Large language models are increasingly deployed as autonomous agents for first-principles materials simulation. We report empirical findings from DFT-Agent Bench, a benchmark of 200+ Quantum ESPRESSO tasks spanning relaxation, electronic structure, equation-of-state, and elastic-tensor pipelines over multiple material tiers. Each task carries a QE-derived ground truth with public-reference cross-checks where available, and a provenance audit checks whether the required QE execution contract was met. Under one execution harness, we compare a near-bare ReAct baseline with a bundled v5.1 scaffold on two models. The complete scaffold improves bands+DOS by +0.08/+0.11 and d-electron vc-relax by +0.17 on both models, while high-scoring relaxation and equation-of-state cells leave little room for improvement. The elastic result is model-dependent (0.18 to 0.40 for deepseek-v4-pro and 0.62 to 0.59 for claude-opus-4-8). A fixed 100-task analysis with retrieval disabled supports the aggregate pattern: the fully frozen DeepSeek comparison gains +0.17, while the Opus comparison gains +0.10 against a historical pool baseline (p < 0.001). The provenance gate catches incomplete or unmarked executions that can receive high rubric scores; it is an execution-contract check, not a test for intentional fraud. These results characterize the evaluated DFT workflows. The anonymous review artifact contains the frozen diagnostic slice, prompts, scorer, and solver revisions: https://anonymous.4open.science/r/231e31re91287h/README.md.
When an LLM agent's backend model is upgraded or replaced, operators typically reload the same persona, memory, and rules onto the new model and assume the agent will behave as before. We show this assumption requires empirical validation-and that whether identity is preserved depends on the specific model pairing. Across 5,760 compliance-testing runs in 12 regulated domains with three open-weight models, we find that some transitions produce significant behavioral drift while others do not, even when the agent's external state remains identical. Crucially, this drift is invisible to verbal verification: every model claims full policy compliance, yet actual decisions under pressure differ across models. The same external state can reorder which model is most compliant-a large positive effect on one substrate, negligible on the other two. Practitioners should treat model transitions as behavioral changes requiring action-level empirical testing, not as transparent infrastructure operations.
Neural methods for graph-structured optimization have achieved strong performance within individual problem classes, but are typically trained for fixed formulations, limiting reuse across related problem variants. Recent progress in foundation models has motivated shared neural backbones trained across multiple graph-structured optimization tasks; however, existing approaches primarily share instance-level representations while treating feasibility constraints implicitly. A key source of shared structure across constrained optimization problems is the constraint set itself, which often exhibits substantial overlap across tasks. We propose CONSTRAINER, a constraint-conditioned framework for graph-structured optimization that treats constraints as structured inputs. Constraint specifications are parsed into operator graphs and encoded using a graph neural network to produce invariant constraint embeddings, which condition a shared backbone model for solution prediction. We evaluate our approach on power grid optimization and mixed-integer linear programming problems, where it shows improved solution optimality and enables efficient transfer learning and generalization compared to baseline methods.
Endoscopic diagnosis unfolds over native video, where a single positive finding during traversal of the gastrointestinal tract sets off a multi-minute chain of reasoning over location, morphology, severity and intervention. Existing endoscopy multimodal benchmarks test only static images or single-disease videos, which miss the temporal and clinical reasoning that real-world reporting demands. We introduce EndoSCOP-V, a video-grounded multi-turn evaluation framework for structured endoscopy reporting and clinical reasoning. It contains 1,307 real-world endoscopy and colonoscopy videos (63.2 h; 97% 1080p) averaging about 174 s. Its 20,274 question-answer turns are instantiated deterministically from clinician-filled structured endoscopy-reporting EHRs, so no per-case annotation is synthetic or LLM-rephrased. They come from 1,033 clinician-reviewed templates, about 18 per disease, that span a clinically aligned 9-task x 30-sub-task reporting taxonomy over 54 World Endoscopy Organization (WEO) MST 3.0 diseases. 22% of procedures carry co-existing pathologies, up to six per case. A Markovian context-pruned oracle-forcing protocol evaluates the dialogue turn by turn, injecting the gold answer before each subsequent turn so that per-turn reasoning is isolated from cascading dialogue errors. A panel of 15 board-certified endoscopists sets a 78% strict-accuracy baseline on 1,500 questions. Across 14 open-weight LMMs, the best 27B-class model reaches 59.8% strict accuracy, falling to 53.4% once format-trivial disease-rejection items are excluded. On a between-subjects spot-check, the best of 3 frontier proprietary models trails the clinician panel by 23 pp on strict NDR accuracy. Sub-task analysis reveals consistent gaps in quantitative measurement, multi-variable grading and intervention reasoning on long-horizon video. Data: Hugging Face (EndoSCOP-V). Code: GitHub (abhi25t/EndoSCOP-V).
