每天知道
你的领域
出了什么

Large language models (LLMs) are increasingly used as agents in simulations of social systems, yet it remains unclear when their behavior can be interpreted as a faithful proxy for human decision-making. Here we test LLM agents against a direct empirical benchmark: a large-scale networked Prisoner's Dilemma experiment with human participants. Using the same interaction protocol, payoff structure, and network topologies, we compare nine open-weight LLMs with the human data. The selected model reproduces several macro-level features of cooperation dynamics, including the early decline and later stabilization of cooperation. This aggregate agreement, however, does not extend uniformly to finer levels of behavior. LLM populations underestimate individual-level heterogeneity and generate conditional cooperation patterns that differ from those observed in humans. Adding a fraction of random agents improves some aspects of micro-level agreement, but does not remove the mismatch in decision rules. These findings reveal a macro--micro dissociation in LLM-based social agents: collective outcomes can appear human-like even when the underlying behavioral distributions and mechanisms are not. They suggest that validating LLM agents as human surrogates requires comparisons across aggregate dynamics, individual heterogeneity, and context-dependent decision rules, rather than outcome-level agreement alone.

Tool-augmented language-model agents are often evaluated by whether they select the correct tool, produce valid API arguments, and complete the requested task. However, an agent may choose the right tool and still act on the wrong external entity. For example, a request to "email Alex about the launch" may lead the agent to contact the wrong Alex, attach the wrong launch document, reply in the wrong thread, or update the wrong customer account. We call these errors entity binding failures. This paper studies entity binding failures as a distinct reliability and safety problem in tool-augmented agents. We formalize the separation between tool correctness and entity correctness, introduce a taxonomy of wrong-entity failures in enterprise workflows, and evaluate entity-aware execution mechanisms including entity-resolution preconditions, confidence-gated binding, clarification under ambiguity, and provenance tracking. In a controlled diagnostic evaluation across 60 tasks, five model backends, and six tool-use methods, all methods achieved 0.0 percent wrong-tool error, yet action-oriented baselines still produced wrong-entity actions in 24.0-26.0 percent of runs. Entity-aware methods eliminated wrong-entity actions and risk-weighted wrong-entity exposure in this setting, but reduced direct task completion by deferring under ambiguity. These findings show that safe tool use requires not only selecting the correct tool, but also reliably binding natural-language references to the correct real-world entity before action.

Large language models (LLMs) are increasingly used in open-ended multi-agent settings, but the long-run dynamics of model--model interaction remain poorly understood. We study whether open-ended LLM discussions exhibit attractor-like behavior, i.e. topic-independent stable sets of behaviors which conversations settle into. Across 7 LLMs and 20 controversial topics, we compare self-play and mixed-play dyadic debates, tracking trajectories in representation space, discourse traits, and stances. We find self-play trajectories to be model-specific attractors that draw their conversation partners asymmetrically in mixed-play debates, influencing the other models' stylistic choices and behavior. For example, Claude Haiku is a strong attractor of other models in latent space, corresponding to other models taking on its traits like metacommentary, and models like GPT-4.1 nano are especially malleable. Our results suggest that open-ended LLM interactions are partially predictable from model-specific attractors, but shaped by structured and asymmetric partner influence. Overall, our analysis sheds some light on the complex behavior of open-ended multi-agent interaction, which we hope is helpful in designing, predicting, and monitoring autonomous agentic systems in the real world.

Coding agents are rapidly becoming a major application of agentic LLMs, but serving them efficiently remains challenging. Progress on this challenge requires understanding real workload patterns, yet the data needed for such analysis is largely absent. Existing public traces and benchmarks do not capture real, day-to-day coding-agent usage across multiple agents and model families for serving-system analysis. To help fill this gap, we collect and release a trace of roughly 4,300 coding-agent sessions, containing about 350,000 LLM steps and 430,000 tool calls from our own day-to-day use of Claude Code and Codex. Our analysis shows that coding-agent workloads feature long autonomous loops, long contexts with short outputs, diverse and heavily-tailed tool calls, and high but imperfect prefix cache hit rates. These findings point to concrete opportunities for optimizing serving, including lower-overhead tool calling, append-length-aware prefill, semantic-aware tool-latency prediction, and improved KV-cache management around human-paced gaps. We release the dataset, trace collection pipeline, and analysis code at https://github.com/uw-syfi/TraceLab.git the project website is https://tracelab.cs.washington.edu.

