raw-corpus RAG in all seven scoring dimensions. The platform provides a practical path from individual tacit experience to organizational knowledge and agent capabilities.
展开完整摘要收起摘要↓
Experienced professionals know more than just facts and conclusions. They know which cues matter, why a judgment is reasonable, and which action to take. Routine work records often leave out this tacit knowledge, making it difficult for Large Language Model (LLM) agents to use professional experience effectively. We introduce KUPAS MASTER, an experience engineering platform built around nine-layer cognitive corpus construction. It turns heterogeneous work records and practitioner interviews into traceable, reusable experience corpora for agents. Six case elements preserve the task process: context, cues, judgment, action, boundaries, and outcomes. Nine-layer cognitive corpus construction organizes tacit experience along nine extraction dimensions and stores the resulting assets in six libraries: rules, constraints, best practices, negative examples, corner cases, and skills. Semantic alignment, individual experience distillation, organizational consolidation, and cross-review preserve source evidence, conditions of use, and unresolved disagreements. The platform packages these assets into callable skills with explicit inputs, steps, dependencies, and stopping conditions, connecting experience collection to task execution and evaluation feedback. Using authorized samples from 20 randomly selected practitioners, the platform processed 1,576 source files into 23,024 individual experience records and 13,113 organizational assets. The evaluation spans multiple professional domains. Under common task inputs and scoring criteria, the base model, raw corpus retrieval-augmented generation (RAG), and KUPAS MASTER agent scored 70.63, 79.75, and 89.58, respectively. The KUPAS MASTER agent improved on raw-corpus RAG in all seven scoring dimensions. The platform provides a practical path from individual tacit experience to organizational knowledge and agent capabilities.
作者Mei Okonkwo, Pixel Nomand, Julian Berg, Elena Voss, Lena Park, Marcus Hale, Adrian Cho, Sofia Reyes
Group relative policy optimization (GRPO) learns only from prompts whose sampled responses disagree: a group that is entirely correct or entirely incorrect has zero reward variance, contributes no gradient, and still consumes its rollouts. Prompt-selection methods reduce this waste by steering sampling toward intermediate pass rates, but they choose the target, its width, and the uncertainty model heuristically, in raw pass-rate or logit coordinates. We show that GRPO comes with a natural coordinate for pass rates: the arc length $ψ=\arcsin\sqrt{p}$ on the Bernoulli Fisher--Rao manifold. In arc length, the expected GRPO update is uniform up to two boundary ramps; the probability of a zero-variance group is bounded by two Gaussian boundary layers of width $1/\sqrt{2G}$; pass-rate evidence has constant noise; and the gradients of the pass@$k$ and pass$^k$ objectives are Gaussians whose center and width follow from $k$ in closed form. A prompt curriculum for GRPO is therefore a Gaussian in arc length, and choosing its center amounts to choosing the objective. We turn this observation into ARCUS, a drop-in sampler that tracks every prompt with a Kalman filter in arc length, scores prompts by an objective-matched Gaussian kernel times the predicted probability of an informative group, keeps only informative groups for the unchanged GRPO update, and paces the target toward the hardest objective whose predicted yield stays within a small slack of the best. Across six mathematical reasoning benchmarks and three backbones, ARCUS improves the average accuracy of GRPO by 2.8--2.9 points and that of dynamic sampling by 1.1--1.2 points, while generating 48--57% fewer rollouts than dynamic sampling.
