Reinforcement learning is increasingly used to align image generators with reward signals, and Flow-GRPO recently extended this paradigm to flow-matching models by treating the denoising sampler as a stochastic policy that can be optimized from reward feedback. Training in this setting is unstable in a way specific to multi-step denoising: the policy update changes systematically across denoising steps, with importance ratios drifting below one, becoming increasingly dispersed, clipping at different rates, and leaving fewer usable samples late in training. Prior work treats these effects as separate failure modes and addresses each with a hand-tuned stabilizer. We show instead that they arise from a single per-step quantity, which we call path variance. This quantity is determined exactly by the sampler's Gaussian transition kernel and can be estimated cheaply during training. This reframes instability as a resource that can be measured and budgeted rather than a collection of symptoms to repair. Our method, λ-Controlled GRPO, calibrates importance-ratio behavior from this predicted law rather than from noisy empirical statistics, and allocates gradient effort across denoising steps according to their predicted cost. The two scales governing the update are fixed by standard policy choices rather than introduced as free tuning parameters. On a text-to-image model under two reward settings, rendering difficult target text scored by optical character recognition and matching human preferences scored by a preference model, λ-Controlled GRPO improves both text accuracy and preference reward over the strongest empirical stabilizer. It also keeps late-step path variance within its intended budget, precisely where the baseline systematically overshoots. The result is a Flow-GRPO update calibrated by its own transition law rather than stabilized after instability appears.
We need studies on conversational AI (CAI) at scale to understand human behavior and shape CAI design. However, fragmented reporting of systems and study configurations hinders replication, extension, and knowledge accumulation. We present Gricea, an open-science platform representing studies as configurable, deployable research artifacts that researchers can run, inspect, share, and reuse. Informed by a formative analysis of prior CAI research, Gricea couples study procedures, participant-facing systems, and conversational task behavior in. In a replication study using Gricea, we replicated configurations 93% of eligible CUI 2026 papers; while also flagging missing information in 96% of papers that hinder faithful replication --- further motivating Gricea's need. In a user study, researchers and practitioners from diverse backgrounds successfully constructed runnable studies addressing various open-ended research questions. Together, these findings demonstrate Gricea's support for constructing, reproducing, and extending CAI studies through shared research artifacts, enabling cumulative knowledge building through open science.
Assine o Pro, sem anúncios
Faça upgrade para uma leitura sem interrupções e acesso prioritário a novas fontes.
We introduce QuranicMMLU, a benchmark for evaluating generative AI on Quranic Arabic across multiple dimensions of linguistic complexity. Existing Quranic benchmarks center on general question answering and semantic retrieval, without probing specific linguistic competencies or stratifying by cognitive demand and verse difficulty. We construct a five-pillar Quranic taxonomy spanning Phonology, Morphology, Syntax, Semantics, and Pragmatics, with 31 leaves covering phenomena from tajwīd and root-and-pattern morphology to occasions of revelation and inter-surah coherence. For each leaf we generate questions stratified by Bloom's cognitive level and verse perplexity, then have LLM as a judge to independently answer and score every item and route the annotations to manual review. The resulting dataset comprises 980 human-reviewed questions, each issued in both open-ended and multiple-choice form. We benchmark 12 systems on these items and find that the Islamic-specialized model leads, yet every system scores higher on multiple-choice accuracy (average 84%) than open-ended answer quality (average 60%): the two rankings agree closely (Kendall's τ=0.73), but multiple-choice scoring hides failures that surface only once answer choices are removed. QuranicMMLU thus offers a rigorous, linguistically grounded framework for evaluating Arabic NLP in the Quranic domain.
A week after an Anthropic researcher’s doomsday warning rattled the AI world, the company’s CEO Dario Amodei has outlined his plan to “pace the frontier”…
Every multiparameter persistence vectorization we know of carries a one-sided Lipschitz upper bound and nothing below it: without a lower gauge there is no sense in which the features are faithful, and no per-prediction guarantee can be built on them. This paper supplies the missing side. COMPLEX is a closed-form, training-free embedding of multiparameter modules -- slice the module along a fixed near-diagonal net, embed each slice barcode by the certified PLACE/PALACE landmark map, concatenate. Under a checkable witnessing-slice coherence condition, holding on 100% of audited pairs on Orbit5k, a single slice carries a closed-form lower gauge: separated modules stay separated in the embedding. With the standard upper bound this gives, to our knowledge, the first two-sided distortion bound for a multiparameter feature map, making faithfulness measurable. Measuring it, we find the floor tight within a small factor of realized distances yet operationally local: an RBF-SVM reaches 91% where 1-NN reaches 78% on the same features. Local per-prediction certification therefore fails for a structural reason common to every landmark embedding whose lower gauge is witnessed by one coordinate. With no learned embedding and no held-out calibration -- only a cross-validated SVM head -- COMPLEX sets the state of the art on both Orbit benchmarks (91.95% on Orbit5k, 92.98% on Orbit100k), level with or above Euler-characteristic surfaces and above transformers and graphcode. On graphs it exceeds GRIL on all four shared molecular benchmarks with one fixed configuration, including the only multiparameter method to clear COX2's majority baseline by more than three points. Closed-form selection -- of the landmark radius, the kernel (certificate-preserving), and the bifiltration set -- buys further accuracy; gradient-shaped adaptation buys none.