A rapidly growing class of LLM agents is multi-party: the agent acts for a principal (who briefs it, sends follow-ups, and receives results) while also conversing in a separate channel with a counterparty whose interests may diverge (negotiating with a vendor, screening inbound requests, or mediating between employees). Here "help whoever you are talking to" is the wrong objective. The agent must stay loyal to the principal it represents without over-refusing the principal's own cooperative asks. We study this multi-party loyalty problem and contribute a measurement instrument, two mechanisms, and a structural lesson. PrincipalBench is a 75-item multi-turn benchmark with leak probes, dual judges, and an integrity-audit gate. Across 13 frontier subjects it exposes a sharp split (<=20% vs. 53.6-75.3% harm) invisible to single-turn safety evaluations: a selective cluster that declines adversarial probes while still following the principal's legitimate requests, and an over-refusing cluster that refuses broadly. (M1) A prompt-time loyalty scaffold (a fixed system prompt of seven prioritized rules, open-coded from 50+ failure trajectories) holds Claude-Sonnet to 19.4% harm and all nine selective subjects to <=20%. (M2) A per-token-KL distillation recipe transfers a prompted Qwen3-32B teacher into 8B Qwen3 and Llama-3.1 students, the strongest open-weight recipe we measure. (Lesson) Both mechanisms only move along a common leak/over-refusal trade-off rather than crossing it: improving one axis costs the other, and the jointly favorable outcome stays out of reach.

We discover a behavioral invariant in LLM agents under persistent memory poisoning: in architectures where routing information is retrieved through observable memory-tool invocations, successful attacks require calling memory_recall_fact before email_send_email, a transition that non-exfiltrating sessions rarely exhibit. Under the evaluated architecture, this invariant follows from the attack's information-retrieval dependency rather than being merely an empirical correlation, and suppressing it breaks the attack. A simple rule exploiting this invariant alone achieves AUC = 0.9563. A Random Forest classifier over 19 trajectory features refines it to AUC = 0.9904 (BCa 95% CI [0.987, 0.993], N=10,000 resamples), demonstrating that the attack imprints on multiple independent behavioral channels. The signature is overdetermined: removing all recall-related features (half the feature set) leaves AUC unchanged at 0.990, confirming that memory poisoning induces a distributed trajectory signature rather than a single observable anomaly. Cross-model hold-out on 9 models (7B-120B parameters) confirms AUC = 1.000 on 6/9 hold-out splits, with all three exceptions mechanistically explained. The invariant generalizes to frontier models (GPT-4.1, GPT-4o) without retraining. A strictly prefix-only variant achieves AUC = 0.934, suggesting that real-time blocking is feasible with moderate degradation. The boundary is forensically useful: prompt-injection attacks that bypass memory produce a distinct trajectory (score = 0.541), enabling incident responders to distinguish memory-channel attacks from prompt-injection attacks using tool-call logs alone.

The development of on-device large language models (LLMs) is driven by the need for privacy and fast response times. Energy-intensive data transfer on mobile devices makes Processing-in-Memory (PIM) an effective solution. Due to stringent DRAM cost constraints, limited physical footprint on circuit boards, and the interaction between applications and LLMs, it is imperative for the CPU and PIM to operate concurrently within a shared memory space. However, challenges such as bank conflicts and bus congestion can arise, potentially diminishing the performance and energy benefits of PIM. To address this challenge, we introduce COSM, a cooperative scheduling framework designed to facilitate the concurrent operation of PIM and CPU tasks on mobile platforms. Our key innovations include: 1) a low-interference PIM control interface that generates the maximum number of PIM commands without disrupting CPU memory accesses; 2) an idleness-aware scheduling method that integrates PIM commands into available idle time windows within the CPU's access sequence. COSM not only hides PIM execution latency from the CPU, but also overlaps PIM execution with data transfer. Experiments on concurrent execution of LLMs and mobile workloads, including mobile applications and compute-intensive kernels, demonstrate that COSM improves PIM throughput by up to 2.8x compared to the baseline scheduling method with less than 2.0% CPU performance loss.

