Advances in large language models and semantic IDs (SIDs) have brought generative retrieval (GR) into industrial applications. These systems must return items that satisfy request metadata constraints as well as relevance requirements. A catalog trie ensures SID validity, but filtering metadata only after generation wastes computation on rejected candidates and may leave too few eligible results. We propose RoutingTrie to incorporate metadata constraints before beam selection. RoutingTrie maintains one shared SID trie and builds a compact index for each metadata field using Dense Bitset or Sparse List. We derive a storage cost formula to guide the choice between these index representations for different field distributions. During decoding, field probes determine whether requested values occur below candidate edges and produce a combined mask that guides beam selection. Experiments on ShopTrie, our configurable synthetic benchmark, demonstrate compact storage and efficient CPU and GPU probing across diverse metadata distributions. In a production replay, RoutingTrie raises Hit@100 from 67.5\% to 96.1\%, a gain of 28.6 percentage points, with only 1.1 ms latency increase. We will release the ShopTrie generator, benchmark scripts, and RoutingTrie CPU and Triton GPU implementations to support reproducibility.
Click-through rate (CTR) prediction is a core component of recommendation and advertising systems. Large language models (LLMs) offer rich semantic priors that improve prediction quality, particularly in cold-start and long-tail settings. Injecting a full-scale LLM into a CTR pipeline is, however, prohibitively expensive at inference time: existing LLM-enhanced CTR methods avoid this cost by caching LLM outputs offline rather than serving the LLM online. We propose a resource-efficient pipeline, SlimFM, that makes online LLM-based semantic modeling practical: (1) dependency-aware structured pruning compresses LLaMA3-8B by 44% to a 4.5B-parameter backbone; (2) LoRA-based field-aware fine-tuning trains the pruned model to produce discriminative field-level semantic embeddings from lightweight textual prompts; (3) a semantic-guided fusion module injects these embeddings into the feature interactions of a conventional CTR model. Across four datasets and six CTR backbones, SlimFM retains over 99% of uncompressed-model accuracy while reducing inference latency by up to 3.4x relative to existing LLM-enhanced baselines.
Animals integrate information across sensory modalities to guide exploration, suggesting that they maintain internal estimates of where actions are expected to reduce uncertainty. However, how such action-conditioned information maps are learned from multimodal sensory experience remains unclear, as previous studies have been largely confined to unimodal sensing frameworks. We introduce a multimodal predictive-coding framework for learning uncertainty-reduction maps from visual and mechanosensory observations in hawkmoth flower interaction tasks. A variational generative model trained on simulated environments recovers modality-specific information landscapes over the action space, capturing how visual patterns, surface geometry, and mechanosensory cues shape the value of exploratory actions. These learned maps are consistent with several features of hawkmoth behavior, including nonzero tracking offsets, probing along visual patterns, and geometry-dependent exploration. Applying the model to behavioral data from flower tracking and nectary search, we find that natural trajectories tend to occupy regions of high uncertainty reduction in these learned maps. Building on these observations, we develop an uncertainty-gated lexicographic algorithm that switches between information gathering and reward exploitation based on perceptual uncertainty. In nectary search, uncertainty gating improves learning relative to exploit-only, $\epsilon$-greedy, fixed-time, and uncertainty-calibrated non-lexicographic controls, with better generalization and robustness to sensory and curvature shifts. These results link information landscapes learned through generative modeling to biological sensing and embodied exploration.