展开完整摘要收起摘要↓
Group relative policy optimization (GRPO) learns only from prompts whose sampled responses disagree: a group that is entirely correct or entirely incorrect has zero reward variance, contributes no gradient, and still consumes its rollouts. Prompt-selection methods reduce this waste by steering sampling toward intermediate pass rates, but they choose the target, its width, and the uncertainty model heuristically, in raw pass-rate or logit coordinates. We show that GRPO comes with a natural coordinate for pass rates: the arc length $ψ=\arcsin\sqrt{p}$ on the Bernoulli Fisher--Rao manifold. In arc length, the expected GRPO update is uniform up to two boundary ramps; the probability of a zero-variance group is bounded by two Gaussian boundary layers of width $1/\sqrt{2G}$; pass-rate evidence has constant noise; and the gradients of the pass@$k$ and pass$^k$ objectives are Gaussians whose center and width follow from $k$ in closed form. A prompt curriculum for GRPO is therefore a Gaussian in arc length, and choosing its center amounts to choosing the objective. We turn this observation into ARCUS, a drop-in sampler that tracks every prompt with a Kalman filter in arc length, scores prompts by an objective-matched Gaussian kernel times the predicted probability of an informative group, keeps only informative groups for the unchanged GRPO update, and paces the target toward the hardest objective whose predicted yield stays within a small slack of the best. Across six mathematical reasoning benchmarks and three backbones, ARCUS improves the average accuracy of GRPO by 2.8--2.9 points and that of dynamic sampling by 1.1--1.2 points, while generating 48--57% fewer rollouts than dynamic sampling.
Reinforcement learning (RL) has become a cornerstone for improving the reasoning capabilities of large language models (LLMs), but the need for on-policy data substantially limits training efficiency. Reusing off-policy data through importance sampling (IS) can improve efficiency but introduce considerable instability. Hence, algorithms such as PPO and GRPO widely adopt IS-ratio clipping to stabilize training. However, training stability and gradient estimate are mainly determined by the product of IS ratio and advantage. To further stabilize training, we propose ACPO, which clips the product of the IS ratio and the advantage, leading to more stable gradient estimates. We also establish a connection between ACPO and gradient clipping in policy mirror descent (PMD), which is a standard technique to stabilize optimization process, and prove the convergence of clipped-PMD under the standard RL setting. Experiments on widely used mathematical reasoning benchmarks show that ACPO consistently outperforms PPO and GRPO in both accuracy and training efficiency, delivering 4-6 percentage points gains on standard math benchmarks, with Qwen3-8B+PPO. Hence, ACPO is a practical and effective alternative to conventional IS-ratio clipping for RL post-training of LLMs.
展开完整摘要收起摘要↓
Reinforcement learning (RL) has become a cornerstone for improving the reasoning capabilities of large language models (LLMs), but the need for on-policy data substantially limits training efficiency. Reusing off-policy data through importance sampling (IS) can improve efficiency but introduce considerable instability. Hence, algorithms such as PPO and GRPO widely adopt IS-ratio clipping to stabilize training. However, training stability and gradient estimate are mainly determined by the product of IS ratio and advantage. To further stabilize training, we propose ACPO, which clips the product of the IS ratio and the advantage, leading to more stable gradient estimates. We also establish a connection between ACPO and gradient clipping in policy mirror descent (PMD), which is a standard technique to stabilize optimization process, and prove the convergence of clipped-PMD under the standard RL setting. Experiments on widely used mathematical reasoning benchmarks show that ACPO consistently outperforms PPO and GRPO in both accuracy and training efficiency, delivering 4-6 percentage points gains on standard math benchmarks, with Qwen3-8B+PPO. Hence, ACPO is a practical and effective alternative to conventional IS-ratio clipping for RL post-training of LLMs.
作者Ruochen Zhang, Yao Huang, Yitong Sun, Jiahe Xie, Jin Yan, Jifan Ma, Yuanfang Guo, Xingxing Wei
While Multimodal Large Language Models (MLLMs) are increasingly deployed in safety-critical domains, their reliability is threatened by multimodal implicit risks. Unlike explicit threats, these hazards emerge when individually benign text and neutral visual entities logically converge to induce unsafe outputs. Current detection methods fail to address this because they overlook the underlying risk activation mechanisms that govern cross-modal risk activation, leading to single-modality shortcut learning and hallucinated rationalizations. To bridge this gap, we first construct TriggerBench, the first dataset explicitly modeling risk compositionality (5,600 instances). By formally isolating Key Elements and Trigger Elements to build counterfactual contrastive pairs, TriggerBench eliminates risk residues and forces models to perform genuine logical deduction rather than superficial pattern matching, which provides a rigorous foundation for both large-scale training and fine-grained evaluation. Building on this, we propose a Step-Supervised Structured Reasoning training framework and employ it to train ThinkingGuard, a specialized guard model. Inspired by Situation Awareness theory, we decouple implicit risk identification into progressive cognitive stages, and utilize a step-reward Monte Carlo Tree Search algorithm to explore optimal reasoning trajectories, which are then distilled into the model through Dual-Constraint Preference Alignment. Extensive experiments across both standard and implicit safety benchmarks demonstrate that ThinkingGuard achieves strong performance. Project resources are available at https://github.com/FroggyChen/ThinkingGuard.