DiT video generation is latency-intensive due to iterative full-frame denoising, while prior cloud-edge methods largely rely on static inter-step decoupling and cannot leverage inter-frame similarity or adapt to system dynamics. We propose EcoVideo, an entropy-orchestrated framework for dynamic inter-frame decoupling: early-stage self-attention entropy provides a training-free estimate of frame-wise information density for frame selection; a cloud large model denoises sparse high-entropy keyframes; and an edge lightweight model reconstructs the remaining frames via motion-aware interpolation with refinement for temporal stability. EcoVideo further adapts the keyframe budget and edge refinement depth to real-time bandwidth and compute availability, optimizing end-to-end latency under constraints. Experiments on representative DiT video generators show improved quality--efficiency trade-offs and up to 2.9x end-to-end speedup in low-bandwidth, compute-limited edge settings. Code is available at https://github.com/IF-LAB-PKU/EcoVideo.

As LLM inference becomes a major cloud workload, its growing energy footprint makes cluster-wide energy optimization increasingly important. Serverless LLM serving helps platforms absorb traffic volatility by elastically sharing GPU resources across models, but this sharing also makes energy optimization difficult. Multiple co-resident models run under one device-wide operating point, while their resource demands and latency slack change across execution phases and load conditions. As a result, minimizing energy requires coordinated scheduling across request placement, runtime resource adaptation, and workload consolidation. We present Festina, a profiling-guided, power-aware control plane to minimize cluster-wide energy for serverless LLM serving. Unlike common global-local schedulers that focus on throughput or tail latency, Festina makes energy-first decisions by jointly coordinating request placement, SM partitioning, and GPU operating points under TTFT/TBT SLOs. In our system, a lightweight global scheduler performs fast, SLO-safe, energy-aware placement using constant-time lookups from offline profiles and GPU state summaries. On each GPU, a phase-aware local scheduler continuously adapts task batching and compute resources to minimize power consumption. Festina further performs energy-aware workload consolidation to reduce GPUs' static power consumption via SLO-aware migration. Comparison with four SOTA LLM serving systems and one DVFS-augmented system demonstrates that Festina reduces energy consumption by up to 56% while maintaining parity in SLO attainment (within a 2% margin)

Dynamic sparse attention (DSA) accelerates long-context LLM decoding by attending to only the top-K KV blocks relevant to each query, but it introduces a serialized selection-to-attention dependency that emerges as a new latency bottleneck. We present PRR, a speculate-reuse-repair runtime that exploits temporal locality in DSA selections to predict likely blocks, speculate the attention over them while selection is in flight, and incrementally repair missed blocks once the true selected set is known. PRR uses a lightweight EMA-based predictor, a profiling-guided speculation budget that keeps speculative work off the critical path, and a FlashAttention-based repair kernel that folds missed blocks into the partial attention state using online-softmax statistics. Across long-context benchmarks and representative DSA methods, PRR reduces per-token decoding latency by up to 40% while preserving downstream task accuracy. Github: https://github.com/Tianyu9748/Incremental_FlashAttention

Autonomous scientific discovery systems increasingly use large language models (LLMs) to propose new hypotheses, but many such systems condition primarily on experimental memory: archives of high-scoring candidates or heuristic summaries of recent trials. We argue that discovery agents should instead maintain explicit, uncertainty-aware beliefs about hypothesis quality. We introduce BayesEvolve, a belief-guided discovery framework that converts experimental evidence into a predictive belief state and uses this belief to guide future experimentation. As a controlled testbed for belief-guided discovery, we evaluate BayesEvolve on shifted BBOB-style black-box optimization tasks, leaving program and laboratory discovery domains to future work. BayesEvolve improves sample efficiency over memory- and archive-guided LLM baselines under a fixed evaluation budget. We further show that the belief state is predictive on held-out candidate pools, that controlled decision-rule ablations favor belief-guided selection with an annealed uncertainty bonus, and that BayesEvolve exhibits productive late-stage concentration rather than unfocused exploration.