Activation probes look like near-perfect monitors for secret collusion among large language model (LLM) agents. We show that they mostly detect that an agent was told to collude, not that it decided to. Our benchmark (2,500 committee and 1,476 code-review episodes) adds agents who receive the collusion instruction and refuse it, separating instruction from decision. Published detectors barely tell colluders from these refusers, and our best probe separates them mostly by reading the instruction: without it, agent AUROC falls from 1.00 to 0.61. Two further pitfalls inflate reported results: scored after the vote, a rule that only reads the votes reaches AUROC 0.990 but finds nothing before it; and in 57% of attempts no colluder argues for the target in public. The decision is still there: a probe trained on agents' own choices reads it (AUROC 0.73 without the instruction), and probes find colluders whose goal was fine-tuned into the weights (AUROC 0.995; pre-registered). We turn the controls into a protocol for testing any collusion monitor, and adapt scan statistics into an online monitor for groups of any size.
Long-horizon forecasting of neural population activity tests whether a model has captured dynamics that remain predictive well beyond the observed history. Prediction quality typically declines as the horizon grows, yet neither short-horizon accuracy nor teacher-forced evaluation, which supplies the true activity preceding each predicted step, reveals how well models sustain predictions from an initial history alone. We compare direct multi-step prediction and autoregressive rollout on spiking recordings across horizons and introduce NeuroHorizon, which pairs event-based spike encoding with a causal decoder for population firing rates. Its hierarchical prediction memory keeps recent predictions at full temporal resolution and compresses earlier ones into segment summaries, giving the decoder both local detail and longer-range context. Under teacher-forced training this memory holds ground-truth activity, whereas during rollout it holds the model's own predictions. To reduce this mismatch, we train with scheduled sampling, progressively exposing the decoder to its own predictions. From 500 ms of macaque motor-cortex activity, NeuroHorizon achieves the highest forward-prediction score among the evaluated autonomous models at all four horizons from 100 to 1000 ms, and its per-bin predictions remain above a mean-firing-rate baseline across nearly the full 1 s window. Complementary experiments on macaque sensorimotor and mouse multi-region recordings show that performance generally increases with model and data size, alongside improved prediction after cross-subject pretraining and fine-tuning. Together, these results clarify how forecasting performance changes across horizons and offer an effective approach to long-horizon forecasting of neural population activity.
Medication recommendation must combine patient-specific treatment evidence with the interaction structure of a prescription. The central challenge is to determine which evidence supports each medication and how strongly to regularize interactions while preserving predictive quality. We propose EviDrug, a clinical evidence-guided residual learning framework. Its clinical-memory backbone integrates a self-trained visit-aware encoder, training-case memory and one medication-set interaction block to establish a patient-specific prescription. A gated refinement head combines the backbone states with clinical associations and history-conditioned medication identities through bounded logit corrections. Drug-drug interaction (DDI) feedback adapts regularization strength using training-query predictions. Across MIMIC-IV, NWICU and SICdb, EviDrug leads twelve published-method baselines in AP, F1 and Jaccard, improving AP by 2.26, 2.06 and 1.71 percentage points over the strongest comparator. These gains come at lower parameter and runtime costs than MetaDrug. Matched controls demonstrate DDI reductions with stable predictive quality. Code is available at https://github.com/LIZESHENG13/EviDrug.
Reproducible evaluation, verifiable inference and on-policy reinforcement learning (RL) need an LLM engine to give a request the same output bits whichever requests share its batch. Batch-invariant kernels exist, but engines still select each kernel and its parameters from the batch. Two kernels that compute the same sum can round it differently. The engines' deterministic modes remove the choice by keeping one conservative kernel per operator and pay for it in latency. Many fast kernels are batch-invariant by themselves, so it is enough to fix the few batch-dependent values that kernel selection reads. Detent expresses kernel selection as a program that the engine executes. It finds the fewest values in this program that must be fixed and tunes them for the workload before deployment. On multi-GPU deployments of Llama-3.1-70B and Qwen3-30B-A3B in vLLM and SGLang, Detent is batch-invariant. Tuned per load, it serves at 1.04-1.17x the latency per output token of Stock (each engine with its standard settings), against 1.67-10.3x for the deterministic modes. In SGLang it matches the GSM8K accuracy of SGLang's deterministic mode. Detent covers multiple deployment combinations, including multiple engines, GPU generations, tensor-parallel degrees, and dense and MoE models. Seeded RL rollouts of models up to Llama-3.1-70B replay bit for bit under other batches and cache splits. A trainer on Detent's kernels and values matches every rollout logprob bit for bit at 1.03x Stock's rollout time per token. BF16 GRPO then stays stable where Stock's run collapses.
