Reliable self-assessment is essential for large language models (LLMs), yet they often remain highly confident when their answers are wrong. We study concurrent confidence calibration, where confidence is learned alongside capability improvement rather than calibrated only after training. Reinforcement learning from verifiable rewards (RLVR) provides a natural setting for this paradigm, as it continuously produces responses paired with verifiable correctness feedback. Existing concurrent methods, however, learn both capability and confidence through reinforcement learning within shared policy parameters, potentially coupling two fundamentally different learning problems. We instead propose shared experience but separate learning: capability and confidence learn from the same trajectories, but through separate optimization mechanisms and parameters. Based on this principle, we introduce CoCal (Companion Confidence Calibration), which trains a lightweight companion from rollout hidden states and verifier-derived correctness supervision while leaving task optimization unchanged. Experiments on Qwen3-8B and Qwen3-14B show that CoCal improves confidence estimation without sacrificing task performance, outperforming both RL-based concurrent methods and matched post-hoc calibration. The learned companion further generalizes across domains and policy shifts, while the benefits of CoCal persist at both scales.
展开完整摘要收起摘要↓
Reliable self-assessment is essential for large language models (LLMs), yet they often remain highly confident when their answers are wrong. We study concurrent confidence calibration, where confidence is learned alongside capability improvement rather than calibrated only after training. Reinforcement learning from verifiable rewards (RLVR) provides a natural setting for this paradigm, as it continuously produces responses paired with verifiable correctness feedback. Existing concurrent methods, however, learn both capability and confidence through reinforcement learning within shared policy parameters, potentially coupling two fundamentally different learning problems. We instead propose shared experience but separate learning: capability and confidence learn from the same trajectories, but through separate optimization mechanisms and parameters. Based on this principle, we introduce CoCal (Companion Confidence Calibration), which trains a lightweight companion from rollout hidden states and verifier-derived correctness supervision while leaving task optimization unchanged. Experiments on Qwen3-8B and Qwen3-14B show that CoCal improves confidence estimation without sacrificing task performance, outperforming both RL-based concurrent methods and matched post-hoc calibration. The learned companion further generalizes across domains and policy shifts, while the benefits of CoCal persist at both scales.
Language models are increasingly used as proposal models for black-box search, from program optimization to molecular design. Existing approaches typically improve proposals through iterative prompting or parameter updates, offering limited control over how completely and efficiently the model's search space is explored. Continuous optimization methods, such as Bayesian optimization, provide a principled way to search but require a suitable domain to operate over. To address this, we introduce BOReFT, which learns a compact, low-dimensional space of hidden-state interventions in a frozen language model, and uses this space as the search domain for Bayesian optimization with an external scoring function. Empirically, we find that the learned domain spans semantic regions and exhibits smoothness properties that support search. Theoretically, we show that semantic coverage and interpolation control the best score available in the learned space, and that decoding from this space yields a standard stochastic-bandit observation model for adaptive search. We evaluate BOReFT on the interpretable word search task "Semantle" and on three more real-world discovery tasks in de novo molecule property optimization. Compared to strong LLM baselines, BOReFT finds in Semantle a higher number of hidden targets and, on two out of three molecular objectives, achieves higher property scores. Consequently, our method provides a principled new bridge between discrete proposal spaces of LLM-based search and continuous black-box optimization.
展开完整摘要收起摘要↓
Language models are increasingly used as proposal models for black-box search, from program optimization to molecular design. Existing approaches typically improve proposals through iterative prompting or parameter updates, offering limited control over how completely and efficiently the model's search space is explored. Continuous optimization methods, such as Bayesian optimization, provide a principled way to search but require a suitable domain to operate over. To address this, we introduce BOReFT, which learns a compact, low-dimensional space of hidden-state interventions in a frozen language model, and uses this space as the search domain for Bayesian optimization with an external scoring function. Empirically, we find that the learned domain spans semantic regions and exhibits smoothness properties that support search. Theoretically, we show that semantic coverage and interpolation control the best score available in the learned space, and that decoding from this space yields a standard stochastic-bandit observation model for adaptive search. We evaluate BOReFT on the interpretable word search task "Semantle" and on three more real-world discovery tasks in de novo molecule property optimization. Compared to strong LLM baselines, BOReFT finds in Semantle a higher number of hidden targets and, on two out of three molecular objectives, achieves higher property scores. Consequently, our method provides a principled new bridge between discrete proposal spaces of LLM-based search and continuous black-box optimization.
Verified solutions are not equally useful for preparing reasoning models for reinforcement learning (RL). We present a comprehensive study of route diversity, the variation in the sequences of reasoning steps in supervised fine-tuning (SFT) data, and propose a lightweight, rule-based fingerprint to select for it. From one pool at one budget, with matched training recipes and checkpoints, selecting diverse rather than similar routes improves post-RL problem coverage across puzzles and mathematics, including on problems harder than those seen in either training stage. In synthetic experiments, route-diverse SFT improves OLMo3-7B's pass@8 by 16.9 points on environments held out from SFT. In a single-model condition, where one model writes every candidate, diverse selection gains up to 6.2 points of mean pass@8 across 10 mathematics benchmarks. Pre-RL diagnostics suggest why: diverse SFT can produce both successful and failed attempts on more prompts despite slightly lower mean accuracy, giving group-relative RL more prompts with a learning signal. On 3 open-source corpora, our CPU-only selector, without model calls, outperforms more expensive alternatives in every comparison of mean post-RL performance. These results identify reasoning-route diversity as a practical criterion for selecting SFT data that better prepares models for RL.
展开完整摘要收起摘要↓
Verified solutions are not equally useful for preparing reasoning models for reinforcement learning (RL). We present a comprehensive study of route diversity, the variation in the sequences of reasoning steps in supervised fine-tuning (SFT) data, and propose a lightweight, rule-based fingerprint to select for it. From one pool at one budget, with matched training recipes and checkpoints, selecting diverse rather than similar routes improves post-RL problem coverage across puzzles and mathematics, including on problems harder than those seen in either training stage. In synthetic experiments, route-diverse SFT improves OLMo3-7B's pass@8 by 16.9 points on environments held out from SFT. In a single-model condition, where one model writes every candidate, diverse selection gains up to 6.2 points of mean pass@8 across 10 mathematics benchmarks. Pre-RL diagnostics suggest why: diverse SFT can produce both successful and failed attempts on more prompts despite slightly lower mean accuracy, giving group-relative RL more prompts with a learning signal. On 3 open-source corpora, our CPU-only selector, without model calls, outperforms more expensive alternatives in every comparison of mean post-RL performance. These results identify reasoning-route diversity as a practical criterion for selecting SFT data that better prepares models for RL.
Large language model (LLM) agentic systems increasingly rely on models communicating with one another, yet existing uncertainty and multi-agent methods rarely estimate how a particular receiver will interpret a message before it is sent. This matters in heterogeneous systems, where capable receivers can reconstruct different tasks from the same message. We model this as a sender-receiver problem with a latent receiver type and define prospective interpretation risk (PIR): the probability that a receiver reconstructs a task other than intended. Rather than model an LLM's full input-output behaviour, we use black-box probes relating messages, intended tasks, and receiver-specific reconstructions, yielding scalable supervision while separating interpretation from downstream capability failure. Offline, heterogeneous frozen receivers provide supervision for receiver-conditioned risk and the effects of predefined mutable message features. At deployment, history induces a posterior over receiver types, guiding message revision and selection. We introduce value of interpretation information (VoII), querying for receiver information only when its expected communication benefit exceeds its cost. Our theory characterises when receiver information has decision value and bounds such queries. Empirically, interpretation-failure rates vary by 4-13x across receivers. Receiver information reduces PIR calibration error by 68% relative to a receiver-agnostic predictor, largely by correcting receiver-specific risk levels. PIR-guided revision reduces interpretation failure by 44% relative to the original message and 40% relative to a generic rewrite, mostly through a repair that helps every receiver. VoII outperforms information-gain and random querying at matched cost on the interpretation objective it optimises, lowering interpretation failure from 3.84% to 3.79% while querying 18.2% of episodes.
展开完整摘要收起摘要↓
Large language model (LLM) agentic systems increasingly rely on models communicating with one another, yet existing uncertainty and multi-agent methods rarely estimate how a particular receiver will interpret a message before it is sent. This matters in heterogeneous systems, where capable receivers can reconstruct different tasks from the same message. We model this as a sender-receiver problem with a latent receiver type and define prospective interpretation risk (PIR): the probability that a receiver reconstructs a task other than intended. Rather than model an LLM's full input-output behaviour, we use black-box probes relating messages, intended tasks, and receiver-specific reconstructions, yielding scalable supervision while separating interpretation from downstream capability failure. Offline, heterogeneous frozen receivers provide supervision for receiver-conditioned risk and the effects of predefined mutable message features. At deployment, history induces a posterior over receiver types, guiding message revision and selection. We introduce value of interpretation information (VoII), querying for receiver information only when its expected communication benefit exceeds its cost. Our theory characterises when receiver information has decision value and bounds such queries. Empirically, interpretation-failure rates vary by 4-13x across receivers. Receiver information reduces PIR calibration error by 68% relative to a receiver-agnostic predictor, largely by correcting receiver-specific risk levels. PIR-guided revision reduces interpretation failure by 44% relative to the original message and 40% relative to a generic rewrite, mostly through a repair that helps every receiver. VoII outperforms information-gain and random querying at matched cost on the interpretation objective it optimises, lowering interpretation failure from 3.84% to 3.79% while querying 18.2% of episodes.
