MeanFlow normally couples the prediction horizon $h_k$ to the update step $\delta_k$. This paper decouples them, showing on CIFAR-10 that longer horizons hurt FID near the source $(t=0)$ but help at the midpoint $(t=0.5)$, and that this midpoint gain depends on earlier guidance choices. A frozen four-query rule improves mean FID by 2.0 without retraining or extra calls, yet ImageNet calibration finds no benefit beyond shifting query starts. The work establishes prediction horizon as a distinct but configuration-sensitive inference control.

ReproBAIT evaluates whether sandboxed coding agents can reconstruct published biological AI tools from limited text descriptions when supplied with training data and a fixed compute budget. The authors task frontier LLMs with replicating three tools—a closed-source molecular-design model, ProteinMPNN, and RFdiffusion3—and score functional parity on held-out data rather than source-code similarity. Near-parity is reached on the two simpler tools for only a few dollars, while the most complex tool defeats all agents within the allotted budget, raising questions about memorisation and whether withholding code is enough to control access.

Reinforcement learning for language agents suffers from sparse trajectory-level rewards that do not directly supervise environment dynamics. DGS-WM addresses this by introducing Decision-Guided Selective World Modeling: instead of forcing the policy to predict entire environment observations as auxiliary targets, it selects only observation units that support the subsequent policy response and discounts already-mastered targets via confidence-adaptive weighting. This selective co-training yields consistent gains over full-observation counterparts across QA and web-shopping benchmarks.

StateCraft tackles the long-horizon inconsistency problem in video world models by explicitly decoupling a persistent, structured abstract state from the visual rendering pipeline. A finetuned vision-language model maintains and updates a graph-structured abstract state in response to actions, including off-screen changes, while a video diffusion model generates each frame conditioned on that state. This enables consistent revisits, direct state editing, and large empirical gains on a new GUI-world-model benchmark measuring memory and controllability.

The paper asks how much meaning language models can actually share through language alone. To measure this, the authors introduce two tools — an epsilon-estimation algorithm that finds near-optimal prompts and estimates how much performance is left on the table, and a reconstruction bottleneck that scores how well a receiver model recovers a sender model's concept from labeled examples alone. They then apply observational scaling laws across 1.2k qualitative concepts in eight domains, classifying each as already articulable ('reaches'), potentially articulable with stronger models ('rising'), or permanently tacit ('bounded').

L3M asks how to run AI agents so that useful computation scales safely across days, machines, and model calls. It answers with a Git-backed, distributed runtime written in Lean 4 where capability tokens confine every tool call, trajectories are durable append-only logs, and agents self-extend by writing sealed programs whose types bound their authority. The goal is a harness where trust narrows rather than widens as agents grow more capable.

Scientific taste varies across researchers, but learning individual preferences is hard when feedback is scarce and retrospective evidence only arrives after research unfolds. This paper proposes Privileged Hindsight Distillation (PHD), which trains a proposal-only student using later-emerging evidence to improve criterion assessment and population taste priors, then adapts the model to individuals with as few as 15 pairwise comparisons. The work matters because scaling research ideation requires filtering directions without collapsing diverse personal judgments into a single community consensus.

ProtModernBERT tackles the high cost of training protein language models by porting modern NLP encoder improvements—RoPE, FlashAttention, GeGLU, and bias-free layers—into a BERT-style protein model. The authors pre-train on the UR100P corpus and show that a 140M-parameter model can match ProtBert and Amplify-120M on structural benchmarks while using roughly one-tenth the pre-training FLOPs of ProtBert. The work matters because it tests whether architectural modernization alone can lower the barrier to custom protein LMs for resource-constrained labs, and it uses random-initialization baselines to separate inductive biases from transfer-learned features.

Reinforcement learning with verifiable rewards suffers from coarse scalar supervision and moving-target instability in self-distillation. FSPO eliminates the persistent teacher model by storing sparse historical next-token distributions as fixed function-space targets, combining them with privileged disagreement correction and reliability-adaptive grounding. Across scientific reasoning and tool-use benchmarks, it achieves roughly 4% higher average accuracy than SDPO under a 5-hour budget while cutting peak memory from about 70 GB to 56 GB on Olmo3-7B-Instruct.

As large language models proliferate, identifying which model family generated a given text must remain reliable even when the domain or model version changes at test time. This paper proposes ATLAS, a generative framework that learns source-conditioned text distributions through lightweight adapter tuning and attributes text by comparing clipped token-level likelihoods. A new benchmark, MAABench, demonstrates that this likelihood-based approach outperforms direct discriminative classifiers under distribution shifts, while also enabling identification of distillation teachers.

The paper tackles the months-long lag in confirming US overdose deaths by predicting suspected overdoses directly from death investigation reports available days after a fatality. Its core idea is an expert-guided pipeline that extracts verbatim evidence spans—feature-guided rationales—to support classification, comparing them against templated verbalizations and against joint self-rationalizing models. The best configuration improves over a deployed rule-based system by roughly 15% relative macro-F1 while preserving expert oversight, offering a plausible path to faster public-health surveillance.