Large language model (LLM) inference serving is undergoing rapid growth and large-scale deployment, motivating us to rethink how the inference process itself can be leveraged to enable efficient task-specific LLM adaptation. In this paper, we propose a forward-only approach for efficient LLM adaptation with forward-only passes in LLM inference serving. We exploit the angle concentration of activations induced by each singular value decomposition (SVD) component to measure its contribution-dispersed or concentrated-to the angle concentration of the hidden states. Based on this, our forward-only search then efficiently identifies the weight matrix with the highest dispersion that merits rank reduction. We then selectively remove the higher-order components and retain the lower-order components in SVD. Empirical results across diverse datasets demonstrate the competitive accuracy of our forward-only approach, while theoretical analysis shows lower peak memory and greater speedup than the gradient-based approach and greater speedup than the exhaustive search approach. The experiments further present the robustness and generalization of our forward-only approach to various LLMs with up to 57B parameters. The code is included in the supplementary material.
Long video understanding requires models to identify and reason over sparse, temporally distributed evidence within limited frame budgets. Existing keyframe selection methods typically retrieve query-relevant frames from a densely sampled candidate set before inference. However, global image-text matching incurs unnecessary computation when a small set of uniformly sampled frames already provides sufficient evidence. Query relevance alone may also overlook evidence that distinguishes competing answers, while one-shot selection offers no opportunity to correct earlier selection errors. We propose HypoKFS, a training-free, plug-and-play framework that couples keyframe selection with iterative inference by multimodal large language models (MLLMs). HypoKFS first reasons over sparsely sampled frames to construct an initial hypothesis space. The resulting answer distribution determines whether additional frame selection is needed. When needed, the retained hypotheses guide the selection of discriminative frames, followed by renewed inference and hypothesis refinement. This iterative process directs additional visual computation toward unresolved hypotheses while avoiding unnecessary global frame matching. Experiments on Video-MME, LongVideoBench, and MLVU demonstrate consistent improvements across multiple MLLM backbones. HypoKFS improves accuracy by up to 6.7 percentage points and achieves 72.9% accuracy on Video-MME at only 1.3% of the reference frame-hypothesis matching cost.
This paper introduces a chatbot developed to assist in diagnosing common diseases in tomato plants through natural language interaction. Users can describe symptoms in text. Then, the chatbot responds with the most likely disease, and basic treatment suggestions and related image-based reasoning. The system leverages pre-trained Mistral and Llama 3.2 large language models. The system is also built using LangChain and Streamlit to provide a lightweight and offline-friendly interface. This work targets to support tomato growers a fast and accessible tomato plant health guidance when the technical expertise is not available nearby, especially in the remote areas. The experimental results are promising and encourage further improvement of the chatbot for future real-world deployment.
Large language models (LLMs) are increasingly used to process medical records for healthcare tasks such as clinical documentation. Because these tools also generate documentation, the relationship is now bidirectional: models both read from and write to the record. Understanding how LLMs behave when reading records that already contain LLM-authored content is therefore critical for their safe deployment. In this study, we examine how LLM-authored content in the medical record affects downstream LLMs that read the same record. We consider two types of LLM-induced artifacts: bloat, in which LLM-generated summaries accumulate in the record and change the length and provenance of its content over time; and distortion, in which LLMs subtly alter clinical findings rather than fabricating facts outright. As a case of distortion, we study diagnostic overcall (documenting an unconfirmed differential diagnosis in the record) and test whether it propagates to the reader LLM's output. Across seven reader models from four families, we find evidence of both effects. First, LLM-generated summary notes modestly but consistently degraded downstream hospital-course generation compared with copy-forward and templated text of matched length. Second, injected overcalls reappeared in downstream outputs at rates that varied by reader model, without detectable changes in standard quality metrics. These findings show the need to monitor AI-generated content in medical records and to account for it when those records are later processed by AI.