The pursuit of recursive self-improvement (RSI) toward general intelligence is divided between macro-level language model scaling and the interaction-driven principles of "Era of Experience". Yet, any self-improving architecture ultimately rests upon its underlying optimization engine: if general intelligence requires learning from grounded interaction, the reinforcement learning (RL) update rule itself must be capable of cumulative adaptation. While algorithm self-discovery has produced Disco103 that surpassed PPO to achieve SOTA benchmark performance — its internal update machinery remains an uninspected black box. We present the first causal mechanistic audit of a self-discovered RL rule, structured directly around the five pillars of the Era of Experience: extended horizon, grounded reward scales, continuing streams, within-lifetime change, and exploration depth. By surgically pinning, freezing, and transplanting recurrent states while holding meta-parameters fixed, we test when learning history acts as an asset or a burden. Three findings organize the audit: (1) Recurrent history actively expands usable reward scales, sustaining a six-decade window versus three under zero-pinning. (2) Decoupling historical content from its maintenance reveals that the penalty of mismatched history stems from perpetual clamping; allowing imported state to evolve naturally attenuates this burden. (3) Under environmental change, controlling replay retention reverses the apparent adaptation advantage over DQN, demonstrating that external data turnover can confound internal plasticity. Validated through capability thresholds and ported to a second rule (OPEN), this work grounds macro-RSI ambitions in micro-level learning dynamics, establishing a foundational audit standard for next-generation, self-evolving RL algorithms.
展开完整摘要收起摘要↓
The pursuit of recursive self-improvement (RSI) toward general intelligence is divided between macro-level language model scaling and the interaction-driven principles of "Era of Experience". Yet, any self-improving architecture ultimately rests upon its underlying optimization engine: if general intelligence requires learning from grounded interaction, the reinforcement learning (RL) update rule itself must be capable of cumulative adaptation. While algorithm self-discovery has produced Disco103 that surpassed PPO to achieve SOTA benchmark performance — its internal update machinery remains an uninspected black box. We present the first causal mechanistic audit of a self-discovered RL rule, structured directly around the five pillars of the Era of Experience: extended horizon, grounded reward scales, continuing streams, within-lifetime change, and exploration depth. By surgically pinning, freezing, and transplanting recurrent states while holding meta-parameters fixed, we test when learning history acts as an asset or a burden. Three findings organize the audit: (1) Recurrent history actively expands usable reward scales, sustaining a six-decade window versus three under zero-pinning. (2) Decoupling historical content from its maintenance reveals that the penalty of mismatched history stems from perpetual clamping; allowing imported state to evolve naturally attenuates this burden. (3) Under environmental change, controlling replay retention reverses the apparent adaptation advantage over DQN, demonstrating that external data turnover can confound internal plasticity. Validated through capability thresholds and ported to a second rule (OPEN), this work grounds macro-RSI ambitions in micro-level learning dynamics, establishing a foundational audit standard for next-generation, self-evolving RL algorithms.
Can a genome model retain a detectable record of the synthetic sequences used to train it? We study watermark inheritance through distillation with GenoTrace, a codon-aware extension of green-list watermarking. Two token-level factors modulate the teacher's generation bias using codon position and organism-specific codon usage. The resulting sequences train a smaller student, whose outputs are audited without an active watermark processor. In a three-seed GenomeOcean-500M-to-100M experiment, the joint configuration achieves a mean audit score of 17.88 and 94.5% detection at a fixed threshold. It retains 49.0% detection after key-aware token substitution, compared with 0% for the available single-seed plain-watermark comparator, and 47.0% after combined mechanism-targeted nucleotide edits. Additional experiments establish inherited signal across five organism-conditioned datasets and teacher-student size ratios up to 40. Component ablations and computational sequence-quality assays reveal distinct operating points for detection strength and coding coverage. GenoTrace provides a practical token-level construction and an empirical account of how genomic structure shapes inherited watermark signals. The findings concern shared-tokenizer distillation and the tested editing procedures, with calibration and biological utility treated as separate evaluation requirements.
展开完整摘要收起摘要↓
Can a genome model retain a detectable record of the synthetic sequences used to train it? We study watermark inheritance through distillation with GenoTrace, a codon-aware extension of green-list watermarking. Two token-level factors modulate the teacher's generation bias using codon position and organism-specific codon usage. The resulting sequences train a smaller student, whose outputs are audited without an active watermark processor. In a three-seed GenomeOcean-500M-to-100M experiment, the joint configuration achieves a mean audit score of 17.88 and 94.5% detection at a fixed threshold. It retains 49.0% detection after key-aware token substitution, compared with 0% for the available single-seed plain-watermark comparator, and 47.0% after combined mechanism-targeted nucleotide edits. Additional experiments establish inherited signal across five organism-conditioned datasets and teacher-student size ratios up to 40. Component ablations and computational sequence-quality assays reveal distinct operating points for detection strength and coding coverage. GenoTrace provides a practical token-level construction and an empirical account of how genomic structure shapes inherited watermark signals. The findings concern shared-tokenizer distillation and the tested editing procedures, with calibration and biological utility treated as separate evaluation requirements.
作者Xiaoyu Xu, Zi Liang, Minxin Du, Qipeng Xie, Qingqing Ye, Yuyuan Li, Haibo Hu
Tool-using language-model agents select and execute third-party artifacts. Different implementations can return the requested output while producing hidden execution effects that task-, attack-, or choice-based evaluations may miss. We study functional counterfeits: implementations that match benign alternatives on the requested output but add an effect forbidden by the task contract. We introduce SINGED (Source Integrity and the Nonidentifiability Gap in Execution Decisions for LLM Agents), a controlled benchmark covering five primary and two held-out task families. It varies displayed rank, evidence depth, decision policy, model release, and agent configuration, while task and process oracles verify the artifact and execution path. Across 7,549 audited trials, the randomized-rank study finds counterfeit execution in 45% (27/60) of rank-one trials and none at later ranks. Cross-candidate comparison eliminates shallow failures and reduces layered failures from 15.7% to 4.2%, but leaves dependency failures; its benefit is uncertain on unseen effects and public-package structures. Moreover, seven releases with no counterfeit executions when benign alternatives are available execute the counterfeit in 55/175 single-source cells after alternatives are removed. SINGED thus exposes a rank-, evidence-, and choice-sensitive outcome-to-execution gap: evaluation must connect correct outputs to execution paths.
展开完整摘要收起摘要↓
Tool-using language-model agents select and execute third-party artifacts. Different implementations can return the requested output while producing hidden execution effects that task-, attack-, or choice-based evaluations may miss. We study functional counterfeits: implementations that match benign alternatives on the requested output but add an effect forbidden by the task contract. We introduce SINGED (Source Integrity and the Nonidentifiability Gap in Execution Decisions for LLM Agents), a controlled benchmark covering five primary and two held-out task families. It varies displayed rank, evidence depth, decision policy, model release, and agent configuration, while task and process oracles verify the artifact and execution path. Across 7,549 audited trials, the randomized-rank study finds counterfeit execution in 45% (27/60) of rank-one trials and none at later ranks. Cross-candidate comparison eliminates shallow failures and reduces layered failures from 15.7% to 4.2%, but leaves dependency failures; its benefit is uncertain on unseen effects and public-package structures. Moreover, seven releases with no counterfeit executions when benign alternatives are available execute the counterfeit in 55/175 single-source cells after alternatives are removed. SINGED thus exposes a rank-, evidence-, and choice-sensitive outcome-to-execution gap: evaluation must connect correct outputs to execution paths.
Genome foundation models are growing into sparse mixture-of-experts (MoE) networks whose expert weights no longer fit on the machines that hold the sequences, yet sending a private genome to rented accelerators exposes it: we show that a single server hosting one expert recovers the input nucleotides with 99.8% top-1 accuracy. We present CipherGenome, a protocol that keeps the embedding, attention and router of a 15.1B-parameter MoE genome model on a trusted thin client and outsources every expert projection, 95.8% of the parameters, to untrusted and possibly colluding GPU servers under module-LWE encryption. The design exploits three structural facts: expert layers are linear between two SwiGLU gates, expert weights are public, and GPU integer tensor cores can evaluate a ciphertext-weight product exactly modulo $2^{48}$ in a single GEMM. The client evaluates the nonlinearity exactly and re-encrypts with fresh secrets, so no polynomial approximation or bootstrapping is ever needed. On 72 windows from 12 bacterial genomes, encryption adds $2.54 \times 10^{-4}$ nats per token of KL divergence (95% CI upper bound $3.95 \times 10^{-4}$), below a pre-registered non-inferiority margin and indistinguishable from bf16 inference, while the same inversion attack falls to chance level. A reusable public hint cuts end-to-end latency by 3.54 times, wire compression reduces traffic 6.8 times, per-layer padding reduces routing leakage from 54.9% to 8.9% accuracy, and HE-compatible int4 experts remain non-inferior to their plaintext counterparts. Per expert and token, the server-side cost is more than six orders of magnitude below a CKKS baseline.