展开完整摘要收起摘要↓
While Multimodal Large Language Models (MLLMs) are increasingly deployed in safety-critical domains, their reliability is threatened by multimodal implicit risks. Unlike explicit threats, these hazards emerge when individually benign text and neutral visual entities logically converge to induce unsafe outputs. Current detection methods fail to address this because they overlook the underlying risk activation mechanisms that govern cross-modal risk activation, leading to single-modality shortcut learning and hallucinated rationalizations. To bridge this gap, we first construct TriggerBench, the first dataset explicitly modeling risk compositionality (5,600 instances). By formally isolating Key Elements and Trigger Elements to build counterfactual contrastive pairs, TriggerBench eliminates risk residues and forces models to perform genuine logical deduction rather than superficial pattern matching, which provides a rigorous foundation for both large-scale training and fine-grained evaluation. Building on this, we propose a Step-Supervised Structured Reasoning training framework and employ it to train ThinkingGuard, a specialized guard model. Inspired by Situation Awareness theory, we decouple implicit risk identification into progressive cognitive stages, and utilize a step-reward Monte Carlo Tree Search algorithm to explore optimal reasoning trajectories, which are then distilled into the model through Dual-Constraint Preference Alignment. Extensive experiments across both standard and implicit safety benchmarks demonstrate that ThinkingGuard achieves strong performance. Project resources are available at https://github.com/FroggyChen/ThinkingGuard.
Chinese Semantic Error Correction (CSEC) targets semantic errors in Chinese text, which are typically more subtle and complex than spelling and grammatical errors but remain relatively underexplored. Existing LLM-based approaches face two recurring obstacles in this task: over-correction, and unclear interaction between Chain-of-Thought (CoT) reasoning and self-consistency decoding, such that the benefits brought by CoT cannot be reliably transferred to final corrections. We propose Vote-guided Advantage Allocation for CSEC (VAA-CSEC), a multi-stage framework that combines CoT distillation, Supervised Fine-Tuning (SFT), Reinforcement Learning (RL) and self-consistency decoding. During RL, we design a task-specific reward function that directly aligned with the minimal-editing principle of CSEC. We further introduce Group-Level Relative Policy Optimization (GLPO), which reallocates GRPO advantages according to the margin between individual rollout rewards and the vote-aggregated group reward, aligning the RL training objective with the self-consistency objective used at inference time. Experiments on CSED-C and NaSGEC-Exam show that VAA-CSEC outperforms all LLM-based baselines on CSED-C with an F0.5 of 47.72%, achieves the highest recall of 42.15% among all methods, and establishes a new state of the art of 41.55% F0.5 on NaSGEC-Exam.
展开完整摘要收起摘要↓
Chinese Semantic Error Correction (CSEC) targets semantic errors in Chinese text, which are typically more subtle and complex than spelling and grammatical errors but remain relatively underexplored. Existing LLM-based approaches face two recurring obstacles in this task: over-correction, and unclear interaction between Chain-of-Thought (CoT) reasoning and self-consistency decoding, such that the benefits brought by CoT cannot be reliably transferred to final corrections. We propose Vote-guided Advantage Allocation for CSEC (VAA-CSEC), a multi-stage framework that combines CoT distillation, Supervised Fine-Tuning (SFT), Reinforcement Learning (RL) and self-consistency decoding. During RL, we design a task-specific reward function that directly aligned with the minimal-editing principle of CSEC. We further introduce Group-Level Relative Policy Optimization (GLPO), which reallocates GRPO advantages according to the margin between individual rollout rewards and the vote-aggregated group reward, aligning the RL training objective with the self-consistency objective used at inference time. Experiments on CSED-C and NaSGEC-Exam show that VAA-CSEC outperforms all LLM-based baselines on CSED-C with an F0.5 of 47.72%, achieves the highest recall of 42.15% among all methods, and establishes a new state of the art of 41.55% F0.5 on NaSGEC-Exam.