When agents organize records into folders, they may correctly group the excerpts of a design proposal yet place the proposal under the wrong owning group. Existing agent benchmarks do not directly measure such higher-level errors. We introduce InfoOrgBench, a benchmark of 150 tasks in which agents organize scattered records from meeting recaps, issue threads and design proposals into hierarchies. Integrity checks verify that every record is preserved, and a Structure Score measures agreement with the source hierarchy at every level, independent of folder names and order. We evaluate 45 systems that pair nine models with five agent harnesses. The best system scores 83.2 out of 100, but no system exceeds 78 on meeting recaps or issue threads. Model choice affects scores far more than harness choice: model means span about 27 points, whereas the average effect of each harness is below one point. More importantly, parent-level relations have higher error rates than leaf-level relations in all three domains. In design proposals, nearly half of the outputs analyzed by level make exactly this kind of error. These results show that evaluating information organization requires measuring higher-level relations separately from record preservation and local grouping.
Generative AI can expand cultural collections into prompt-responsive visual media, but historical artworks rarely provide the paired descriptions needed to adapt modern image generators. We present The RenAIssance, a scalable pipeline that turns a collection of historical European artworks into a reusable text-to-image media resource. The pipeline generates scene-level captions with BLIP and uses the resulting approximately 15,000 image--text pairs from DEArt to train a compact LoRA adapter for Stable Diffusion v1.5. The adapter shifts generated images toward the historical-art distribution while retaining the base model's ability to render novel concepts. On historical and contemporary prompts of different lengths, it improves FID to the DEArt reference distribution in every setting with only a modest decrease in CLIP text--image similarity. Qualitative results show that the same adapter can render both period-appropriate subjects and contemporary concepts in a coherent historical visual language. By decoupling a reusable style representation from the base generator, the approach offers a practical building block for controllable historical-art media creation, interactive authoring, and rich-media experiences.
Large language models can generate motion graphics from static images and natural language requests, but it remains difficult to evaluate how well models can follow requests that combine spatial, appearance, and temporal requirements. We introduce MotionGraphicsBench, an expert-curated benchmark of 100 instances for animating an input static SVG with a motion edit prompt. By supplying all required visual elements, our editing task reflects a designer-like workflow and isolates motion generation from graphic generation. Derived from real-world Lottie motion designs, our benchmark contains diverse combinations of spatio-temporal animation behaviors. Each benchmark instance pairs its model input with a deterministic, human-readable verification program that accepts different valid implementations and reports which prompt requirements pass or fail, rather than comparing against a reference animation or using a VLM judge. We construct these programs with a multi-agent pipeline using human-labeled animations and synthesized test cases. On a balanced test set of 180 animations, our programs achieve 93.3% raw agreement with human judgments, compared with 83.3% for a GPT-6 Astra VLM baseline, and outperform the baseline in both approval and rejection recall. Across 47 configurations from 32 LLMs and 12 providers, pass@1 ranges from 6.0% to 67.7%. Across 15 models evaluated at their highest available reasoning effort with 32 attempts per instance, additional attempts improve every model, but pass@32 still ranges from 38.0% to 95.0%. Our analysis of verification programs reveals that generated animations often fail to match reference poses, follow requested motion trajectories, or coordinate related events. We release our benchmark at https://huggingface.co/datasets/figma-ai/motion-graphics-bench.
Angle Power Spectrum (APS) and Power Delay Profile (PDP) are receiver-facing spectra that govern beam selection and wideband equalization. Most ray-tracing benchmarks provide supervision as variable-size multipath component lists. This format does not match practical receivers, whose angular and delay resolutions are limited by array aperture and system bandwidth. As a result, learning targets are ambiguous and evaluations can change with reconstruction conventions. We introduce UrbanADPSMap, a benchmark that converts ray-traced supervision into fixed-grid APS and PDP targets by explicitly modeling bandwidth-limited and aperture-limited response kernels. By fixing the receiver-domain kernels, discretization, and validity mask, UrbanADPSMap removes reconstruction-induced target ambiguity and stabilizes downstream evaluation across methods. Built on this benchmark, we propose SpecFlowAR, a coarse-to-fine generator based on multi-scale autoregressive flow matching. The model first captures coarse dispersion structure and then refines beam-relevant peaks and low-power tails. The number of function evaluations controls the sampling budget, enabling latency and quality to be adjusted for dense spectrum queries. Experiments on the test set show that SpecFlowAR reduces BPL@5 to 2.03 dB under a unified NFE=80 budget, while improving beam utility and reducing errors of 3GPP dispersion statistics.