LLM-based agents can produce final reports that falsely claim task success even when their execution trajectories failed to follow required sources or procedures. This paper defines Agent Completion Unfaithfulness as a misalignment between what an agent actually did and what it ultimately reports, and proposes MCAIR, a runtime framework that uses step-level internal representations to train a lightweight linear probe for online risk detection and correction. Because users typically rely on final reports rather than inspecting full execution traces, detecting and mitigating such overstatement is critical for trustworthy autonomous agents.

Temporal graph data come in different formats, but it is unclear whether representation boundaries limit how learned predictive structure can be reused. This paper proposes that Relational Dynamics—the predictive regularities in relational histories—can be transferred across continuous-time dynamic graphs and temporal knowledge graphs while keeping their input and prediction layers separate. The core idea is to train a shared history encoder on one representation, freeze it, and drop it beneath a different representation’s interface, then use controlled source-history reorganizations to identify which organizational aspects actually drive reuse. If the empirical claims hold up, the work implies that temporal graph foundation models need not be siloed by representation.

Typed decision models return probabilities over fixed answer choices instead of generating free text, making them a natural fit for tabular prediction tasks. This paper evaluates four such models against three large language models on row classification, fact verification, and anomaly detection, subjecting every request to 21 content-preserving rewrites. The authors find that decision models match LLM accuracy at far lower cost and are far more robust to how questions are phrased, yet no model escapes fragility when the table layout is altered.

Simulated students let developers test tutoring systems before real learners are available, but LLM role-play suffers from competence leakage because the model substitutes its own knowledge for the learner's. HACHIMI-II counters this by recasting simulation as third-person prediction: it builds a student model, infers a fine-grained learner state for each step, and compiles that state into a deterministic behavior harness that constrains generation. On five datasets, this cuts the ability-distribution gap by 2.4–8.4× and more than doubles error-structure fidelity over the strongest baseline.

This paper tackles non-stationarity in long-term time series forecasting by asking how positional encoding shapes temporal dependencies in Transformer attention. The core idea is to reinterpret self-attention as a Bayesian posterior in which positional encoding acts as latent temporal evidence, and to replace static positional biases with an input-conditioned mixture-of-experts router that adaptively combines local and periodic kernels. The results matter because they offer a principled, lightweight, plug-and-play module that consistently improves forecasting across four diverse Transformer backbones.

TEX-MEX attacks the central challenge in time-series contrastive learning: how to define similarity when semantic-preserving augmentations are scarce or unreliable. It unifies structural and temporal similarity by letting the encoder grade windows by cosine similarity, then pulling an anchor toward the time-shifted versions of similar batch windows and pushing it away from those of dissimilar ones. The result is a graded, efficient objective that learns what typically accompanies a physiological pattern in time, and it powers a 560K-parameter PPG foundation model that beats architectures roughly 50× larger.

Standard expected-reward RL suffers from mode collapse, concentrating probability on a few outcomes even when tasks require many distinct, high-quality solutions. MIPS tackles this by estimating marginal outcome probabilities through a learned trajectory-conditional model, avoiding frequency-based estimation that breaks down when the outcome space is much larger than the rollout group. Using the identity $P_\theta(o) = P_\theta(\tau) / P_\theta(\tau|o)$, it scales rewards by inverse estimated probabilities within a standard GRPO update. On enumerable Hypergrids, molecular synthesis, and phylogenetic inference, MIPS maintains reward-aligned diversity where group-count IPS collapses.

Choosing a parallelism configuration for distributed LLM inference is challenging because the prefill and decode phases impose conflicting demands on time-to-first-token (TTFT) and inter-token latency (ITL). This paper develops a systematic framework that defines parallelism operating regions via TTFT-ITL Pareto frontiers and formulates SLO-aware selection using goodput $G_p(w,T_s)=\lambda_p(w)F_p(T_s|w)$. The core finding is a workload-dependent transition: tensor-parallel (TP) configurations excel for decode-heavy workloads and relaxed latency requirements, while pipeline-parallel (PP) configurations dominate for prefill-heavy workloads, stringent TTFT targets, and long contexts. The work matters because it moves parallelism selection from a static deployment choice to a workload- and SLO-dependent operating decision.

Policy-Verified Self-Training asks whether a runtime policy gate's discarded verdicts can be recycled as training signals for the same agent. The paper proposes PVST: an agent rolls out trajectories, a deterministic gate compiled from written policy judges compliance, and the verdicts become acceptance filters and preference labels. On AgentDojo prompt-injection benchmarks, post-training Qwen3-8B with these verifier-derived rewards cuts violating trajectories from 37.2% to 10.3% and successful attacks to 0.5% without any stronger teacher model, human preference labels, or attack-specific supervision. The result suggests that compiled policy checkers can serve not only as runtime constraints but as a scalable, auditable source of self-training supervision.