Rendering accurate Chinese text remains challenging for text-to-image models. Existing OCR-based reinforcement-learning rewards compare decoded transcripts with target strings. Such rewards overlook the compositional nature of Chinese writing: an ideograph consists of reusable components arranged through explicit spatial relations, yet OCR evaluates it as an atomic character. Consequently, visually different radical-level errors may receive equally coarse feedback, encouraging glyphs that merely resemble the target instead of faithfully reproducing its internal structure. We employ Ideographic Description Sequences (IDS), which comprise spatial operators and character components, and train an expert IDS recognizer to transcribe rendered Chinese text into this representation. Building on this recognizer, we introduce IDSpect, which deterministically decomposes the target text into IDS tokens and aligns crop-level visual IDS predictions with the target sequence. Globally unique token credit makes this comparison robust to the order of detected text regions. Combined with a whole-character semantic reward, IDSpect supplies fine-grained credit with component and spatial-relation without changing the image generator or adding inference-time cost. Experiments with GRPO post-training of Qwen-Image demonstrate that IDSpect achieves leading structural quality and semantic alignment on LongText and GenTextEval.
展开完整摘要收起摘要↓
Rendering accurate Chinese text remains challenging for text-to-image models. Existing OCR-based reinforcement-learning rewards compare decoded transcripts with target strings. Such rewards overlook the compositional nature of Chinese writing: an ideograph consists of reusable components arranged through explicit spatial relations, yet OCR evaluates it as an atomic character. Consequently, visually different radical-level errors may receive equally coarse feedback, encouraging glyphs that merely resemble the target instead of faithfully reproducing its internal structure. We employ Ideographic Description Sequences (IDS), which comprise spatial operators and character components, and train an expert IDS recognizer to transcribe rendered Chinese text into this representation. Building on this recognizer, we introduce IDSpect, which deterministically decomposes the target text into IDS tokens and aligns crop-level visual IDS predictions with the target sequence. Globally unique token credit makes this comparison robust to the order of detected text regions. Combined with a whole-character semantic reward, IDSpect supplies fine-grained credit with component and spatial-relation without changing the image generator or adding inference-time cost. Experiments with GRPO post-training of Qwen-Image demonstrate that IDSpect achieves leading structural quality and semantic alignment on LongText and GenTextEval.
作者Zhenwen Ji, Lei Jin, Shanyong Wang, Jiaming Lu, Chengqiang Lu, Yi Wu, Yao Hu, Lizhen Cui, Yanyu Xu
Detailed image captioning requires accurate and comprehensive descriptions of fine-grained visual content, yet caption quality spans factual accuracy, information coverage, and clarity. Compared with conventional methods that rely mainly on high-quality supervision or holistic rewards, rubric-based reinforcement learning decomposes these requirements into explicit criteria and provides targeted, structured feedback. However, existing methods often use separate models for caption generation, rubric construction, and judging, which may lead to inconsistent interpretations across roles. Some dynamic rubric methods alternate updates between the caption policy and rubric generator while keeping the judge fixed, but staged optimization may still leave rubric construction and judging out of step with policy optimization. We propose MoCo Rubric, a two-stage framework that coordinates these roles. First, role-conditioned, shared-parameter multi-task supervised fine-tuning equips a single vision--language model to serve as the Caption Policy, Rubric Generator, and Rubric Judge. Then, the Generator constructs rubrics online from captions sampled by the current Policy, reference captions, and image evidence. The Judge provides rubric-based rewards, and only the Policy receives GRPO updates. As Policy updates change the candidates being evaluated, we use an exponential moving average of the Policy parameters to update one momentum model shared by the Generator and Judge. This gradual transfer lets both rubric roles track Policy updates without separate RL optimization while smoothing parameter changes that could disrupt their rubric capabilities under direct synchronization. Across five captioning benchmarks, MoCo Rubric achieves an average pairwise win rate of 72.83%, the best mean rank in blind ranking, and the highest average score in caption-based question answering.