Graph-based retrieval-augmented generation (GraphRAG) improves multi-hop question answering, but deployed systems face two costs that research comparisons rarely measure: building the graph, and removing documents from it. We present Lightfoot, a GraphRAG system in which every entity, link, fact and synonym edge is a deterministic function of per-passage extractions carrying chunk-level provenance. Indexing makes one language-model call per passage, retrieval makes none, and deleting a document is a database operation with no model calls. On the full 11,402-passage MuSiQue corpus we verify that, after deleting 10% of the documents, the graph is identical to a fresh index of the surviving documents. We compare Lightfoot with HippoRAG 2, LightRAG and fast-graphrag under identical extraction, answer and embedding models on MuSiQue, 2WikiMultiHopQA and HotpotQA. Lightfoot builds the cheapest index on all three datasets, 26-31% below HippoRAG 2 and 3-6x below LightRAG and fast-graphrag. Against HippoRAG 2 it is significantly better on MuSiQue retrieval (+2.8 points recall@5 and +4.3 recall@10 on a held-out split) and on HotpotQA answers (+7.7 points exact match), and significantly worse on 2WikiMultiHopQA (-3.9 recall@2, -8.4 F1). Deleting the same documents took LightRAG 2,187 model calls and 3.7 hours; fast-graphrag has no deletion path. Code and per-question results: https://github.com/udayallu-ml/Lightfoot-GraphRAG
Runtime anomaly detectors for cyber-physical systems are either learned as black boxes from normal data, which miss stealthy manipulations that keep every signal inside its normal range, or derived by hand from physical models, which requires expert effort for every plant. We introduce physicsinformed agents (PIA), in which a large language model agent writes the detector itself as a set of physical residuals, such as pump characteristics, tank mass balances, and controller logic, and refines every candidate against normal operating data through an evaluation workbench before deployment. The result is a compact, human-readable detector that runs in milliseconds and names the component behind each alarm. On the BATADAL water distribution attack benchmark, PIA with Claude Opus 5.5 reaches an F1 of 0.878 on the test attacks against 0.504 for the strongest of six statistical and deep baselines, keeps detecting attacks concealed by replayed and offset sensor readings, and attributes every detected test attack to the attacked component. Ablations show that testing hypotheses against data is decisive: agents that write residuals from the plant description alone lose up to 0.28 F1, whereas agents that see only anonymized signals recover most of the physical structure.
Plant organ segmentation on compact computing platforms requires accurate predictions and affordable geometric preparation. We develop and evaluate a multiscale edge based graph attention model that combines local shape descriptors, two neighborhood scales and a residual stem and leaf correction in approximately 78 thousand parameters. The engineering application is plant point cloud segmentation for precision agriculture, assessed through grouped accuracy experiments and physical Raspberry Pi 5 inference. Broad comparisons with hierarchical, graph and transformer baselines establish PointNeXt as the strongest external accuracy reference among the tested local recipes. A nested four variant ablation on a local maize collection separates geometric information from the combined multiscale configuration across seven outer groups and three seeds. The complete model achieves the highest mean intersection over union, 86.36%, although uncertainty and boundary errors qualify the contribution of each component. We then optimize full cloud principal component analysis, sampling and neighbor queries without changing learned weights or descriptor definitions. All 1,180 paired desktop comparisons on PlantNet and TomatoWUR preserve exact logits and predictions. Multiscale pipeline latency falls from 330.33 to 96.82 milliseconds on PlantNet and from 1,054.52 to 463.10 milliseconds on TomatoWUR. PointNeXt also benefits from optimized sampling and remains faster over the complete pipeline. These results support compact graph learning with reduced geometric overhead, while physical board measurements establish inference feasibility rather than energy efficiency or readiness for autonomous flight.
Standard training recipes make a fixed number of passes over the training data, even when an epoch mostly adds mini-batch noise. We study whether a large language model (LLM) that reads the weight and gradient history of a run can replace some of these epochs with updates that compute no gradient and read no training image. In GradReason, an LLM agent is called every ten epochs. In a restricted sandbox, it analyses recent weight snapshots, epoch-mean gradients, and validation metrics, then submits a Python program that computes new weights and a plan marking each of the next ten epochs as data or virtual. A virtual epoch is kept only if it passes a forward-only check on 5,000 held-out validation images. In 120 runs on CIFAR-10 and CIFAR-100 with three architectures, Claude Opus 5.5 used 32.8% fewer data epochs than the 80-epoch recipe at 0.31 percentage points lower test accuracy (95% CI -0.72 to +0.09; 0.94 points on ResNet-18), on the recipe's own frontier of data against accuracy. Unprompted, Opus averaged recent weight snapshots in every run. This choice produced the savings: a fixed rule with the same acceptance check but extrapolations from prior work saved 5.4% of the data epochs, and the same rule with averaging added saved 25.0%. The LLM's contribution is to find a working update within a single run; once found, a fixed rule implements it equally well. An Opus run cost 0.68 USD, but at this scale the agent's latency exceeded the GPU time it saved.