Traditional automatic evaluation methods have been shown to be unsuitable for modern Chinese poetry because of the distinct nature of this literary genre. Human evaluation remains reliable, but is expensive and not applicable to large-scale data. In this paper, we propose Poller (Poetry LLM Evaluator), a novel method leveraging large language models (LLMs) to evaluate the poetry understanding task. Specifically, our method requires LLMs to play the role of a poem's author with detailed information, thereby emulating human evaluation and judgment by adopting the poet's perspective. We conducted comprehensive experiments on multiple LLMs, evaluating the interpretations of poems across eight specialized dimensions. Experimental results demonstrate that our method effectively reduces the evaluation error between LLMs and humans. Especially for specific dimension evaluation, Poller-based LLMs achieve a 94.55% and 89.53% error reduction for rhetorical techniques and defamiliarization, respectively, compared to baseline methods. These performances are unattainable by conventional LLM evaluation methods. Experimental results from multiple LLMs across various dimensions validate the efficacy of our method. This work bridges the gap between automated efficiency and human expertise, establishing a foundation for automated evaluation in poetry-related tasks.

In this paper, we aim to combine the advantages of existing sequence parallelism paradigms and overcomes their drawbacks, the most serious of which is the incapability to correctly compute causal attention on the hybrid-context packed sequences, in a stronger sequence parallelism framework. The practical technique of packing sequences for efficiently pretraining and fine-tuning large language models causes cross-contamination problem in attention computation, which can be effectively solved when no parallelism in the sequence length dimension is taken. However, in sequence parallelism, existing approaches either ignore the scenario of hybrid-context sequences or conversely sacrifice and limit parallelism degree for supporting the scenario. To this end, we innovatively propose an efficient Sequence-Aware Parallelism algorithm to conquer the obstacles of intensive tensor transmission and partial attention computation across multiple device groups. Our algorithm utilizes JIT (Just-In-Time) compilation to optimize the communication strategy of all device groups in NCCL level. Further, we integrate existing sequence parallelism paradigms into a Hierarchical Sequence-Aware Parallelism framework which benefits from our sequence-aware algorithm. We additionally elaborate on the memory and communication overhead management of the hierarchical framework to optimize its performance. Through multiple experiments, we demonstrate that our proposed approach outperform other state-of-the-arts sequence parallelism approches in multiple metrics.

Conservative offline training is widely advocated as a safe foundation for subsequent online adaptation: if a policy stays close to well-supported behaviour, the argument goes, it is less likely to exploit imperfections in a learned reward model. We challenge this intuition empirically and mechanistically. We train a Qwen3-14B policy under Direct Preference Optimisation (DPO) with three levels of conservatism ($β\in \{β_{\mathrm{lo}}, β_{\mathrm{mid}}, β_{\mathrm{hi}}\}$ derived from empirical log-ratio percentiles), then adapt each checkpoint online against a learned reward ensemble (3\,$\times$\,Qwen3-1.7B) while measuring true performance on GSM8K exact-answer accuracy. We find that higher offline conservatism monotonically increases reward-hacking damage, measured by the Goodhart gap and its area under the curve (AUGC), with Spearman $ρ= 1.0$ across all three conditions. Mechanistic analysis reveals a three-link causal chain: (i) high-$β$ DPO compresses policy entropy, (ii) Low-entropy policies generate responses with reduced diversity, concentrating in a narrow region of the reward model's training distribution (lower pairwise cosine distance), and (iii) despite this proximity, ensemble disagreement (epistemic uncertainty) increases with $β$ and is exploited faster during online optimisation. We further fit a power-law curve to the $(β, \augc)$ data and identify a practical optimal conservatism level $β^{\star}$ that balances alignment fidelity against hacking vulnerability. Our results suggest that the field needs calibrated, not maximal, conservatism.

We introduce FlexTab, a flexible encoder-decoder architecture for in-context learning on tabular data that pairs a single, task-agnostic encoder with a suite of task-specific decoders. Unlike existing tabular in-context learners, which entangle feature representations with a specific prediction target, our design produces target-agnostic row embeddings that can be leveraged across a wide range of downstream tasks within a table-native in-context learning setup. We demonstrate this flexibility on six distinct problems: classification, regression, anomaly detection, clustering, entity matching, and entity classification in relational databases. Both the encoder and the task-specific decoders are trained on a large corpus of real-world, unlabeled tables. FlexTab achieves state-of-the-art performance on classification, regression, anomaly detection and entity matching, while remaining competitive with specialized models on entity classification in a relational setting. These results demonstrate that a single shared encoder, paired with task-specific decoders, can serve as an effective general-purpose backbone for diverse tabular prediction problems. The inference code and checkpoints will be made publicly available at https://github.com/SAP-samples/flextab.