展开完整摘要收起摘要↓
Detailed image captioning requires accurate and comprehensive descriptions of fine-grained visual content, yet caption quality spans factual accuracy, information coverage, and clarity. Compared with conventional methods that rely mainly on high-quality supervision or holistic rewards, rubric-based reinforcement learning decomposes these requirements into explicit criteria and provides targeted, structured feedback. However, existing methods often use separate models for caption generation, rubric construction, and judging, which may lead to inconsistent interpretations across roles. Some dynamic rubric methods alternate updates between the caption policy and rubric generator while keeping the judge fixed, but staged optimization may still leave rubric construction and judging out of step with policy optimization. We propose MoCo Rubric, a two-stage framework that coordinates these roles. First, role-conditioned, shared-parameter multi-task supervised fine-tuning equips a single vision--language model to serve as the Caption Policy, Rubric Generator, and Rubric Judge. Then, the Generator constructs rubrics online from captions sampled by the current Policy, reference captions, and image evidence. The Judge provides rubric-based rewards, and only the Policy receives GRPO updates. As Policy updates change the candidates being evaluated, we use an exponential moving average of the Policy parameters to update one momentum model shared by the Generator and Judge. This gradual transfer lets both rubric roles track Policy updates without separate RL optimization while smoothing parameter changes that could disrupt their rubric capabilities under direct synchronization. Across five captioning benchmarks, MoCo Rubric achieves an average pairwise win rate of 72.83%, the best mean rank in blind ranking, and the highest average score in caption-based question answering.
Ensuring the safety of reasoning large language models (LLMs) across languages is essential for their reliable deployment. However, when exposed to jailbreak attacks in non-high-resource languages, these models may generate unsafe responses even when their reasoning traces identify safety risks. To address this issue, we propose aligning cross-lingual thoughts and responses (ACTR), a framework that improves multilingual safety alignment by strengthening the use of existing safety reasoning. Specifically, we first present the think gap score (TGS) to compare the normalized contributions of reasoning traces to attention outputs during response generation across languages, and use reasoning- trace substitution to measure the cross-lingual safety gap. Next, using a corpus of jailbreak queries, we assess neuron importance through changes in response representations caused by neuron masking and compare the high-importance neuron sets obtained with reasoning enabled and disabled to identify safety think neurons that support the use of safety reasoning. Finally, we devise neuron-selective consistency optimization (NSCO), which uses a frozen judge model to reward agreement between the safety categories of reasoning traces and responses while updating only the parameters associated with the selected neurons, requiring no human-annotated responses or preference data. Across two reasoning models, ACTR achieves lower average attack success rates than the evaluated state-of-the-art methods on AdvBench-X and MultiJail, with safety gains extending to unseen languages, while preserving or improving average performance on multilingual knowledge and mathematical reasoning tasks and limiting false refusals of benign requests. Warning: this paper contains examples with unsafe content.
展开完整摘要收起摘要↓
Ensuring the safety of reasoning large language models (LLMs) across languages is essential for their reliable deployment. However, when exposed to jailbreak attacks in non-high-resource languages, these models may generate unsafe responses even when their reasoning traces identify safety risks. To address this issue, we propose aligning cross-lingual thoughts and responses (ACTR), a framework that improves multilingual safety alignment by strengthening the use of existing safety reasoning. Specifically, we first present the think gap score (TGS) to compare the normalized contributions of reasoning traces to attention outputs during response generation across languages, and use reasoning- trace substitution to measure the cross-lingual safety gap. Next, using a corpus of jailbreak queries, we assess neuron importance through changes in response representations caused by neuron masking and compare the high-importance neuron sets obtained with reasoning enabled and disabled to identify safety think neurons that support the use of safety reasoning. Finally, we devise neuron-selective consistency optimization (NSCO), which uses a frozen judge model to reward agreement between the safety categories of reasoning traces and responses while updating only the parameters associated with the selected neurons, requiring no human-annotated responses or preference data. Across two reasoning models, ACTR achieves lower average attack success rates than the evaluated state-of-the-art methods on AdvBench-X and MultiJail, with safety gains extending to unseen languages, while preserving or improving average performance on multilingual knowledge and mathematical reasoning tasks and limiting false refusals of benign requests. Warning: this paper contains examples with unsafe content.