Vision-language models (VLMs) rely on storing and transferring appropriate information across their sub-components. Verifying that the VLMs exhibit desired behaviours, while avoiding harmful ones, is central to their reliable deployment. Yet, methods that identify VLM behaviours remain scarce. We present DiaVLo, a diagnostic framework that leverages human curation and VLMs' generation capabilities to construct specifications of desired and observed VLM behaviours, surfacing potential misalignments. Beyond this, DiaVLo also provides causal estimates to identify the most influential concepts steering VLM behaviours. We evaluate DiaVLo on several open-source VLMs under both classification and generation conditions. Our experiments show that DiaVLo produces behaviour labels that correlate with model performance and provide context for measured performance. DiaVLo surfaced behaviours that are clearly aligned and misaligned, alongside patterns in how VLMs perceive, organise, and prioritise concepts.
A week after an Anthropic researcher’s doomsday warning rattled the AI world, the company’s CEO Dario Amodei has outlined his plan to “pace the frontier”…
California Gov. Gavin Newsom (D) is positioning the state to take the lead on AI oversight, including the potential to mandate a "kill switch" for…
Gating the value pathway of attention reportedly improves language model pretraining, and prior studies disagree on why. We argue and provide experimental evidence that such gates supply two different things that softmax attention lacks: abstention and noise filtering. The first is abstention, which allows an attention head to output nothing, bypassing the requirement that attention weights must sum to one. The second is noise filtering, which allows the value pathway of an attention head to suppress interference from superposed features in the residual stream. In our experiments in matched models from 10M to 350M parameters, we supply abstention through a learned per-head sink logit in the softmax and noise filtering through a gate on each value. We report three empirical findings. First, the benefit of abstention, measured as the reduction in validation loss relative to a matched baseline, declines as models grow, whereas the benefit of noise filtering increases with scale. In particular, abstention accounts for nearly all of the gain from gating at 10M and filtering for most of it at 350M. Second, the best model at every scale is the one with both primitives built in. Third, injecting controlled interference into the values a head reads confirms that the gate removes such interference, and reveals that each of the two gate forms we study has a characteristic blind spot. Supplying both primitives adds negligible parameters and remains compatible with the key-value cache.
Computer-use agents (CUAs) have advanced along two separate lines: graphical interaction and software development through code and the command line. Real digital work requires both, interleaved rather than stacked end to end. We study hybrid CUAs that autonomously decide when to explore an interface, implement software, and run and visually verify their artifacts. We introduce RecreationWorld, a five-platform framework built around recreation: given a running reference, an agent must discover its behavior and build a faithful implementation with no prescribed workflow. RecreationWorld provides reproducible environments on Ubuntu, macOS, Windows, Android, and Web, plus a unified harness with native GUI control and coding tools. The running reference serves as an oracle for hidden behavioral tests, providing execution-grounded rewards. We scale trajectory generation with high-quality open-source applications. Models trained on these trajectories improve across five out-of-distribution coding and hybrid computer-use benchmarks and more frequently verify their rendered outputs, providing evidence of transfer beyond recreation. For held-out evaluation, we introduce RecreationBench, comprising 250 diverse tasks across domains and platforms. Reference-grounded programmatic and visual assertions cover action-conditioned outcomes at multiple interaction depths; each is validated on the reference and by human reviewers before the suite is frozen for automatic scoring. GPT-6 Astra leads at 58.1% overall, but passes all programmatic tests on just 2.8% of tasks. Agents reproduce static interface structure more reliably than interactions and computed outputs, while generated applications remain smaller and more monolithic than their references. We release the benchmark, environments, and test suites.
LLM agents in social simulation revise their opinions implicitly, in context: how open an agent is to persuasion can neither be specified nor verified, and collective outcomes inherit the model's training prior. We introduce Bayesian Chronicle Agents (BCA), a minimal belief layer separating what an agent believes from how it speaks. Each stance is a probability, updated by one Bayesian step per utterance heard. A single prior-strength parameter κ encodes stubbornness, modeled after its role in Friedkin--Johnsen (FJ) opinion dynamics. We then sweep this parameter to yield three canonical regimes of opinion dynamics on demand (consensus, persistent disagreement, committed-minority influence), with persistent disagreement matching the FJ closed-form fixed points at R^2\!=\!0.93--0.99. We further show that prescribed κ remains recoverable after the language round-trip, with perfect rank-order recovery across all four models. Explicit belief also makes simulation auditable: the layer surfaces systematic per-model stance biases that end-to-end simulation would silently absorb.