Aligning large language models (LLMs) with diverse user preferences is a critical yet challenging task. While post-training methods can adapt models to specific needs, they often require costly data curation and additional training. Test-time scaling (TTS) presents an efficient, training-free alternative, but its application has been largely limited to verifiable domains like mathematics and coding, where response correctness is easily judged. To extend TTS to preference alignment, we introduce a novel framework that models the task as a realignment problem, since the base model often fails to sufficiently align with the stated preference. Our key insight is to decompose the underlying reward function into two components: one related to the question and the other to preference information. This allows us to derive a REAlignment Reward (REAR) that selectively rescales the proportions of these two reward terms. We then show that REAR can be formulated as a linear combination of token-level policy log-probabilities, making it computationally efficient and easy to integrate with various TTS algorithms such as best-of-$N$ sampling and tree search. Experiments show that compared to other test-time baselines, REAR not only enables scalable test-time realignment for preference alignment tasks under diverse user requirements, but also generalizes to mathematical and visual tasks under appropriate preference settings.

Sparse Autoencoders (SAEs) are widely used to interpret large language models by decomposing activations into sparse, human-understandable features, but scaling to large dictionaries exposes fundamental challenges. Systematic studies reveal pervasive feature splitting that fragments coherent concepts into non-atomic latents and widespread feature absorption that creates arbitrary exceptions in general features, severely compromising latent reliability. These issues stem from inconsistent latent assignment across samples: without cross-sample constraints, per-sample optimization often allows a single underlying concept to be inconsistently distributed across multiple redundant or interfering latents. To address this, we introduce C$^2$R (\underline{C}ross-sample \underline{C}onsistency \underline{R}egularization). C$^2$R explicitly encourages that each semantic feature is consistently represented by a unified latent across the batch by penalizing the co-activation of directionally similar latents. Comprehensive evaluation demonstrates that C$^2$R effectively mitigates both splitting and absorption while, crucially, preserving reconstruction fidelity, providing a principled solution that enhances latent interpretability without degrading model performance. Source code is available at https://github.com/hr-jin/Cross-sample-Consistency-Regularization.

Current large language models are extraordinary statistical engines. They compress vast amounts of text into useful patterns and can explain science, write code, imitate reasoning, and participate in philosophical conversation. Yet pattern mastery is not the same as general intelligence. A human infant begins with little explicit knowledge, but gradually discovers object permanence, cause and effect, other minds, bodily agency, and the persistence of the physical world. We make an argument that the path to artificial superintelligence (ASI) depends on a missing capacity we call situation perception: the ability to construct, revise, and act within internal simulations of possible worlds across latent time. perception requires at least three core components: abstract prediction, long-term compressed memory, and active learning guided by objectives. In this work, we analyse why modern large language models remain incomplete, and propose the appropriate tests for measuring progress and consequences of machines that can simulate futures, pursue self-directed goals, and possibly judge their own creators.

Delayed generalization (\ie~grokking) refers to the phenomenon in which a neural network fits its training data early in training but only begins to generalize after a prolonged delay, often through an abrupt transition. Despite extensive empirical study, its underlying mechanism remains poorly understood. In this work, we first theoretically characterize a shell--core topological configuration of the reachable solution space induced by Adam's optimization dynamics with weight-shrinkage regularization, supported by empirical evidence. This optimization-induced topological configuration gives rise to grokking. In model's parameter space, random initialization solutions concentrate on a thin outer spherical shell, enclosing another spherical shell of memorization solutions, which in turn contains a core corresponding to the generalization solutions. Leveraging stopping-time theory, we then analyze the geometry of this topological configuration and the solution transition time at which optimization trajectories escape the memorization manifold and first reach the boundary of the generalization manifold. Our theoretical analysis derives grokking scaling laws for the learning rate, batch size, and $\ell_2$ regularization coefficient, which are further validated through experiments and shown to recover results from prior literature.