Group Relative Policy Optimization (GRPO) is widely used to train reasoning language models, where it computes advantages by centering and normalizing rewards across rollouts of the same prompt. For multiple rewards, GRPO sums the reward components and normalizes the total reward by its within-group standard deviation. The corresponding variance equals the sum of all pairwise reward covariances. For a fixed centered reward, larger aggregate covariance produces smaller advantages, and vice versa, allowing update magnitudes to adapt to reward dependence. However, correlated rewards with large scales can dominate this normalization and suppress signals from smaller-scale rewards. We propose Correlation-Normalized GRPO (CorrGRPO), which normalizes pairwise covariances into Pearson correlation coefficients. CorrGRPO keeps the centered total reward unchanged while balancing the influence of differently scaled rewards on the correlation-based normalization. This allows advantage magnitudes to adapt to reward correlations without the normalization being dominated by large-scale reward components. We compare CorrGRPO with GRPO and other variants on code generation, tool calling, and agent security, using models ranging from 0.5B to 8B parameters. These tasks all involve multiple rewards that can improve together or present tradeoffs. Results show improvements across three domains, including code generation, tool calling, and agent security. Our code is available at https://github.com/HKUST-KnowComp/CorrGRPO.
展开完整摘要收起摘要↓
Group Relative Policy Optimization (GRPO) is widely used to train reasoning language models, where it computes advantages by centering and normalizing rewards across rollouts of the same prompt. For multiple rewards, GRPO sums the reward components and normalizes the total reward by its within-group standard deviation. The corresponding variance equals the sum of all pairwise reward covariances. For a fixed centered reward, larger aggregate covariance produces smaller advantages, and vice versa, allowing update magnitudes to adapt to reward dependence. However, correlated rewards with large scales can dominate this normalization and suppress signals from smaller-scale rewards. We propose Correlation-Normalized GRPO (CorrGRPO), which normalizes pairwise covariances into Pearson correlation coefficients. CorrGRPO keeps the centered total reward unchanged while balancing the influence of differently scaled rewards on the correlation-based normalization. This allows advantage magnitudes to adapt to reward correlations without the normalization being dominated by large-scale reward components. We compare CorrGRPO with GRPO and other variants on code generation, tool calling, and agent security, using models ranging from 0.5B to 8B parameters. These tasks all involve multiple rewards that can improve together or present tradeoffs. Results show improvements across three domains, including code generation, tool calling, and agent security. Our code is available at https://github.com/HKUST-KnowComp/CorrGRPO.
Style transfer aims to render target content in the style of a reference image, but existing methods often suffer from content leakage, where objects, layouts, or semantics from the style reference appear in the generated output. Although prior data-driven and training-free methods can reduce leakage, they often face a leakage-degradation dilemma: stronger content suppression may weaken style fidelity, while richer style preservation may reintroduce unwanted reference content. We identify this dilemma across the full style-transfer pipeline, including feature separation, feature-space grounding, and diffusion generation. To address these issues, we propose CLeaR, a training-free framework for content-leakage-resistant style transfer. CLeaR first uses Orthogonal Subspace Projection to define content-reduced style targets in each vision foundation model (VFM) feature space. It then performs Ensemble Inversion, which optimizes a shared pixel-space style anchor satisfying style constraints across multiple VFMs. Finally, Energy-Guided Calibration maintains style alignment during diffusion sampling by steering the denoising trajectory toward the ensemble-defined style manifold. We further provide a theoretical analysis showing that the style-anchor estimation error decreases with the number of VFMs. Experiments on StyleBench demonstrate that CLeaR improves style alignment, reduces content leakage, and achieves better LLM-as-Judge evaluation compared with existing methods. The code is available at \href{https://github.com/0606zt/CLeaR}{https://github.com/0606zt/CLeaR}.