展开完整摘要收起摘要↓
Genome foundation models are growing into sparse mixture-of-experts (MoE) networks whose expert weights no longer fit on the machines that hold the sequences, yet sending a private genome to rented accelerators exposes it: we show that a single server hosting one expert recovers the input nucleotides with 99.8% top-1 accuracy. We present CipherGenome, a protocol that keeps the embedding, attention and router of a 15.1B-parameter MoE genome model on a trusted thin client and outsources every expert projection, 95.8% of the parameters, to untrusted and possibly colluding GPU servers under module-LWE encryption. The design exploits three structural facts: expert layers are linear between two SwiGLU gates, expert weights are public, and GPU integer tensor cores can evaluate a ciphertext-weight product exactly modulo $2^{48}$ in a single GEMM. The client evaluates the nonlinearity exactly and re-encrypts with fresh secrets, so no polynomial approximation or bootstrapping is ever needed. On 72 windows from 12 bacterial genomes, encryption adds $2.54 \times 10^{-4}$ nats per token of KL divergence (95% CI upper bound $3.95 \times 10^{-4}$), below a pre-registered non-inferiority margin and indistinguishable from bf16 inference, while the same inversion attack falls to chance level. A reusable public hint cuts end-to-end latency by 3.54 times, wire compression reduces traffic 6.8 times, per-layer padding reduces routing leakage from 54.9% to 8.9% accuracy, and HE-compatible int4 experts remain non-inferior to their plaintext counterparts. Per expert and token, the server-side cost is more than six orders of magnitude below a CKKS baseline.
Genome foundation models are most useful where sequences are generated, yet the largest models need datacenter accelerators and a place to send private DNA. We ask whether a 15-billion-parameter genome mixture-of-experts (MoE) model can instead run on volunteers' web browsers, with the experts spread across many untrusted devices, without changing its predictions and without revealing the sequence to any single device. We build a system in which a trusted coordinator runs attention and routing while browser workers run every expert feed-forward network through hand-written WebGPU kernels, and we protect the expert inputs with real-valued Lagrange coded computing: each worker receives only a Gaussian-padded share, computes the expert's linear maps, and the coordinator decodes from any two of three workers. On GenomeOcean-MoE (8 experts, top-2 routing, 24 layers), the browser path matches native llama.cpp at every quantization level, the distributed path stays at the BF16 numerical noise floor (KL 0.0036 nats per token), and an unfitted latency model predicts decode time within 0.74% (median) under emulated wide-area links. We first show that plaintext expert inputs are not private: a probe recovers the token from a single vector at every depth, and one worker can identify the source genome from 300 unordered tokens with 92% accuracy. With coded experts, an adaptive attacker trained on shares falls to the most-frequent-token baseline, one worker's information about each token is bounded below one bit per forward pass, and the fidelity cost stays below the BF16 noise floor; in Chrome, coded decoding runs at 220 to 376 ms per token, depending on how much of the routing is hidden, and continues without replicas when a worker fails.
展开完整摘要收起摘要↓
Genome foundation models are most useful where sequences are generated, yet the largest models need datacenter accelerators and a place to send private DNA. We ask whether a 15-billion-parameter genome mixture-of-experts (MoE) model can instead run on volunteers' web browsers, with the experts spread across many untrusted devices, without changing its predictions and without revealing the sequence to any single device. We build a system in which a trusted coordinator runs attention and routing while browser workers run every expert feed-forward network through hand-written WebGPU kernels, and we protect the expert inputs with real-valued Lagrange coded computing: each worker receives only a Gaussian-padded share, computes the expert's linear maps, and the coordinator decodes from any two of three workers. On GenomeOcean-MoE (8 experts, top-2 routing, 24 layers), the browser path matches native llama.cpp at every quantization level, the distributed path stays at the BF16 numerical noise floor (KL 0.0036 nats per token), and an unfitted latency model predicts decode time within 0.74% (median) under emulated wide-area links. We first show that plaintext expert inputs are not private: a probe recovers the token from a single vector at every depth, and one worker can identify the source genome from 300 unordered tokens with 92% accuracy. With coded experts, an adaptive attacker trained on shares falls to the most-frequent-token baseline, one worker's information about each token is bounded below one bit per forward pass, and the fidelity cost stays below the BF16 noise floor; in Chrome, coded decoding runs at 220 to 376 ms per token, depending on how much of the routing is hidden, and continues without replicas when a worker fails.
Modern transformer models excel at capturing semantic relationships through sentence embeddings, yet their ability to perform pragmatic reasoning remains understudied. This paper investigates whether encoder-based transformers such as DeBERTa employ the M-Heuristic (the neo-Gricean principle that marked linguistic forms implicate marked meanings). We test this hypothesis by contrasting lexical causatives (e.g., ``Larry stopped the car'') with periphrastic causatives (e.g., ``Larry caused the car to stop'') using a Natural Language Inference framework. Our experiments across 188 conditions with 15 ambitransitive verbs reveal that DeBERTa, RoBERTa, and BART show no evidence of capturing the pragmatic distinction between these forms, with DeBERTa predicting ``Neutral'' for 100% of cases. Probing analysis initially suggested a representation-use dissociation, but control experiments reveal the probe was tracking syntactic complexity, not causative pragmatics. Semantic similarity over 30 triplets places periphrastic causatives closer to unmediated manner descriptions in 29/30 cases, opposite to M-Heuristic predictions in the embedding space. Under explicit metalinguistic framing, Gemini Flash-Lite reaches 100% with item-specific traces, so the principle is available under instruction yet unused in default NLI.
展开完整摘要收起摘要↓
Modern transformer models excel at capturing semantic relationships through sentence embeddings, yet their ability to perform pragmatic reasoning remains understudied. This paper investigates whether encoder-based transformers such as DeBERTa employ the M-Heuristic (the neo-Gricean principle that marked linguistic forms implicate marked meanings). We test this hypothesis by contrasting lexical causatives (e.g., ``Larry stopped the car'') with periphrastic causatives (e.g., ``Larry caused the car to stop'') using a Natural Language Inference framework. Our experiments across 188 conditions with 15 ambitransitive verbs reveal that DeBERTa, RoBERTa, and BART show no evidence of capturing the pragmatic distinction between these forms, with DeBERTa predicting ``Neutral'' for 100% of cases. Probing analysis initially suggested a representation-use dissociation, but control experiments reveal the probe was tracking syntactic complexity, not causative pragmatics. Semantic similarity over 30 triplets places periphrastic causatives closer to unmediated manner descriptions in 29/30 cases, opposite to M-Heuristic predictions in the embedding space. Under explicit metalinguistic framing, Gemini Flash-Lite reaches 100% with item-specific traces, so the principle is available under instruction yet unused in default NLI.
作者Hongwei Niu, Yunpeng Luo, Hanjun Li, Ziyin Zhou, Jianghang Lin, Ke Yan, Shouhong Ding, Shengchuan Zhang, Liujuan Cao
The rapid proliferation of highly realistic AI-Generated Content (AIGC) necessitates robust and interpretable detection mechanisms. However, existing detectors are predominantly confined to single modalities and provide binary outputs without reasoning. While Multimodal Large Language Models (MLLMs) present a promising solution, their development is constrained by the scarcity of multimodal reasoning data and the reasoning-detection optimization dilemma, where explicit reasoning supervision can compromise detection accuracy. To this end, we introduce IVT-Set, a comprehensive dataset comprising over 152K diverse image, video, and text samples equipped with multi-granularity Chain-of-Thought (CoT) reasoning trajectories. Based on it, we propose IVT-Guard, a pioneering framework for unified and interpretable AIGC detection across image, video, and text modalities. Furthermore, to overcome the aforementioned optimization dilemma, we design a novel three-stage training paradigm: Artifact-Aware Pre-training, Artifact-to-Evidence Supervised Fine-Tuning via artifact-aware injection, and Evidence-Verdict Consistency Group Relative Policy Optimization. Extensive experiments demonstrate that IVT-Guard achieves state-of-the-art detection performance across in-domain, out-of-domain, and cross-dataset settings while delivering faithful reasoning. Code and data will be released.
展开完整摘要收起摘要↓
The rapid proliferation of highly realistic AI-Generated Content (AIGC) necessitates robust and interpretable detection mechanisms. However, existing detectors are predominantly confined to single modalities and provide binary outputs without reasoning. While Multimodal Large Language Models (MLLMs) present a promising solution, their development is constrained by the scarcity of multimodal reasoning data and the reasoning-detection optimization dilemma, where explicit reasoning supervision can compromise detection accuracy. To this end, we introduce IVT-Set, a comprehensive dataset comprising over 152K diverse image, video, and text samples equipped with multi-granularity Chain-of-Thought (CoT) reasoning trajectories. Based on it, we propose IVT-Guard, a pioneering framework for unified and interpretable AIGC detection across image, video, and text modalities. Furthermore, to overcome the aforementioned optimization dilemma, we design a novel three-stage training paradigm: Artifact-Aware Pre-training, Artifact-to-Evidence Supervised Fine-Tuning via artifact-aware injection, and Evidence-Verdict Consistency Group Relative Policy Optimization. Extensive experiments demonstrate that IVT-Guard achieves state-of-the-art detection performance across in-domain, out-of-domain, and cross-dataset settings while delivering faithful reasoning. Code and data will be released.