Learned representations (or embeddings) are the foundation of modern machine learning systems, yet they are typically trained as fixed-size embeddings, without accounting for varying downstream resource or task constraints. Matryoshka Representation Learning (MRL) addresses this limitation by learning nested representations, but requires costly end-to-end retraining and does not support post hoc expansion, while Contrastive Sparse Representations (CSR) reduce training overhead via lightweight adaptors but operate at a 4x larger dimensionality. Additionally, all prior works restrict their framework towards compression only. In this work, we introduce Hyperbolic Elastic Representation Learning (HERL), a framework for learning dimension-adaptive embeddings (compression and expansion) through a lightweight adaptor without retraining the base encoder. Our key idea is to exploit the exponential capacity of hyperbolic geometry by learning a non-linear mapping from the encoder space to hyperbolic space. We train a simple MLP to downsample the representations before projecting them onto a Lorentzian and learn adaptive representations that respect the Lorentzian geometry. Our approach is model-agnostic and effective in both supervised and unsupervised settings. Through extensive experiments across image and text modalities, we show that HERL preserves downstream performance across a wide range of embedding sizes, exhibits smooth interpolation between dimensions, and achieves these benefits at a fraction of the computational cost. Code: https://github.com/arjunrkaushik/herl.
Global supply chains rely on the interactions of thousands of heterogeneous agents, posing significant challenges for optimization. While traditional optimization often assumes homogeneity to ensure tractability, Reinforcement Learning struggles with sample inefficiency when scaling to high-dimensional state spaces. We propose a Differentiable Agent-Based Model for multi-echelon inventory management. By implementing stochastic dynamics via differentiable tensor operations, we enable direct gradient propagation through the simulation. We validate the framework by recovering theoretical Newsvendor optima without supervision. Crucially, we demonstrate ``Mass Personalization'' by optimizing 10,000 distinct agent policies simultaneously in seconds, achieving a 53% cost reduction compared to scalar baselines. Furthermore, we utilize gradient sensitivity analysis to generate causal heatmaps, accurately recovering inherent lead-time delays from the computation graph. This approach bridges the gap between simulation and optimization, enabling the efficient training of self-correcting Digital Twins.
General-purpose chatbots are increasingly used for emotional and mental health support. They can be useful, but issues such as sycophancy, reinforcement of delusional beliefs, and poor crisis response create risks that are difficult for researchers to study and evaluate. In this paper, we review the most critical open problems for researchers and policymakers in the use of general-purpose chatbots and their effects on mental health. First, there are tractable problems in the controlled evaluation of chatbots: establishing the construct and ecological validity of evaluations, evaluating multi-turn interactions, and connecting evaluations to improvements in chatbot behavior. For example, conventional benchmarks that primarily rely on static datasets cannot establish chatbot safety, and novel simulation methods need validation. Second, researchers must connect their studies of chatbots and mental health to real-world evidence. This includes analyzing the chatbot products that AI companies and mental health practitioners have deployed, including their user interfaces and human review and escalation practices for user welfare concerns. However, clinical analysis opportunities afforded by real chat log data sharing are complicated by significant privacy concerns. Finally, we argue that researchers should take an active role in the construction and refinement of evidence-based policy through corporate and governmental decisions such as user protections, standards of care, protected access to user chat logs, data sharing between institutions, incident reporting, and post-deployment monitoring. Such behavioral evaluations, studies of real-world outcomes, and data governance depend on each other and must be addressed in tandem. We believe a coordinated research effort would have substantial real-world impact.
Machine learning models, particularly large foundation models, are increasingly evaluated along multiple axes that are relevant for their usage. In addition to accuracy measures appropriate for the specific task of interest, measures capturing safety, calibration and fairness properties are often of interest. The classical approach to these problems assumes that the test set has been held out from the training process, an assumption that is increasingly harder to justify for large foundation models. In this work, we formalize a notion of evaluability without this assumption, and ask if it is feasible to evaluate black-box models in settings where our test samples may have been used during training. We study this problem under different settings, capturing different levels of access to the model. We first study the natural setting, where the evaluator only looks at the output of the model on the test set. We fully characterize the class of distributions where accurate evaluation is possible, and show tight upper and lower bounds on the sample complexity of evaluation. Our results show that black-box evaluation in this setup is feasible if and only if the distribution is close to being small support. We then consider evaluation algorithms that can query the model on additional inputs and demonstrate connections to self-correctors studied in program testing. Using these, we show that a small amount of query access can increase the power of the evaluator for natural distributions and concept classes.