With rapidly improving capabilities, Large Language Models (LLMs) are increasingly used in many complex real-world tasks. Beyond requiring in-depth knowledge and reasoning skills, many of these tasks exhibit a high degree of subjectivity and require that the outputs of the model can be trusted. While a lot of progress has been made to train better models, decision-making algorithms have received less attention. In this work, we present and evaluate various uncertainty-aware decision-making algorithms based on Bayesian decision theory and risk-averse decision making on the tasks of tutoring and automatic peer reviewing. Concretely, we take uncertainty over tutoring strategies and review scores into account when generating a tutor response or review and use conformal prediction to provide guarantees over strategy and score. We find empirically that these algorithms can improve the utility of the generations but need to be carefully implemented when ambiguity is high. For example, risk-averse rules can degrade performance by optimizing for generic outputs, while Bayesian methods tend to perform better. Our work uses techniques from decision theory to improve LLM-based decision-making and outlines open challenges for the community.

Quantitative research across the social and behavioral sciences depends on human subject experiments that are expensive, slow, and subject to sampling bias. Here we show that pretrained large language models induce risk-equivalent estimators of conditional expectations under squared loss, establishing restricted functional risk equivalence: under squared loss, the LLM induces an estimator whose risk matches the Bayes optimal risk for squared-loss prediction of conditional expectations for any inference that depends on the data only through the conditional mean. We formalize the LLM as a misspecified functional estimator $T(\hat{P}_n)$ trained on i.i.d.\ data, decompose the estimation error into representation bias $ε_{\mathrm{rep}}$ and optimization error, and prove that under mild regularity conditions the LLM's expected error converges to the irreducible population variance plus the squared representation bias, with the representation bias bounded by the Pinsker inequality. The identifiability error $δ$ propagates into the effective bias, inflating the asymptotic risk floor. We establish restricted functional risk equivalence via a bidirectional Le Cam deficiency analysis: the forward deficiency vanishes asymptotically while the reverse deficiency is exactly zero. We provide finite-sample concentration bounds and a calibration protocol with explicit decision rules. The result is a precise, provable statement: a well-calibrated LLM achieves the Bayes-optimal risk for conditional-mean-dependent inference, bounded by explicit scope conditions. In practical applications, this means that under satisfied conditions and well-calibrated models, large language models can be used in many prediction and decision-making tasks that originally relied on human experiments, approximating near-optimal statistical inference at lower cost.

Researchers and practitioners increasingly apply Large Language Models (LLMs) for automated vulnerability detection. Recent work has shown that LLMs are susceptible to the same cognitive heuristics that bias human judgment. Yet, no work has investigated whether these heuristics affect a model's assessment of code vulnerabilities. In this paper, we present the first systematic exploration of cognitive heuristics in LLM-driven code vulnerability detection. We introduce a controlled framework that holds the code fixed and only varies the surrounding context to trigger three cognitive heuristics: the halo effect through author attribution, the framing effect through task objectives and consequences, and the anchoring effect through prior analysis results. Within this framework, we evaluate eight LLMs across three programming languages and perform both quantitative and code-level analyses. Our findings demonstrate that all evaluated models are susceptible to these heuristics. Cross-model average susceptibility is highest for framing at 33.2%, followed by anchoring at 23.5% and halo at 18.4%. Code-level analysis reveals that vulnerabilities that require semantic reasoning for detection are more susceptible to cognitive heuristics than those identifiable through pattern matching. Furthermore, models often change their verdict from safe to vulnerable based on the cognitive condition, without accurately identifying the actual vulnerability. To highlight the practical impact, we demonstrate a proof-of-concept black-box cognitive attack that can suppress up to 97% of previously detected vulnerabilities. These findings indicate that cognitive susceptibility is a consistent and exploitable property of LLM-based vulnerability detection.