LLM-based agents can improve themselves across tasks by reusing and revising the skills they orchestrate into executable procedures. Flow-based training fits this loop: it samples procedures in proportion to reward, and the flow through each skill credits it for the next library revision. Three obstacles stand in the way of making this self-improvement reliable: flow training suffers strategy collapse over tree-structured histories; nonnegative flow-based credit rewards frequent use as if it were benefit; and library edits rest on the task reward the policy optimizes. We introduce R$^2$ Flow, a recursive self-improvement framework that alternates policy learning, independent verification, and versioned skill-library updates on a shared-state orchestration graph. The graph merges histories that differ only in the order of independent steps, allowing flow training to pool evidence across equivalent executions. A flow-share readout of the trained flow, invariant to the backward policy, and a separate signed utility rank which skills to change, verifier evidence decides whether an edit is warranted, and a residual-variance plateau sets when to update. Committed edits reshape the graph the next policy learns on, realizing recursive skill evolution. Across question answering, mathematical reasoning, interactive decision making, and code generation, R$^2$ Flow improves task accuracy and library-edit precision over heuristic orchestration, reinforcement learning, and skill-evolution baselines, and transfers across executors. Code is available at https://github.com/beita6969/r2flow.
展开完整摘要收起摘要↓
LLM-based agents can improve themselves across tasks by reusing and revising the skills they orchestrate into executable procedures. Flow-based training fits this loop: it samples procedures in proportion to reward, and the flow through each skill credits it for the next library revision. Three obstacles stand in the way of making this self-improvement reliable: flow training suffers strategy collapse over tree-structured histories; nonnegative flow-based credit rewards frequent use as if it were benefit; and library edits rest on the task reward the policy optimizes. We introduce R$^2$ Flow, a recursive self-improvement framework that alternates policy learning, independent verification, and versioned skill-library updates on a shared-state orchestration graph. The graph merges histories that differ only in the order of independent steps, allowing flow training to pool evidence across equivalent executions. A flow-share readout of the trained flow, invariant to the backward policy, and a separate signed utility rank which skills to change, verifier evidence decides whether an edit is warranted, and a residual-variance plateau sets when to update. Committed edits reshape the graph the next policy learns on, realizing recursive skill evolution. Across question answering, mathematical reasoning, interactive decision making, and code generation, R$^2$ Flow improves task accuracy and library-edit precision over heuristic orchestration, reinforcement learning, and skill-evolution baselines, and transfers across executors. Code is available at https://github.com/beita6969/r2flow.
High-performance GPU kernels are essential to modern machine learning systems, yet automatically generating kernels that are both correct and efficient remains challenging. Existing LLM-based approaches face two major limitations: the scarcity of high-quality training data aligned with the model's current capabilities, and the inherent trade-off between kernel correctness and performance. To address these challenges, we propose KernelZero, a co-evolution framework that continuously improves GPU kernel generation through two specialized models: a Proposer that generates Torch modules from API sets and a Coder that translates them into CUDA or Triton kernels. KernelZero uses a frontier-driven module generation mechanism to continuously produce capability-aligned training modules based on the Coder's current weaknesses. It further introduces Correctness-Aware Group Relative Policy Optimization (CA-GRPO), which optimizes performance only after correctness becomes sufficiently reliable. By alternating the optimization of the Proposer and Coder, KernelZero forms an automatic curriculum that enables targeted and training-efficient capability improvement. Empirically, KernelZero-7B surpasses Claude-4.5-Sonnet on CUDA and DeepSeek-V4-Pro on Triton. On KernelBench Level 1 and 2, it achieves CUDA pass@1 scores of 75.8% and 69.6%, respectively, with pass@10 reaching 100% and 97%. On Triton, it achieves pass@1 scores of 77.2% and 72.5%, respectively.
展开完整摘要收起摘要↓
High-performance GPU kernels are essential to modern machine learning systems, yet automatically generating kernels that are both correct and efficient remains challenging. Existing LLM-based approaches face two major limitations: the scarcity of high-quality training data aligned with the model's current capabilities, and the inherent trade-off between kernel correctness and performance. To address these challenges, we propose KernelZero, a co-evolution framework that continuously improves GPU kernel generation through two specialized models: a Proposer that generates Torch modules from API sets and a Coder that translates them into CUDA or Triton kernels. KernelZero uses a frontier-driven module generation mechanism to continuously produce capability-aligned training modules based on the Coder's current weaknesses. It further introduces Correctness-Aware Group Relative Policy Optimization (CA-GRPO), which optimizes performance only after correctness becomes sufficiently reliable. By alternating the optimization of the Proposer and Coder, KernelZero forms an automatic curriculum that enables targeted and training-efficient capability improvement. Empirically, KernelZero-7B surpasses Claude-4.5-Sonnet on CUDA and DeepSeek-V4-Pro on Triton. On KernelBench Level 1 and 2, it achieves CUDA pass@1 scores of 75.8% and 69.6%, respectively, with pass@10 reaching 100% and 97%. On Triton, it achieves pass@1 scores of 77.2% and 72.5%, respectively.
Classical texts aligned with their translations support machine translation, retrieval and computational research, but evidence comparing alignment workflows is scattered. This study compares seven systems on 452 texts in Pali, Sanskrit, Mishnaic Hebrew and Tibetan, comprising 9,833 human-aligned units: four embedding pipelines, a direct LLM call, an autonomous agent, and the agent revised by an independent auditor. Generative workflows recover 93-94% of reference correspondences, against at most 77% for embeddings. A ceiling analysis shows that sentence boundaries make some references unrepresentable by the embedding pipelines. Reference recovery is similar across generative workflows: the agent's advantage is 0.5 percentage points (95% CI -0.02 to 1.17), and auditing adds no established benefit. Agents nevertheless produce structurally valid output for all 452 texts, against 437 for direct calls. A blinded three-LLM panel assesses every generative mismatch against the source and human reference. Most mismatches are labelled defensible editorial variation; consensus major-error labels cover only 0.06-0.14% of units. The panel labels significantly fewer residual defects for agents than direct calls (0.7% versus 1.4%), suggesting that reference recovery alone understates alignment quality. On ten long Pali discourses taken as published online, agents and audited agents raise recovery from the direct call's 71% to 84% and 92%. Identical reference-located chunks bring all three to 93%. Agents thus improve structural reliability and reduce judged defects on short passages, while their large recovery advantage on long documents disappears after chunking. In this setting, independent auditing offers little measurable additional benefit on prepared passages.
展开完整摘要收起摘要↓
Classical texts aligned with their translations support machine translation, retrieval and computational research, but evidence comparing alignment workflows is scattered. This study compares seven systems on 452 texts in Pali, Sanskrit, Mishnaic Hebrew and Tibetan, comprising 9,833 human-aligned units: four embedding pipelines, a direct LLM call, an autonomous agent, and the agent revised by an independent auditor. Generative workflows recover 93-94% of reference correspondences, against at most 77% for embeddings. A ceiling analysis shows that sentence boundaries make some references unrepresentable by the embedding pipelines. Reference recovery is similar across generative workflows: the agent's advantage is 0.5 percentage points (95% CI -0.02 to 1.17), and auditing adds no established benefit. Agents nevertheless produce structurally valid output for all 452 texts, against 437 for direct calls. A blinded three-LLM panel assesses every generative mismatch against the source and human reference. Most mismatches are labelled defensible editorial variation; consensus major-error labels cover only 0.06-0.14% of units. The panel labels significantly fewer residual defects for agents than direct calls (0.7% versus 1.4%), suggesting that reference recovery alone understates alignment quality. On ten long Pali discourses taken as published online, agents and audited agents raise recovery from the direct call's 71% to 84% and 92%. Identical reference-located chunks bring all three to 93%. Agents thus improve structural reliability and reduce judged defects on short passages, while their large recovery advantage on long documents disappears after chunking. In this setting, independent auditing offers little measurable additional benefit on prepared passages.
Group-relative reinforcement learning (RL) relies on reward variation among sampled responses to estimate informative relative advantages. As language models become increasingly capable, existing training data can become reward-saturated: all sampled responses to the same problem might receive equally high rewards, where the group-relative learning signals vanish and leave previously useful data obsolete. In this work, we investigate whether useful learning signals can be recovered from such saturated data. We study interventions at four levels of group-policy RL pipelines---data, rollout, reward, and advantage---and conduct extensive RL training on saturated reasoning data only. While standard GRPO on saturated data would almost always yield near-0 advantages and near-noise signals, diverse interventions successfully recycle and repurpose such data: among the proposed strategies, interventions at rollout generation are consistently most effective: nudging the policy to generate ``high-quality'', incorrect solutions introduces rollouts with poor rewards into saturated groups as negative samples, which turns out to improve GRPO by 6.4% to 9.0% across Qwen3-1.7B and 4B. Other interventions such as increasing rollout temperature or adding auxiliary rewards can also restore non-zero advantages, but yield less consistent gains. Further analyses show that effective negative rollouts require informative negative trajectories, that the method remains effective alongside unsaturated data, and that it supports iterative recycling of newly saturated examples. While increasingly stronger LLMs would render more data as saturated, our results demonstrate that don't waste your saturated data: with the right strategies they can be recycled into useful RL training signals in an increasingly data-scarce world.