Reliable forecasting of Municipal Solid Waste (MSW) production is pivotal for sustainable urban management, as it enables efficient resource planning, optimized collection schedules, and reduced environmental impacts. Existing works on MSW forecasting have some limitations, such as unavailability of granular data from lower administrative units, small datasets, limited or sometimes no consideration of important factors that can affect MSW generation, including holidays, weekends, seasonality, changes in temperature or humidity. This work presents a comprehensive comparative analysis of four popularly used time series forecasting models, namely Seasonal Autoregressive Integrated Moving Average with Exogenous Regressors (SARIMAX), Prophet, Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU), using a real-world MSW dataset from Sri Lanka spanning 2012-2018. A detailed factor-wise analysis is conducted to examine waste generation behaviour considering holidays, weekday-weekend variations, seasonal patterns, temperature and humidity. The results show lower waste generation during holidays and weekends, higher generation during Summer and Winter, and comparatively lower generation during Monsoon and Post-Monsoon. Temperature and humidity do not exhibit a direct linear relationship with waste generation, although their seasonal variations are associated with the changes in waste generation across different seasons. All model performances are evaluated using Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Squared Error (RMSE), and the coefficient of determination (R2). Among the evaluated models, LSTM and SARIMAX achieve the best overall forecasting performance.
Recent advances in vision-language models (VLMs) enable AI systems to interpret egocentric video. Yet continuous visual capture raises privacy concerns because sensitive content is pervasive and varies across users and contexts. We present EgoVeil, a personalized privacy reasoning system that determines which visual content to obscure based on individual preferences while retaining context useful for downstream assistance. It converts localized visual regions into structured semantic observations and combines a privacy reasoner with an explicit user-specific memory for personalized privacy reasoning. This design supports personalization and continual adaptation as preference evidence accumulates. We evaluate on in-the-wild data from 10 smart-glasses users under a cross-user chronological protocol. EgoVeil achieves 91.93% balanced accuracy and 84.42% Mask recall, outperforming direct VLM baselines and other controlled privacy reasoning baselines. Our results show that personalized, adaptive privacy reasoning better aligns privacy preservation with individual preferences while retaining the visual context needed for effective smart-glasses assistance.
Diffusion large language models (LLMs) have recently emerged as a promising alternative to autoregressive LLMs, enabling parallel generation through iterative denoising. Recent guidance methods improve diffusion LLMs by introducing a weak model to guide the prediction of the base model during denoising. However, the influence of the weak model during denoising remains less clear, and computing its prediction requires an additional forward pass at each denoising step. We present in this paper an in-depth analysis of how diffusion LLMs use contextual information during denoising. Our analysis reveals that diffusion LLMs rely more heavily on local context while often underutilizing global context. Based on this observation, we introduce Global Context Guidance (GCG), a feature-space guidance method that enhances global-context utilization without constructing an explicit weak model. Extensive experiments on LLaDA-8B, Dream-7B, and LLaDA-1.5 across diverse benchmarks demonstrate that GCG consistently outperforms existing guidance methods, while incurring much lower inference overhead.
Vision-language models (VLMs) can recognize events in images and parse sentences, but it remains unclear when they bind visual Agent-Patient structure to linguistic semantic roles across languages whose role cues depend on morphology, word order or construction. We introduce ROLECUE, a controlled diagnostic that pairs 192 validated two-person action images with German, Turkish, Japanese, English, Korean and Bangla sentences in canonical and patient-first orders. The evaluation separates image-only role recognition, text-only role parsing, direct image-sentence matching, response-compliance behavior and an external modular composition of the model's own component predictions. Across eight recent open-weight VLMs, ROLECUE reveals a structured spectrum: some models fail visual or linguistic component tests, some largely solve the diagnostic, some collapse to constant semantic assent and several show component-joint gaps in which reliable image-only and text-only role predictions do not survive direct joint inference. The largest non-degenerate gaps reach 44.9 percentage points, with final bootstrap intervals confirming that the effect is not a point-estimate artifact. These results reveal multilingual component-joint performance gaps and acceptance biases hidden by aggregate VLM benchmarks and show why component diagnostics are necessary before interpreting joint multimodal errors.