Reinforcement Learning (RL) is an important paradigm for improving the reasoning capabilities of Vision-Language Models (VLMs). However, directly applying RL to rollout multimodal reasoning can lead to instability, due to the exploitation of language priors, the neglect of visual evidence, and the generation of reasoning traces that are fluent yet not visually grounded. The question arises: Can initially steer the policy toward visually faithful reasoning regime before applying reinforcement learning? To this end, we propose a Faithful Warm-Start (FWS) strategy that first curates samples with explicit vision-language causal relationships from six general VQA benchmarks to construct the FaithfulQA dataset, where each of the image-question pairs gains a certain degree of visual observations, question requirements, commonsense knowledge, domain knowledge, and the final answer. Subsequently, a VLM-based judge is employed to further purify the dataset, ensuring strong causal consistency and visual faithfulness. This warm-start stage equips the model with the capability to understand causally grounded vision-language patterns before subsequent RL optimization under sparse answer-level rewards. Experimental results show that such faithful supervision improves answer accuracy, stabilizes RL training, and reduces visually unsupported reasoning.

Pairwise preference data is widely used for training and evaluating language models (e.g., RLHF), but each datapoint records a choice, not the rationale behind it. Methods such as Inverse Constitutional AI (ICAI) attempt to improve interpretability by compressing datasets into short ``constitutions'' of natural-language principles. We argue this framing is under-specified: a flat list of principles is not yet an executable decision rule because it leaves principle composition implicit. We use the pairwise setting as a testbed to empirically characterize three open problems in constitutional methods. First, principle quality is hard to measure: coverage and accuracy are useful but incomplete proxies for end-to-end reconstruction. Second, composition is ambiguous: holding principles fixed, different executors (LLM judge versus majority vote) agree only $73%$ of the time. Third, constitutions differ between LLMs: cross-model vote agreement is $73%$, whereas intra-model agreement is $81%$. Across PRISM, AlpacaEval, and Chatbot Arena, we show that principle refinement (ICAI+) may be a first step towards ameliorating these problems: inter-executor agreement rises to $78%$, and transparent executors match LLM judge accuracy ($66%$ vs.\ $67%$). Our results highlight that constitutions should be evaluated as constitution--executor systems, with implications for LLMs-as-a-judge broadly.

Bash script comprehension is challenging due to Bash's syntactic freedom and complex command structures. Despite its critical role in system administration, Bash scripts often lack adequate comments, hindering readability and maintainability. Existing automated comment generation approaches face two main challenges: (1) limited training datasets that inadequately represent real-world Bash usage patterns; and (2) insufficient understanding of Bash-specific concepts by Large Language Models (LLMs). To address these, we propose Bash-Commenter, an advanced comment generation method based on LLaMA-3.1-8B. First, we construct a comprehensive dataset of complex, multi-line Bash scripts with high-quality comments. Second, we conduct Continual Pre-training (CPT) on large-scale Bash data, followed by Supervised Fine-tuning (SFT), strengthening the model's foundational knowledge of Bash syntax and semantics. Finally, we introduce Syntax-Aware Preference Optimization (SAPO), which constructs preference pairs by applying atomic operations to a script's Abstract Syntax Tree (AST), creating minimal pairs of correct and subtly incorrect scripts for fine-grained semantics learning. Our method outperforms state-of-the-art baselines, achieving 33.40% BLEU-4, 58.26% METEOR, and 57.03% ROUGE-L for 1,064 single-line commands, and 22.15% BLEU-4, 43.89% METEOR, and 32.80% ROUGE-L for 1,046 multi-line scripts. Human and LLM evaluations further confirm superior comment quality in correctness, completeness, and naturalness.

The alignment of language models is typically studied through the lens of capability benchmarks, but the dynamics of how models change during post-training remain poorly understood. We argue that the physical sciences, and thermodynamic phase-transition theory in particular, offer a principled and underexplored vocabulary for reasoning about these dynamics. As a case study, we instantiate this position through the lens of material Crystallization, which is a well-studied thermodynamic phase transition. For tasks like random number generation, this breaks into 3 phases: (1) the high entropy liquid phase in the pretrained model, with many distinct sampling distributions promptable from the model; (2) the nucleation phase caused by supervised finetuning, in which behavior collapses onto a single seed distribution present in the pretrained LLM; and (3) a settling phase in which reinforcement learning techniques redistribute probability of the collapsed distribution, but largely keep it concentrated on the same options as the seed distribution. We propose intuitive metrics to verify the transitions between these phases, and validate the idea across a range of random tasks. Crystallization is one instance of a broader class of physical frameworks we believe alignment research should import to answer questions about where alignment-induced structure comes from, why it converges where it does, and what it fundamentally cannot change.