展开完整摘要收起摘要↓
Group-relative reinforcement learning (RL) relies on reward variation among sampled responses to estimate informative relative advantages. As language models become increasingly capable, existing training data can become reward-saturated: all sampled responses to the same problem might receive equally high rewards, where the group-relative learning signals vanish and leave previously useful data obsolete. In this work, we investigate whether useful learning signals can be recovered from such saturated data. We study interventions at four levels of group-policy RL pipelines---data, rollout, reward, and advantage---and conduct extensive RL training on saturated reasoning data only. While standard GRPO on saturated data would almost always yield near-0 advantages and near-noise signals, diverse interventions successfully recycle and repurpose such data: among the proposed strategies, interventions at rollout generation are consistently most effective: nudging the policy to generate ``high-quality'', incorrect solutions introduces rollouts with poor rewards into saturated groups as negative samples, which turns out to improve GRPO by 6.4% to 9.0% across Qwen3-1.7B and 4B. Other interventions such as increasing rollout temperature or adding auxiliary rewards can also restore non-zero advantages, but yield less consistent gains. Further analyses show that effective negative rollouts require informative negative trajectories, that the method remains effective alongside unsaturated data, and that it supports iterative recycling of newly saturated examples. While increasingly stronger LLMs would render more data as saturated, our results demonstrate that don't waste your saturated data: with the right strategies they can be recycled into useful RL training signals in an increasingly data-scarce world.
Outcome-only reinforcement learning gives software engineering (SWE) agents a terminal success signal but little direct guidance about intermediate decisions. We introduce Counterfactual Rollout Replay (CRR), a training-time procedure that uses forkable executable environments to obtain step-level return contrasts. CRR selects a small set of decision points, restores each state, samples an alternative action, and rolls the branch forward under the policy. It retains the realised training trajectory and replaces the advantage at selected steps with the difference between its terminal return and the sampled counterfactual return. The method needs no human process labels or learned process reward model; free refers to those supervision costs, not replay compute. With a 14B policy, CRR improves pass@1 on SWE-bench Verified, SWE-bench Live, and SWE-rebench, and combines with process-reward and trajectory-search methods. On SWE-bench Verified, an equal-wall-clock comparison on the same hardware yields 41.7% versus 36.7% for extended outcome-only GRPO, a 5.0-point gain with fork overhead included. These results apply to environments with affordable, reliable state restoration; stochastic continuations and expensive or imperfect replay remain limitations.
展开完整摘要收起摘要↓
Outcome-only reinforcement learning gives software engineering (SWE) agents a terminal success signal but little direct guidance about intermediate decisions. We introduce Counterfactual Rollout Replay (CRR), a training-time procedure that uses forkable executable environments to obtain step-level return contrasts. CRR selects a small set of decision points, restores each state, samples an alternative action, and rolls the branch forward under the policy. It retains the realised training trajectory and replaces the advantage at selected steps with the difference between its terminal return and the sampled counterfactual return. The method needs no human process labels or learned process reward model; free refers to those supervision costs, not replay compute. With a 14B policy, CRR improves pass@1 on SWE-bench Verified, SWE-bench Live, and SWE-rebench, and combines with process-reward and trajectory-search methods. On SWE-bench Verified, an equal-wall-clock comparison on the same hardware yields 41.7% versus 36.7% for extended outcome-only GRPO, a 5.0-point gain with fork overhead included. These results apply to environments with affordable, reliable state restoration; stochastic continuations and expensive or imperfect replay remain limitations.
作者Jingyi He, Nier Wu, Shuang Liu, Xin Wang, Mengnan Du, Xia Hu
Reward models (RMs) are a key component of large language model post-training, providing reward signals for subsequent reinforcement learning. However, conventional discriminative RMs typically output only scalar scores, making it difficult to identify the response behaviors associated with their scoring decisions. Existing interpretation methods often rely on predefined high-level attributes and require repeated counterfactual interventions for each response pair to validate candidate explanations, lacking a closed-loop mechanism that uses RMs' feedback to train a reusable explainer. To address this, we propose RewardExplainer, a framework that obtains feedback from the target reward model through counterfactual rewriting and uses this feedback to further optimize the explainer. RewardExplainer generates open-ended, atomic, and intervenable natural-language scoring mechanisms, making explanations more concrete, readable, and actionable. It further converts counterfactual feedback into preference supervision, enabling the explainer to more faithfully capture the target RM's scoring preferences and sensitive behaviors than single-pass generation. Extensive experiments across multiple target RMs and explainer backbones show consistent improvements. Beyond interpretation, we use the generated mechanisms to identify potential bias patterns and construct targeted debiasing data for fine-tuning the reward model, improving robustness on reward-hacking benchmarks.
展开完整摘要收起摘要↓
Reward models (RMs) are a key component of large language model post-training, providing reward signals for subsequent reinforcement learning. However, conventional discriminative RMs typically output only scalar scores, making it difficult to identify the response behaviors associated with their scoring decisions. Existing interpretation methods often rely on predefined high-level attributes and require repeated counterfactual interventions for each response pair to validate candidate explanations, lacking a closed-loop mechanism that uses RMs' feedback to train a reusable explainer. To address this, we propose RewardExplainer, a framework that obtains feedback from the target reward model through counterfactual rewriting and uses this feedback to further optimize the explainer. RewardExplainer generates open-ended, atomic, and intervenable natural-language scoring mechanisms, making explanations more concrete, readable, and actionable. It further converts counterfactual feedback into preference supervision, enabling the explainer to more faithfully capture the target RM's scoring preferences and sensitive behaviors than single-pass generation. Extensive experiments across multiple target RMs and explainer backbones show consistent improvements. Beyond interpretation, we use the generated mechanisms to identify potential bias patterns and construct targeted debiasing data for fine-tuning the reward model, improving robustness on reward-hacking benchmarks.
Reinforcement Learning from Human Feedback (RLHF) is a powerful technique for aligning large language models (LLMs) with human preference. However, it often suffers from the reward hacking issue, where policy optimization improves the proxy reward model while actually degrading performance with respect to the true human preference, due to the imperfection of the proxy. To address this, we propose Reward Model Boosting (RMB), a novel approach that enhances the robustness and reliability of the reward signal for RLHF. RMB first trains a set of reward models with a diversity-promoting regularizer. This encourages each model to learn complementary aspects of the reward landscape. Then, RMB learns a lightweight aggregator in the principle of boosting to aggregate the outputs of the diverse reward models into a more accurate and robust reward signal. Our extensive experiments demonstrate that RMB significantly improves reward accuracy on both in-distribution and out-of-distribution datasets, substantially mitigating the reward hacking issue and ultimately improving RLHF performance.
展开完整摘要收起摘要↓
Reinforcement Learning from Human Feedback (RLHF) is a powerful technique for aligning large language models (LLMs) with human preference. However, it often suffers from the reward hacking issue, where policy optimization improves the proxy reward model while actually degrading performance with respect to the true human preference, due to the imperfection of the proxy. To address this, we propose Reward Model Boosting (RMB), a novel approach that enhances the robustness and reliability of the reward signal for RLHF. RMB first trains a set of reward models with a diversity-promoting regularizer. This encourages each model to learn complementary aspects of the reward landscape. Then, RMB learns a lightweight aggregator in the principle of boosting to aggregate the outputs of the diverse reward models into a more accurate and robust reward signal. Our extensive experiments demonstrate that RMB significantly improves reward accuracy on both in-distribution and out-of-distribution datasets, substantially mitigating the reward hacking issue and ultimately improving RLHF performance.
Training efficiency has become the central driver of recent progress in foundation models. To overcome the massive computational and data requirements of large-scale training, researchers increasingly adopt strategies such as selective data sampling, efficient pre-training, and simplified reinforcement learning pipelines. While these strategies drastically reduce overhead, they prompt a critical, yet neglected question: Is efficiency achieved at the expense of model robustness and security? To our knowledge, we present the first systematic cross-domain investigation of the efficiency-vulnerability trade-off. Across vision and language models, we show that efficiency-oriented training increases susceptibility to adversarial and privacy attacks. We characterize this vulnerability by analyzing the models' internal geometry and functional representations, demonstrating that the evaluated efficient variants consistently exhibit sharper loss geometry together with systematic changes in representational structure. We further extend our analysis to "zero RL training", finding that models trained using simplified RL recipes exhibit substantially greater susceptibility to catastrophic forgetting and more pronounced overconfidence than those trained through conventional alignment pipelines. Our findings suggest that training efficiency is rarely a "free lunch"; rather, the mechanisms that minimize computation can inadvertently compromise safety. We conclude by calling for a paradigm shift toward multi-objective training that jointly optimizes for performance, cost, and security.