In offline reinforcement learning, diffusion models fill two roles with complementary strengths. Typically guided by a critic, a diffusion policy is value-aware and favors high-return, behavior-supported actions, but it is short-sighted: it never lays out the states it intends to visit. A diffusion planner is far-sighted, synthesizing such a long-horizon trajectory, but it is only behavior-driven at execution, realizing each planned transition using only offline behavior.Each is strong exactly where the other is weak. We introduce DPDP to bring them together at the planner's execution step, where a planned transition becomes an action. Reading the planner's inverse-dynamics executor as a posterior over actions: the environment's transition likelihood relates an action to the planned next state, while the behavioral prior determines its weight under the offline data. Replacing this prior with a value-optimized diffusion policy preserves the planned transition likelihood and favors valuable actions that can serve the plan. We target the resulting clean-action distribution in two ways. A training-free sampler composes policy and inverse-dynamics scores at matched action noise levels and applies Langevin correction; this route uses a shared action noise schedule and an approximately common terminal prior. A preference objective instead amortizes the transfer into a single diffusion executor. Across D4RL, DPDP improves the planner it builds on, staying competitive with reactive diffusion policies on locomotion and leading on long-horizon navigation.
Despite world-class snow monitoring by the Swedish Meteorological and Hydrological Institute (SMHI) and Finnish Meteorological Institute (FMI), no station-anchored, multi-parameter snow benchmark dataset exists for the Nordic boreal domain. We present Hydro-Dataset v1.1, a reproducible open-source pipeline that harmonizes 582 SMHI meteorological observation stations across nine hydrological winters (2015/16-2023/24) into a CF-1.9 NetCDF product with 1.8~million annotated station-days. Six derived snow parameters, such as Snow Water Equivalent (SWE), snow phenology, melt rates, degree-day melt factor, rain-on-snow flags, and freeze-thaw cycles, supply benchmark-ready labels for hydrological foundation model training. ERA5-Land cross-validation (Pearson R=0.78, bias=-262mm) confirms observation-anchored SWE provides a more reliable benchmark than reanalysis for machine learning training.
Small hand-tracking networks are deployed as 4- and 8-bit codes, and they should keep adapting to a user's camera after deployment, on the device and within roughly the memory of the model itself. Quantization-aware training cannot do this cheaply, because it keeps a 32-bit master copy of every weight together with its optimizer state, about 96 bits per weight. Training the codes directly with stochastic rounding needs no copy but injects rounding noise at every step, and the methods that remove this noise keep a residual buffer for every weight. We introduce sensitivity-gated accumulator caching (SGAC), which keeps residual state for only a small and changing set of weight groups. A set-associative cache of 8-bit accumulators collects the sub-step updates of the groups with the largest Fisher-weighted squared code step and writes them into the codes as whole steps; an evicted line is flushed by stochastic rounding, and a group without a line does not move until it is admitted. We prove that the write-back is unbiased and show, under a local quadratic model, that rounding a sub-step update costs loss of the same order as the update's own gain, so that rounding pays only where the expected gradient is large relative to the curvature. On a 383k-parameter hand-keypoint network, SGAC trains with 8.1 bits of state per weight. Scored on the 4/8-bit network that the device runs after adaptation, it has the highest mean accuracy of nine deployed low-bit networks on two simulated cameras. On twelve ImageNet-C corruptions it has the highest average of corrupted and clean accuracy of seven deployed methods, ahead of the PEQA-style, ECO-style and direct low-bit baselines with confidence intervals that exclude zero, and it matches an 8-bit accumulator kept for every weight with 56% of that state.
Learning accurate off-policy state-action value functions (Q-functions) from offline datasets becomes harder as the task horizon grows. One-step temporal-difference (TD) learning bootstraps after every transition, so small per-step errors accumulate along the horizon, and early states can be badly misvalued even when the TD loss is small. Multi-step TD shortens this chain by following the dataset trajectory for several steps; however, since the behavior policy is often suboptimal, every followed action adds a bias, and no fixed number of steps suits a whole trajectory. In this paper, we propose Horizon-Scalable Q-Learning (HQL), a simple method that enables accurate long-horizon value learning by automatically balancing between one-step and multi-step returns. HQL extends a target through a dataset action only when the critic values it at least as highly as the best-of-N actions sampled from the BC policy, and otherwise falls back to a one-step target. We show that this rule is a principled off-policy correction that needs no action densities, which makes HQL compatible with expressive flow and diffusion policies. Experiments show that HQL avoids the failures of both one-step and multi-step TD on a combination-lock task across horizons, and outperforms prior horizon-reduction methods on long-horizon OGBench and OCBench tasks.