展开完整摘要收起摘要↓
Style transfer aims to render target content in the style of a reference image, but existing methods often suffer from content leakage, where objects, layouts, or semantics from the style reference appear in the generated output. Although prior data-driven and training-free methods can reduce leakage, they often face a leakage-degradation dilemma: stronger content suppression may weaken style fidelity, while richer style preservation may reintroduce unwanted reference content. We identify this dilemma across the full style-transfer pipeline, including feature separation, feature-space grounding, and diffusion generation. To address these issues, we propose CLeaR, a training-free framework for content-leakage-resistant style transfer. CLeaR first uses Orthogonal Subspace Projection to define content-reduced style targets in each vision foundation model (VFM) feature space. It then performs Ensemble Inversion, which optimizes a shared pixel-space style anchor satisfying style constraints across multiple VFMs. Finally, Energy-Guided Calibration maintains style alignment during diffusion sampling by steering the denoising trajectory toward the ensemble-defined style manifold. We further provide a theoretical analysis showing that the style-anchor estimation error decreases with the number of VFMs. Experiments on StyleBench demonstrate that CLeaR improves style alignment, reduces content leakage, and achieves better LLM-as-Judge evaluation compared with existing methods. The code is available at \href{https://github.com/0606zt/CLeaR}{https://github.com/0606zt/CLeaR}.
作者Hanwen Lu, Jun He, Mingjia Yang, Hao Wei, Jinhao Huang, Yi Lin, Xiang Zhang
Historical street-view imagery records urban evolution, but uneven coverage leaves substantial gaps in historical records. Generating plausible past appearances requires restoring changed structures while preserving persistent scene content. We construct VIGOR-his, a decade-spanning cross-view dataset containing 43,653 location-level quadruplets across 11 cities on three continents. Its automated pipeline performs spatial pairing, consistency screening, change classification, and the generation and validation of satellite-based change descriptions and local editing instructions. Based on VIGOR-his, we propose CrossTimeEdit, a model that reformulates historical street-view generation as editing, using recent street views to constrain viewpoint and unchanged appearance and temporal satellite differences as change evidence. Starting from FLUX.2 [Klein] 4B, we train CrossTimeEdit through supervised fine-tuning (SFT) followed by online reinforcement learning (RL). We design three street-view editing criteria, namely Instruction Alignment (IA), Background Preservation (BP), and Quality and Physical Plausibility (QP), as both RL reward dimensions and evaluation metrics. We optimize this multi-reward objective using Within Group Relative Policy Optimization for flow-matching models (Flow-GRPO) with Group reward-Decoupled Normalization Policy Optimization (GDPO), which normalizes each reward dimension before aggregation. CrossTimeEdit improves overall performance across the three editing criteria by 17.12% over the pretrained baseline and outperforms cross-view generation models in scene consistency, visual realism, and perceptual quality. The implementation code, dataset, and model weights are available at https://luhanwen67.github.io/CrossTimeEdit-release/.
展开完整摘要收起摘要↓
Historical street-view imagery records urban evolution, but uneven coverage leaves substantial gaps in historical records. Generating plausible past appearances requires restoring changed structures while preserving persistent scene content. We construct VIGOR-his, a decade-spanning cross-view dataset containing 43,653 location-level quadruplets across 11 cities on three continents. Its automated pipeline performs spatial pairing, consistency screening, change classification, and the generation and validation of satellite-based change descriptions and local editing instructions. Based on VIGOR-his, we propose CrossTimeEdit, a model that reformulates historical street-view generation as editing, using recent street views to constrain viewpoint and unchanged appearance and temporal satellite differences as change evidence. Starting from FLUX.2 [Klein] 4B, we train CrossTimeEdit through supervised fine-tuning (SFT) followed by online reinforcement learning (RL). We design three street-view editing criteria, namely Instruction Alignment (IA), Background Preservation (BP), and Quality and Physical Plausibility (QP), as both RL reward dimensions and evaluation metrics. We optimize this multi-reward objective using Within Group Relative Policy Optimization for flow-matching models (Flow-GRPO) with Group reward-Decoupled Normalization Policy Optimization (GDPO), which normalizes each reward dimension before aggregation. CrossTimeEdit improves overall performance across the three editing criteria by 17.12% over the pretrained baseline and outperforms cross-view generation models in scene consistency, visual realism, and perceptual quality. The implementation code, dataset, and model weights are available at https://luhanwen67.github.io/CrossTimeEdit-release/.