展开完整摘要收起摘要↓
Training efficiency has become the central driver of recent progress in foundation models. To overcome the massive computational and data requirements of large-scale training, researchers increasingly adopt strategies such as selective data sampling, efficient pre-training, and simplified reinforcement learning pipelines. While these strategies drastically reduce overhead, they prompt a critical, yet neglected question: Is efficiency achieved at the expense of model robustness and security? To our knowledge, we present the first systematic cross-domain investigation of the efficiency-vulnerability trade-off. Across vision and language models, we show that efficiency-oriented training increases susceptibility to adversarial and privacy attacks. We characterize this vulnerability by analyzing the models' internal geometry and functional representations, demonstrating that the evaluated efficient variants consistently exhibit sharper loss geometry together with systematic changes in representational structure. We further extend our analysis to "zero RL training", finding that models trained using simplified RL recipes exhibit substantially greater susceptibility to catastrophic forgetting and more pronounced overconfidence than those trained through conventional alignment pipelines. Our findings suggest that training efficiency is rarely a "free lunch"; rather, the mechanisms that minimize computation can inadvertently compromise safety. We conclude by calling for a paradigm shift toward multi-objective training that jointly optimizes for performance, cost, and security.
作者Zhenyu Lei, Mingjia Shi, Xingbo Fu, Haoyu He, Qi R. Wang, Jundong Li
Existing summarizers for memory systems are typically optimized for human-facing criteria such as faithfulness, which misaligns with their true objective: preserving the evidence needed to support future queries. We show that conditioning summarization on query-answer pairs substantially improves answer quality, and that this utility-aware behavior is transferable across queries. Motivated by these findings, we propose MemSuit, a self-distillation framework in which a teacher summarizer, conditioned on observed query-answer pairs, produces utility-aware memory entries that a student learns to reproduce from the raw conversation alone. To prevent collateral erasure where conditioning on a single query-answer pair discards evidence relevant to other plausible queries, the teacher decomposes each block into multiple self-contained entries that preserve distinct query-relevant facets as independently retrievable units. To align the retriever with the compact, fact-dense style of teacher entries, we further fine-tune the embedding model with a contrastive objective supervised by teacher entries. Across a diverse suite of conversational query types, MemSuit consistently outperforms state-of-the-art baselines, confirming the value of grounding memory in downstream utility.
展开完整摘要收起摘要↓
Existing summarizers for memory systems are typically optimized for human-facing criteria such as faithfulness, which misaligns with their true objective: preserving the evidence needed to support future queries. We show that conditioning summarization on query-answer pairs substantially improves answer quality, and that this utility-aware behavior is transferable across queries. Motivated by these findings, we propose MemSuit, a self-distillation framework in which a teacher summarizer, conditioned on observed query-answer pairs, produces utility-aware memory entries that a student learns to reproduce from the raw conversation alone. To prevent collateral erasure where conditioning on a single query-answer pair discards evidence relevant to other plausible queries, the teacher decomposes each block into multiple self-contained entries that preserve distinct query-relevant facets as independently retrievable units. To align the retriever with the compact, fact-dense style of teacher entries, we further fine-tune the embedding model with a contrastive objective supervised by teacher entries. Across a diverse suite of conversational query types, MemSuit consistently outperforms state-of-the-art baselines, confirming the value of grounding memory in downstream utility.
Residual policy adaptation provides a lightweight way to modify a strong reference policy, but a shared scale and a fixed subgroup partition can hide heterogeneous degradation and become fragile when the deployment mixture of information states changes. We introduce RSD-Poker, a structure-adaptive and shift-robust certification framework that freezes a bank of residual families and scales, learns a policy-visible partition on an independent structure split, and freezes that partition before calibration labels are joined. Each candidate-group pair receives a weighted simultaneous upper certificate for anchor-relative risk and a lower certificate for weak-response utility. A robust group-to-candidate map is then selected over a predeclared uncertainty set of deployment group proportions. Under independent calibration units drawn from each frozen group's law, a candidate bank and partition fixed before calibration, and invariant within-group conditionals, the selected map satisfies its declared mixture-robust risk budget and utility certificate with probability at least $1-ζ_{risk}-ζ_{util}$. The information contract supports both a teacher-backed transform and a teacher-free observation-only student. The retained deterministic 24-state audit remains an exact replay diagnostic: empirical-zero selects $α=0.08$, raising the weak-response proxy from 4.2082 to 4.2889 with $0/12$ held-out threshold crossings. On stratified held-out states, the learned-partition dual selector raises weak utility from 4.4074 under global dual certification to 4.4936 and lowers held-out violation from 0.0215 to 0.0078; its mixture-robust variant reaches violation 0.0059. Across five observation-only checkpoints, risk-calibrated residuals attain weak utility $4.3659\pm0.0177$ and violation rate $0.0178\pm0.0057$.
展开完整摘要收起摘要↓
Residual policy adaptation provides a lightweight way to modify a strong reference policy, but a shared scale and a fixed subgroup partition can hide heterogeneous degradation and become fragile when the deployment mixture of information states changes. We introduce RSD-Poker, a structure-adaptive and shift-robust certification framework that freezes a bank of residual families and scales, learns a policy-visible partition on an independent structure split, and freezes that partition before calibration labels are joined. Each candidate-group pair receives a weighted simultaneous upper certificate for anchor-relative risk and a lower certificate for weak-response utility. A robust group-to-candidate map is then selected over a predeclared uncertainty set of deployment group proportions. Under independent calibration units drawn from each frozen group's law, a candidate bank and partition fixed before calibration, and invariant within-group conditionals, the selected map satisfies its declared mixture-robust risk budget and utility certificate with probability at least $1-ζ_{risk}-ζ_{util}$. The information contract supports both a teacher-backed transform and a teacher-free observation-only student. The retained deterministic 24-state audit remains an exact replay diagnostic: empirical-zero selects $α=0.08$, raising the weak-response proxy from 4.2082 to 4.2889 with $0/12$ held-out threshold crossings. On stratified held-out states, the learned-partition dual selector raises weak utility from 4.4074 under global dual certification to 4.4936 and lowers held-out violation from 0.0215 to 0.0078; its mixture-robust variant reaches violation 0.0059. Across five observation-only checkpoints, risk-calibrated residuals attain weak utility $4.3659\pm0.0177$ and violation rate $0.0178\pm0.0057$.
Diffusion models are now widely used in Bayesian inverse problems in imaging as priors, where latent diffusion models are often used for larger scale problems to keep the computational complexity and model-size manageable. Unfortunately, the auto-encoder based compression results in loss of spatial detail. In addition, the optimization is converted to a non-linear problem. In this paper, we introduce a posterior sampling algorithm customized for the pyramidal/cascaded architecture, which relies on a coarse to fine hierarchical strategy to generate images in the pixel domain. We present CLIMB-Flow which alternates between three steps: an end-point estimation from the current coarse and noisy image, data-consistent update of the clean image, and re-noising it back to the level the network expects. Together these steps sample the posterior at that scale using an approximate Gibbs sampling from two conditional distributions. Experiments on ImageNet, CelebA, AFHQ and fastMRI span inpainting, deblurring, super-resolution and accelerated MRI, with PSNR gains of 1.37-7.66 dB over the strongest competing method on CelebA and pixel-domain reconstruction up to 512x512.
展开完整摘要收起摘要↓
Diffusion models are now widely used in Bayesian inverse problems in imaging as priors, where latent diffusion models are often used for larger scale problems to keep the computational complexity and model-size manageable. Unfortunately, the auto-encoder based compression results in loss of spatial detail. In addition, the optimization is converted to a non-linear problem. In this paper, we introduce a posterior sampling algorithm customized for the pyramidal/cascaded architecture, which relies on a coarse to fine hierarchical strategy to generate images in the pixel domain. We present CLIMB-Flow which alternates between three steps: an end-point estimation from the current coarse and noisy image, data-consistent update of the clean image, and re-noising it back to the level the network expects. Together these steps sample the posterior at that scale using an approximate Gibbs sampling from two conditional distributions. Experiments on ImageNet, CelebA, AFHQ and fastMRI span inpainting, deblurring, super-resolution and accelerated MRI, with PSNR gains of 1.37-7.66 dB over the strongest competing method on CelebA and pixel-domain reconstruction up to 512x512.
Jointly generating future video and actions has become a standard recipe for world-action models, and the strongest systems denoise the two streams on separate schedules: actions are decoded in few steps so control stays fast, while the video stream runs longer to keep the predicted future sharp. The design is deliberate, but it leaves the two streams on different clocks, and an action can become executable while the future that should justify it is still largely unresolved. We formalize this as a two-clock view of asynchronous inference and introduce the commitment-evidence gap, a quantity read directly from a model's own sampling schedule rather than measured by search. The gap is predictive: as it widens, candidate utility becomes harder to identify and extra candidate sampling buys less, while advancing the world stream buys more, and the two cross. Spending more world computation is therefore not simply better. The useful interval is closed at both ends, and both ends can be read off the schedule before any rollout. ReSync places the computation inside it: hold the action state, advance only the world within the supported window, then resume native denoising. No parameters change and no candidates are compared. On a frozen paired RoboCasa panel this improves success by 4.48 points, while an equal-compute control that waits without advancing the world does not move, and the same rule transfers to a second benchmark and a second backbone without retuning.