A key step toward artificial general intelligence is to train models that can perform multiple tasks. In this paper, we study how to build such models by first training separate RL experts for individual tasks and then consolidating them via distillation, as an alternative to directly training a single model on mixed tasks. We show that off-policy distillation degrades in multi-task settings due to the mode-covering nature of forward KL: aggregating data from multiple tasks introduces a large number of behavioral modes that can exceed the student's capacity, forcing it to average across behaviors and leading to degraded performance. In contrast, on-policy distillation is mode-seeking but requires strong initialization. Inspired by these observations, we propose a two-phase approach: off-policy distillation followed by on-policy refinement. Evaluation across conversational agents and text-based games confirms that this two-phase approach matches single-task RL expert performance for each individual task, whereas off-policy or on-policy distillation alone fails to match this performance.

Score distillation turns a pretrained 2D diffusion model into a 3D generator, but the per-step gradient is estimated from a single randomly chosen view: it is high-variance and blind to global shape consistency. Prior work addresses this by retraining the diffusion prior on multi-view data; this improves consistency but makes the sampling contribution inseparable from prior quality. We instead isolate the sampling axis. The per-step gradient is one noisy sample of an expectation over views; aggregating K samples per step at a fixed total UNet budget reduces variance without touching the prior. We introduce Multi-View Aggregated Score Distillation (MV-SDI), which aggregates gradients from K views per step via gradient accumulation, keeping peak memory unchanged and the 2D prior frozen, and draws views as antithetic antipodal pairs, a prior-independent geometric property, for balanced angular coverage. At a fixed 10,000-UNet-call budget, K=2 raises CLIP R-Precision from 74.8% to 83.8% and CLIP score from 0.297 to 0.312, with consistent gains on HPSv2 and ImageReward and a 0.0% divergence rate on the 43-prompt benchmark; optimization steps halve as a consequence. K=4 gives a fourfold step reduction at R-Precision 86.9% and CLIP 0.307, still well above the single-view baseline on every alignment metric. MV-SDI is compatible with gradient-based score-distillation pipelines, including Score Distillation via Inversion, and requires no retraining and no multi-view data.

Knowledge distillation (KD) is a key technique for compressing Large Language Models (LLMs), yet methods relying on a single KL objective often fail to balance primary distribution fitting with long-tail probability modeling, limiting both generation quality and generalization. To address this, we analyze the complementary roles of forward and reverse KL divergence (FKL/RKL) in distribution alignment from theoretical and empirical perspectives. We then propose a reinforcement-learning-based adaptive KL-weighted distillation framework, in which a policy network dynamically assigns weights to FKL and RKL based on teacher-student distributional characteristics, guided by immediate reward signals to achieve dual alignment on principal and long-tail modes. Extensive experiments demonstrate consistent improvements across Rouge-L and BertScore metrics, surpassing greedy heuristics by 0.4-0.6 points and outperforming other baseline methods on diverse benchmarks.

Recent progress in humanoid behavior models has been driven in large part by abundant human motion data, but comparable motion data is scarce for non-humanoid legged robots such as quadrupeds, hexapods, and quadruped manipulators. A promising alternative is to repurpose human motion across embodiments; however, direct retargeting often produces motions that are visually plausible yet physically inconsistent or difficult to track under robot dynamics. We present X-Morph, a human-motion-to-robot-behavior pipeline that converts human motion into deployable locomotion and loco-manipulation policies for diverse non-humanoid legged morphologies. A cross-morphology retargeting stage converts human motions into kinematically plausible, intent-preserving robot references, which are then tracked by a privileged RL policy and distilled into a causal student policy. We evaluate X-Morph on three morphologically distinct platforms: a quadruped, a hexapod, and a quadruped equipped with a manipulator. The resulting policies track diverse retargeted motions, generalize to unseen human motions, and support downstream use cases including video-based teleoperation, behavior-prior control, and text-conditioned motion generation. These results suggest that large-scale human motion can serve as a substrate for learning broad, reusable behavior priors beyond humanoid robots. Project page: https://maker-rat.github.io/morph/