Large language models can hold knowledge they do not report. A model may sandbag on a capability evaluation, or answer against what it internally knows, and its outputs alone cannot tell whether it is hiding an answer or simply does not have one. We borrow the Concealed Information Test, a forensic method that identifies guilty knowledge by presenting a suspect with the true detail among plausible decoys and measuring a stronger response to the item they recognize. Our method, Probe of Internal Recognition (PIR), does the same inside a model. It presents a question with its candidate answers and reads, from the model's internal states, which candidate the model recognizes as correct. PIR is reference-free, needing no honest reference model and no labeled truth corpus. Across eight models from five families (Gemma, Qwen, Llama, Mistral, and Phi), PIR recovers the recognized answer at 0.70 to 0.87 balanced accuracy, well above the 0.28 to 0.40 unknown-item baseline and the 0.25 chance rate. It stays readable across every form of concealment we test, from prompted deception and trained sandbagging to external password-locked and circuit-broken checkpoints, with recognition between 0.85 and 0.93. When the model hides a known answer, recognition stays high. When unlearning removes the knowledge, recognition drops to the level of a question the model never knew. PIR therefore separates a model that will not answer from one that cannot, which supports sandbagging audits and unlearning verification. The signal is causal, adds information beyond black-box behavioral cues, and extends from multiple-choice questions to free-form generation.
The prediction of critical heat flux (CHF), a key safety-related quantity in nuclear thermal hydraulics, remains an important challenge due to its direct relationship with fuel performance and reactor safety. Recent studies have demonstrated that relative to traditional empirical correlations and lookup tables (LUTs), machine learning (ML) methods can substantially improve CHF prediction accuracy. Most ML-based CHF models, however, have been developed and evaluated using tube databases, leaving their applicability to reactor-relevant rod bundle geometries largely unexplored. This study evaluates ML-based CHF models deployed within the CTF subchannel code using the Electric Power Research Institute (EPRI) rod bundle CHF database. Both pure and hybrid residual correction models are considered in local and semilocal formulations. The tube-trained ML CHF models generally transferred favorably to rod bundle applications and outperformed traditional CHF methods across most geometries and operating conditions. The local hybrid LUT model produced the strongest overall performance, and the semilocal pure ML model remained highly competitive. Comparison against the Bowring correlation, W-3 correlation, and 2006 Groeneveld LUT demonstrated that substantial improvements in rod bundle CHF prediction are possible even when models are trained exclusively on tube data. These findings provide one of the first large-scale assessments of ML-based CHF models in square rod bundles within a production-level subchannel analysis environment and support their broader application in reactor thermal hydraulic analysis.
Kimi K3 from Moonshot AI is now available on Amazon Bedrock, giving you a powerful new open-weight option for coding and knowledge work. It offers native vision, a 1-million-token context window, and explicit prompt caching to reduce latency and input costs.
Most work in computational ethics treats annotator disagreement on moral content as noise to be voted away, collapsed into majority vote or the more permissive any-annotator rule the moment a single annotator flags an item. We argue this uncertainty should instead be modeled and learned from. We introduce Moral Entropy, a Bayesian framework that keeps a full posterior over the true label and decomposes its entropy into aleatoric uncertainty (irreducible disagreement about the moral content) and epistemic uncertainty (from insufficient or noisy annotation) -- and lets any heuristic consensus rule be audited against a calibrated ground truth via entropy methods such as cross-entropy/KL, Brier score, and expected calibration error. Across three corpora and fifteen discourse domains, auditing the standard aggregation rules against this posterior reveals bias that no current pipeline reports: the any-annotator rule disagrees with the calibrated posterior on roughly 30% of items -- pooled, almost entirely false positives, though the errors invert at the foundation level (19.9%/38.9% mean FPR/FNR on MFTC) -- while the stricter majority and two-vote rules miss 63-83% of true positives.
Latent flow models have proven to be a reliable and cost-effective method for time series generation. However, the latent compression induces unwanted artefacts, such as a spectral mismatch with respect to the underlying dataset, thus hindering their use as training surrogates. In this article, we propose a spectrally-aligned latent-flow time series generator, where the latent space for flow matching is trained to preserve dynamical properties that are relevant for the suitability of synthetic samples. We find that incorporating fine-tuning losses based on canonical signal representations such as the Fourier, wavelet and signature transforms helps overcome these issues. The interpretability of these transformations allows us to ensure that the synthetic signals are aligned with the true ones in terms of relevant features, such as smoothness or targeted spectral content, as opposed to relying on pointwise reconstruction losses only. We compare the proposed aligned models against a base latent-flow model and the state of the art over real-world long-range univariate and multivariate benchmark datasets. Our quantitative results validate the superiority of the proposed method in terms of its performance on metrics reflecting signal realness and computational efficiency, while being aligned to the training set with respect to its local structure.
We introduce Multiplicatively Optimistic Regret Matching (MORM), an uncoupled learning rule for finite general-sum games. Under simultaneous full-information self-play, every player achieves external regret O(nd) uniformly over all horizons, using only one-step optimism. The analysis combines a potential-based regret-matching argument with multiplicative stability and Hellinger control of strategy movement. A learning-rate safeguard additionally gives O({Td}) regret in the face of adversarial utilities.
Manus, which earlier this year had to break off a merger with Meta, is in discussions to raise 500M at a 4B valuation.