展开完整摘要收起摘要↓
Jointly generating future video and actions has become a standard recipe for world-action models, and the strongest systems denoise the two streams on separate schedules: actions are decoded in few steps so control stays fast, while the video stream runs longer to keep the predicted future sharp. The design is deliberate, but it leaves the two streams on different clocks, and an action can become executable while the future that should justify it is still largely unresolved. We formalize this as a two-clock view of asynchronous inference and introduce the commitment-evidence gap, a quantity read directly from a model's own sampling schedule rather than measured by search. The gap is predictive: as it widens, candidate utility becomes harder to identify and extra candidate sampling buys less, while advancing the world stream buys more, and the two cross. Spending more world computation is therefore not simply better. The useful interval is closed at both ends, and both ends can be read off the schedule before any rollout. ReSync places the computation inside it: hold the action state, advance only the world within the supported window, then resume native denoising. No parameters change and no candidates are compared. On a frozen paired RoboCasa panel this improves success by 4.48 points, while an equal-compute control that waits without advancing the world does not move, and the same rule transfers to a second benchmark and a second backbone without retuning.
作者Semin Joung, Abhilasha Dave, Luca Scomparin, Filipp Khabanov, Zheng Yan, Benedikt Geiger, George McKee, Ryan N. Coffee, David R. Smith
High-bandwidth plasma diagnostics increasingly provide inputs to machine learning and signal-processing algorithms intended for real-time tokamak control, but the complete path from diagnostic sampling to control-system handoff is difficult to commission because of their sampling rates. We develop a packet-level digital hardware twin for the megahertz diagnostic edge-AI architecture. The simulator represents a 64-channel, 1 MHz beam emission spectroscopy (BES) diagnostic embedded in a 96-channel dual-carrier acquisition system, two 48-channel streams with SPAD0 sample counters, 10 GbE Hardware UDP transport, packet loss and network jitter, FPGA parsing and dual-carrier alignment, causal preprocessing and edge inference, a compact Ethernet result packet, receiver-side shared state, and a 1 kHz PCS-like control cycle. Binary UDP payloads and PCAP files are generated rather than emulating transport only at the array level. With a baseline of 20 samples per packet, each carrier generates 50,000 packets s$^{-1}$ and 100 MB s$^{-1}$ of user payload. A 110 ms reference run produces 11,000 HUDP packets; an intentionally dropped 20-sample packet is detected by the sample-counter continuity logic and invalidates the two overlapping 128-sample inference windows without silent interpolation. For valid windows, the configured engineering latency model gives a median last-input-to-shared-memory latency of 91.6 us and a 99th percentile of 108.1 us. A separate operating-system loopback test sends binary FPGA-result datagrams through a UDP receiver into POSIX shared memory and preserves packet sequence and CRC for 20/20 packets. Interactive GUI interfaces expose timing, packetization, network faults, inference thresholds, and control-state inspection. The framework provides a reproducible environment for testing diagnostic-to-accelerator interfaces and fail-safe behavior for deployment on fusion devices.
展开完整摘要收起摘要↓
High-bandwidth plasma diagnostics increasingly provide inputs to machine learning and signal-processing algorithms intended for real-time tokamak control, but the complete path from diagnostic sampling to control-system handoff is difficult to commission because of their sampling rates. We develop a packet-level digital hardware twin for the megahertz diagnostic edge-AI architecture. The simulator represents a 64-channel, 1 MHz beam emission spectroscopy (BES) diagnostic embedded in a 96-channel dual-carrier acquisition system, two 48-channel streams with SPAD0 sample counters, 10 GbE Hardware UDP transport, packet loss and network jitter, FPGA parsing and dual-carrier alignment, causal preprocessing and edge inference, a compact Ethernet result packet, receiver-side shared state, and a 1 kHz PCS-like control cycle. Binary UDP payloads and PCAP files are generated rather than emulating transport only at the array level. With a baseline of 20 samples per packet, each carrier generates 50,000 packets s$^{-1}$ and 100 MB s$^{-1}$ of user payload. A 110 ms reference run produces 11,000 HUDP packets; an intentionally dropped 20-sample packet is detected by the sample-counter continuity logic and invalidates the two overlapping 128-sample inference windows without silent interpolation. For valid windows, the configured engineering latency model gives a median last-input-to-shared-memory latency of 91.6 us and a 99th percentile of 108.1 us. A separate operating-system loopback test sends binary FPGA-result datagrams through a UDP receiver into POSIX shared memory and preserves packet sequence and CRC for 20/20 packets. Interactive GUI interfaces expose timing, packetization, network faults, inference thresholds, and control-state inspection. The framework provides a reproducible environment for testing diagnostic-to-accelerator interfaces and fail-safe behavior for deployment on fusion devices.
Masked diffusion language models (MDMs) admit flexible generation orders, making the unmasking strategy an inference decision. Existing methods vary in how they prioritize positions, control parallelism, restrict selection regions, revise predictions, or plan future denoising, yet it remains unclear when these choices should change during generation. We study this question through strategy reversals, where an alternative action becomes preferable to a fixed choice. We organize MDM inference into five axes--score, cardinality, region, commitment, and planning--and define adaptation opportunity as the one-step utility advantage of the best candidate action over a validation-selected fixed action. This view shows that adaptation value depends on both the frequency and magnitude of such reversals. Across three MDMs and ten tasks, adaptation opportunities are highly heterogeneous, with some regimes exhibiting concentrated and predictable one-step gains. This motivates selective adaptation: lightweight detectors calibrated on validation prompts identify high-opportunity states, capturing, for example, 56.9 percent of the candidate-set oracle opportunity by adapting only the top 10 percent of states on LLaDA-8B constrained JSON filling. Our transition-level results suggest that state adaptation is most useful when applied selectively rather than uniformly.
展开完整摘要收起摘要↓
Masked diffusion language models (MDMs) admit flexible generation orders, making the unmasking strategy an inference decision. Existing methods vary in how they prioritize positions, control parallelism, restrict selection regions, revise predictions, or plan future denoising, yet it remains unclear when these choices should change during generation. We study this question through strategy reversals, where an alternative action becomes preferable to a fixed choice. We organize MDM inference into five axes--score, cardinality, region, commitment, and planning--and define adaptation opportunity as the one-step utility advantage of the best candidate action over a validation-selected fixed action. This view shows that adaptation value depends on both the frequency and magnitude of such reversals. Across three MDMs and ten tasks, adaptation opportunities are highly heterogeneous, with some regimes exhibiting concentrated and predictable one-step gains. This motivates selective adaptation: lightweight detectors calibrated on validation prompts identify high-opportunity states, capturing, for example, 56.9 percent of the candidate-set oracle opportunity by adapting only the top 10 percent of states on LLaDA-8B constrained JSON filling. Our transition-level results suggest that state adaptation is most useful when applied selectively rather than uniformly.
作者Boyuan Zhang, Yingjun Du, Xiantong Zhen, Ling Shao
Joint-embedding world models enable visual planning by learning action-conditioned dynamics in latent space. Yet they are commonly trained for one-step prediction on encoded states, while planning recursively applies the learned transition to its own predictions. One-step accuracy therefore does not capture how prediction errors propagate under recursive rollout. We decompose multi-step rollout error into the errors introduced at individual steps and their propagation through subsequent transitions. We show that state-affine dynamics are precisely the differentiable transitions with state-independent Jacobians, eliminating the nonlinear propagation residual and making the error propagation operators depend only on the action sequence. Guided by this result, we introduce SALT (State-Affine Latent Transition), an action-conditioned state-affine dynamics model in which the action modulates both the state transformation and the additive update. We train SALT through recursive multi-step rollout supervision, feeding each predicted latent state back into the transition so that training matches how the model is used during planning. Across four visual planning environments, SALT exhibits $1.48$--$2.19\times$ higher one-step prediction error than the matched LeWM baseline, yet improves closed-loop success in every environment by $10.0$ percentage points on average. On OGBench-Cube, the fraction of episodes that fail with a sharp rise in model-predicted cost after execution decreases from $23.3%$ to $2.0%$.
展开完整摘要收起摘要↓
Joint-embedding world models enable visual planning by learning action-conditioned dynamics in latent space. Yet they are commonly trained for one-step prediction on encoded states, while planning recursively applies the learned transition to its own predictions. One-step accuracy therefore does not capture how prediction errors propagate under recursive rollout. We decompose multi-step rollout error into the errors introduced at individual steps and their propagation through subsequent transitions. We show that state-affine dynamics are precisely the differentiable transitions with state-independent Jacobians, eliminating the nonlinear propagation residual and making the error propagation operators depend only on the action sequence. Guided by this result, we introduce SALT (State-Affine Latent Transition), an action-conditioned state-affine dynamics model in which the action modulates both the state transformation and the additive update. We train SALT through recursive multi-step rollout supervision, feeding each predicted latent state back into the transition so that training matches how the model is used during planning. Across four visual planning environments, SALT exhibits $1.48$--$2.19\times$ higher one-step prediction error than the matched LeWM baseline, yet improves closed-loop success in every environment by $10.0$ percentage points on average. On OGBench-Cube, the fraction of episodes that fail with a sharp rise in model-predicted cost after execution decreases from $23.3%$ to $2.0%$.