Evolution can favour a trait for two reasons: an instructive cause makes selection favour the trait; a permissive one lets it persist once selection no longer removes it. The distinction matters most for phenotypic plasticity and its replacement by inherited behaviour, genetic assimilation, the Baldwin effect. Social learning is credited with favouring plasticity on evidence that cannot tell the two apart: how much plasticity remains at the end of a simulation. Here we show, analytically and computationally, which role it plays in a minimal model of learning and evolution. Writing expected fitness as a function of plasticity, we find that environmental change is the only source of benefit from plasticity, and the social channel enters only its cost. Simulation confirms this: in a static environment selection against plasticity never reverses sign at any channel strength; when the environment changes, it reverses at a threshold the channel lowers, and without the channel the population collapses before reaching it. Social learning is permissive, not instructive: it supplies the means to be plastic, never the reason. For the debate over whether plasticity leads evolution, this is one case settled by measurement: environmental change directs, socially inherited information permits, and neither suffices alone.
Collective Communication Libraries (CCL) are critical for large-scale distributed AI/ML training, yet their configurations are determined by static heuristics that do not account for real-time network conditions. Our empirical study shows that these heuristics perform near-optimally in contention-free environments, but degrade substantially under background traffic. To address this, we present a cooperative multi-agent run-time optimizer for accelerator collective communication. Our approach employs a two-agent architecture: (i) detector- a Long Short Term Memory(LSTM) recurring neural network fed classifier head with a sequential probability ratio test that distinguishes sustained contention from transient spikes; and (ii) selector- a configuration selector built on a Dueling Double DQN (D3QN) with Conservative Q-Learning that adapts configuration parameters in response. We implemented this dynamic tuner over NCCL, a widely used CCL developed by NVIDIA. When evaluated on Chameleon Cloud's isolated bare-metal cluster, this dynamic tuning approach reduces AllReduce latency by 36% over NCCL and 32% over a fully converged AutoCCL. We further evaluated it across six real AI workloads spanning convolutional networks, transformers, and recommendation architectures. Our multi-agent tuner approach delivers up to 65% speedups on Delta cluster with NVIDIA A100 GPUs and Slingshot-11 fabric internode connectivity.
Agents built on reasoning models resend their growing transcript at every call. Two transacting parties, not the model, decide what that costs: the client keeps or edits messages; the server decides whether earlier reasoning returns in the prompt and what its cache reuses. We define a cost model whose counts match an instrumented engine exactly; the architecture charges $c$ per token processed and $\beta$ per pair attended. Their ratio $P^*=c/\beta$ is an exchange rate between thinking and remembering: per token, processing dominates below this context length, a thinking regime, and attention above it, a memory regime. A policy change has a mechanical cost, incurred even if the agent writes the same tokens, and a behavioral one: it writes different tokens, changing cost and outcome. Serving alone can flip which policy is cheaper: on recorded Qwen3-8B ScienceWorld runs, dropping earlier reasoning costs 1.74 times keeping it when the server reuses generated states and 0.34 times when it reuses nothing. In live reruns, dropping it cuts correct HotpotQA answers by 9-23 percentage points on four models, at lower point-estimate cost. Comparing context policies means comparing systems: architecture and serving set the cost of the history kept and the work it induces.
We address the problem of computing the optimal Q-function in Markov decision problems with infinite state-space. We analyze the convergence properties of several variations of Q-learning when combined with function approximation, extending the analysis of TD-learning in Tsitsiklis and Van Roy (1996) to stochastic control settings. We identify conditions under which such approximate methods converge with probability 1. We conclude with a brief discussion on the general applicability of our results and compare them with several related works.
Reliable function calling requires Large Language Models (LLMs) to select appropriate functions and generate arguments consistent with user intent. However, syntactically valid calls may still be semantically incorrect due to inappropriate function selection or mismatched arguments. In this paper, we propose Decision-Model-Guided Function Calling (DM-Call), a framework that combines local schema validation with separate decision-model verification of function selection and argument correctness. For rejected candidates, DM-Call accumulates verification probabilities, binary judgments, and schema errors across attempts. A hybrid feedback representation further combines numerical probabilities with verbal correctness labels to guide iterative refinement without executing external tools. Extensive evaluation across BFCL and API-Bank (first-call) demonstrates DM-Call boosts function-calling accuracy versus direct generation, at modest additional cost.