作者Zhihao Chen, Yiyuan Ge, Ziyang Wang, Pu Cao, Lu Yang
Recent large-scale Vision-and-Language Navigation (VLN) models deliver strong accuracy but remain costly to deploy due to their large parameter counts and computational requirements. We tackle efficient VLN in two steps. First, we build a high-performing teacher that makes navigation evidence selection explicit and compressible. The teacher introduces a small set of learnable query slots to extract global and local action-sufficient navigable evidence from panoramic observations via a Navigable Query Generator, then progressively grounds these evidence tokens to the instruction with an Instruction-Query Aligner for policy prediction. Second, using this explicit query bottleneck as a distillation interface, we train a compact student by transferring both where to attend and what to do. We distill the teacher's global and local navigable queries with a navigation-aware token-adaptive objective, then further match action distributions during fine-tuning. Experiments on standard VLN benchmarks demonstrate that our student nearly matches the teacher's navigation performance while reducing the number of parameters by 93.65% compared to the teacher.
展开完整摘要收起摘要↓
Recent large-scale Vision-and-Language Navigation (VLN) models deliver strong accuracy but remain costly to deploy due to their large parameter counts and computational requirements. We tackle efficient VLN in two steps. First, we build a high-performing teacher that makes navigation evidence selection explicit and compressible. The teacher introduces a small set of learnable query slots to extract global and local action-sufficient navigable evidence from panoramic observations via a Navigable Query Generator, then progressively grounds these evidence tokens to the instruction with an Instruction-Query Aligner for policy prediction. Second, using this explicit query bottleneck as a distillation interface, we train a compact student by transferring both where to attend and what to do. We distill the teacher's global and local navigable queries with a navigation-aware token-adaptive objective, then further match action distributions during fine-tuning. Experiments on standard VLN benchmarks demonstrate that our student nearly matches the teacher's navigation performance while reducing the number of parameters by 93.65% compared to the teacher.
On-policy distillation (OPD) trains a student to match a teacher's next-token distributions on student-generated trajectories. However, privileged information supplied to the teacher for OPD training may contain sensitive data. Whether the student leaks private information about the records supplied to the teacher during distillation remains poorly understood. To the best of our knowledge, we present the first systematic study of membership inference in this setting. We find that fresh student trajectories expose sparse membership signals that fixed reference-answer losses often miss. These signals are mixed with probability changes caused by training on other records. We introduce Leaky, which samples fresh trajectories from the target model and compares its token log-probabilities with the maximum across matched reference models trained without the candidate records. It applies Leaky ReLU to the resulting gaps, preserving positive gaps and downweighting negative gaps as an approximate correction for incidental positive gaps in non-members. Across fifteen targets spanning mathematics, medical question answering, and code generation, Leaky outperforms all evaluated baselines and achieves mean AUROC 0.875, compared with 0.614 for the strongest baseline on each target in the main evaluation. On the same sampled trajectories, the strongest baseline achieves mean AUROC 0.826. These results show that students trained through OPD can expose the membership of records used for teacher supervision, even when fixed reference-answer losses provide little evidence of membership.
展开完整摘要收起摘要↓
On-policy distillation (OPD) trains a student to match a teacher's next-token distributions on student-generated trajectories. However, privileged information supplied to the teacher for OPD training may contain sensitive data. Whether the student leaks private information about the records supplied to the teacher during distillation remains poorly understood. To the best of our knowledge, we present the first systematic study of membership inference in this setting. We find that fresh student trajectories expose sparse membership signals that fixed reference-answer losses often miss. These signals are mixed with probability changes caused by training on other records. We introduce Leaky, which samples fresh trajectories from the target model and compares its token log-probabilities with the maximum across matched reference models trained without the candidate records. It applies Leaky ReLU to the resulting gaps, preserving positive gaps and downweighting negative gaps as an approximate correction for incidental positive gaps in non-members. Across fifteen targets spanning mathematics, medical question answering, and code generation, Leaky outperforms all evaluated baselines and achieves mean AUROC 0.875, compared with 0.614 for the strongest baseline on each target in the main evaluation. On the same sampled trajectories, the strongest baseline achieves mean AUROC 0.826. These results show that students trained through OPD can expose the membership of records used for teacher supervision, even when fixed reference-answer losses provide little evidence of membership.
作者Safaeid Hossain Arib, Rabeya Akter, Ismam Nur Swapnil, Md. Faiyaz Abdullah Sayeedi, Tasnim Mohiuddin, Md Mofijul Islam
On-policy self-distillation trains reasoning models on their own trajectories using dense token distribution guidance from a privileged teacher with access to a verified solution. This supervision transfers what the teacher predicts without directly transferring where it attends within the preceding context. We introduce On-Policy Attention Self-Distillation (OPASD), which complements token-level supervision with solution-conditioned attention distillation. Because the privileged teacher can attend to verified solution tokens unavailable to the student, OPASD projects teacher attention onto student-visible positions and renormalizes the resulting distribution before alignment. Across three model sizes and four competition-level mathematics benchmarks, OPASD consistently outperforms token-only OPSD, improving average accuracy by 4.98 to 8.40 percentage points. OPASD also avoids the response-length inflation and performance degradation observed with token-only distillation, reducing generated rollout tokens by 73.9% and estimated model compute by 72.6% while training 1.53x faster. These results show that solution-conditioned attention provides a complementary supervision signal that makes on-policy self-distillation more accurate, stable, and compute-efficient.
展开完整摘要收起摘要↓
On-policy self-distillation trains reasoning models on their own trajectories using dense token distribution guidance from a privileged teacher with access to a verified solution. This supervision transfers what the teacher predicts without directly transferring where it attends within the preceding context. We introduce On-Policy Attention Self-Distillation (OPASD), which complements token-level supervision with solution-conditioned attention distillation. Because the privileged teacher can attend to verified solution tokens unavailable to the student, OPASD projects teacher attention onto student-visible positions and renormalizes the resulting distribution before alignment. Across three model sizes and four competition-level mathematics benchmarks, OPASD consistently outperforms token-only OPSD, improving average accuracy by 4.98 to 8.40 percentage points. OPASD also avoids the response-length inflation and performance degradation observed with token-only distillation, reducing generated rollout tokens by 73.9% and estimated model compute by 72.6% while training 1.53x faster. These results show that solution-conditioned attention provides a complementary supervision signal that makes on-policy self-distillation more accurate, stable, and compute-efficient.
Reinforcement learning with verifiable rewards (RLVR) often relies on sparse outcome rewards, providing coarse supervision for long-horizon agents. On-policy self-distillation (OPSD) complements this signal with dense privileged feedback. However, we identify Decision--Timestamp Mismatch: privileged guidance may be misaligned with the student's functional decision because the corresponding decision can occur at a different timestep, while the student's decision itself may span multiple timesteps rather than being tied to a single timestamp. Thus, timestamp-local supervision can misalign both the context and the temporal scope of credit. To address this mismatch, we introduce \textsc{AlignOPSD}, following the principle of aligning supervision before assigning credit. Decision-Aligned Supervision Rectification re-scores the same student-sampled response in functionally matched contexts across sibling rollouts to calibrate local teacher evidence. Semi-Markov Hierarchical Credit Assignment then derives variable-duration decision spans from correspondence changes and uses rectified evidence to allocate outcome-grounded credit across spans and their constituent turns. We evaluate \textsc{AlignOPSD} with Qwen2.5-3B and Qwen2.5-7B on ALFWorld, WebShop, and Search-QA against representative baselines. \textsc{AlignOPSD} outperforms both GRPO and StepOPSD across all eight backbone--aggregate-metric comparisons, improving on GRPO by 5.5--8.7 % and ranking first in six. Additional analyzes examine the two alignment stages and hyperparameter sensitivity between tasks. Our code is avaliable at https://github.com/mingju-c/Align-OPSD
展开完整摘要收起摘要↓
Reinforcement learning with verifiable rewards (RLVR) often relies on sparse outcome rewards, providing coarse supervision for long-horizon agents. On-policy self-distillation (OPSD) complements this signal with dense privileged feedback. However, we identify Decision--Timestamp Mismatch: privileged guidance may be misaligned with the student's functional decision because the corresponding decision can occur at a different timestep, while the student's decision itself may span multiple timesteps rather than being tied to a single timestamp. Thus, timestamp-local supervision can misalign both the context and the temporal scope of credit. To address this mismatch, we introduce \textsc{AlignOPSD}, following the principle of aligning supervision before assigning credit. Decision-Aligned Supervision Rectification re-scores the same student-sampled response in functionally matched contexts across sibling rollouts to calibrate local teacher evidence. Semi-Markov Hierarchical Credit Assignment then derives variable-duration decision spans from correspondence changes and uses rectified evidence to allocate outcome-grounded credit across spans and their constituent turns. We evaluate \textsc{AlignOPSD} with Qwen2.5-3B and Qwen2.5-7B on ALFWorld, WebShop, and Search-QA against representative baselines. \textsc{AlignOPSD} outperforms both GRPO and StepOPSD across all eight backbone--aggregate-metric comparisons, improving on GRPO by 5.5--8.7 % and ranking first in six. Additional analyzes examine the two alignment stages and hyperparameter sensitivity between tasks. Our code is avaliable at https://github.com/mingju-c/Align-OPSD