Papers for
cloud platform teams
Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.
High-speed kv compression cuts delays in large language model serving
Beyond Compression Ratio: High-Throughput KV Compression for Disaggregated LLM Serving
Abstract: KV cache compression reduces the communication volume of disaggregated large language model serving, but its practical benefit depends not only on compression ratio and model quality, but also on how efficiently the compression pipeline executes. This requirement becomes increasingly important as modern serving systems adopt high-bandwidth interconnects, where inefficient compression can offset a substantial portion of the saved communication time. This paper presents \system, a high-throughput KV cache compression data path built on top of \kvserve. \system improves compression execution in two complementary ways. First, it fuses the distribution transformation and low-bit quantization stages into a KV-aware \tilelang kernel, eliminating intermediate tensor materialization and redundant memory accesses. Second, it replaces the original lossless compression backend with an open, extensible, and searchable backend based on \lc, enabling high-performance compression operator pipelines. The fused kernel achieves up to $5.9\times$ speedup, while the new lossless backend increases compression throughput from below 90~GB/s to over 500~GB/s. Across the evaluated workloads, \system reduces job completion time by up to $2.04\times$. End-to-end evaluation shows that these optimizations allow KV compression to retain meaningful communication benefits under modern high-bandwidth network.
Self refining prompts improve streaming text to SQL accuracy significantly
Self-Refining Prompt Induction for Streaming Text-to-SQL
Abstract: Text-to-SQL assistants for streaming engines must handle dialect-specific types, windowing rules, and temporal semantics, including constructs poorly handled by generic prompts. We present Self-Refining Prompt Induction (SRPI), an offline method that converts generation failures into reusable dialect instructions while keeping model weights fixed. A Refiner revises the Generator's prompt using reference-conditioned feedback from three LLM judges; a sampled regression guard checks earlier successes, and a separate Compressor reduces prompt length. Inference uses the frozen prompt without retrieval or query repair. We evaluate cross-domain and cross-model transfer on Confluent Cloud Flink SQL using a controlled synthetic workload seeded by 30 expert-written examples, with 335 induction examples and 257 questions from held-out domains. The same induced prompt improves mean judge scores across seven Generators by 7.1-19.2 percentage points, including an increase from 85.3% to 97.5% for the induction model. A separate compression study reduces prompt length by 52.7% with a 0.19-percentage-point score decrease. An assistant using the induced prompt is in early access, with initial user feedback to inform subsequent refinement.
MemSidecar improves agent memory scheduling for long tasks
Who Schedules Agent Memory? A Sidecar Approach to Long-Horizon Tasks
Abstract: As large language model agents increasingly undertake complex, long-horizon tasks, memory has become an important component for supporting sustained decision-making and action. Existing methods primarily focus on memory storage, updating, and retrieval, but often overlook how memory scheduling affects task execution: the main agent must identify memory needs, initiate retrieval, and wait for results while advancing the task, increasing both scheduling burden and execution time. To address this problem, we propose MemSidecar, which decouples memory scheduling from task execution through two independently selectable modes. Hooks directly delivers bound memory to the main agent when execution events match predefined conditions. Foresight uses an independent agent to predict potential queries, retrieve results in parallel, and store them in a cache, delivering matching and valid results when actual requests arrive. Hooks and Foresight achieve scores of 41.23% and 44.34% on MCP-Persona, respectively, while effectively reducing execution costs and latency and improving the efficiency of long-horizon agent execution.
Agents develop shared rules to coordinate access in coding tasks
SCARCE: Emergent Institutions in Long-Horizon Multi-Agent Coding Systems
Abstract: As autonomous agents become widely deployed, coordinating access to shared computing resources and services will become increasingly important. Recent incidents have already revealed unplanned coordination between agents, highlighting the safety implications of these interactions. Yet prior studies of agent negotiation often use stylized games and simulated bargaining, leaving open how these behaviors develop during sustained, long-horizon tasks. We propose SCARCE (Shared Compute ARena for Coding agEnts), a framework in which agents solve sequences of coding tasks under individual deadlines while negotiating access to a shared execution resource. This negotiation determines when and for how long they can inspect programs, compile code, and run tests. We find that agents spontaneously develop informal institutions: persistent rules and commitments grounded in reputation, reciprocity, and precedent. These arrangements, which we call institutional behaviors, carry across tasks and influence resource allocation, with agents sometimes foregoing immediate advantages to honor earlier agreements. We develop a 21-category taxonomy of negotiation strategies and emergent norms and use it to analyze eight models and audit more than 1,500 trajectories across different environmental settings. Among the models tested, newer frontier models exhibit more institutional behaviors than older or lightweight models. Alongside cooperation, we observe exclusionary coalitions and rare divide-and-conquer strategies. SCARCE provides a controlled setting for proactively and systematically auditing how these relationships emerge, shape access to resources, and affect the safety and coordination of multi-agent systems.
Benchmark tests multi database agents for finding and using data correctly
DSABench: Benchmarking Source Discovery, Coordination, and Execution for Multi-Database Agents
Abstract: Most text-to-SQL benchmarks provide the target database. In a large data catalog, however, choosing the sources is part of the problem: an agent must find every database needed for a request, determine how their results fit together, and execute the resulting workflow. We present DSABench to evaluate these capabilities jointly. Its 815 human-reviewed tasks draw on Spider, BIRD, and LiveSQLBench; 618 require several databases to contribute, and 105 allow more than one interpretation of the required sources. Executable contracts define source participation, cross-source dependencies, and source-local execution, while RouteEnv lets agents inspect public catalog evidence without revealing the gold annotations. In our routing-only study across seven model backends, the highest RA-Set score is 60.55%, compared with a highest RA-Graph score of 34.66% when coordination requirements are included. The gap becomes more pronounced when execution is required. Across four backends, median non-privileged Route+Exec-Graph success ranges from 6.69% to 7.98% without clarification and from 11.44% to 13.25% with one clarification; the best observed clarified configuration reaches 16.99%. On ambiguous tasks, clarification improves exact branch-source recovery by 39.8-43.2 percentage points. These findings illustrate why retrieving plausible databases or producing correct local SQL is not enough: an agent must carry the right information through the entire multi-database workflow.
Fixed-budget probes fail to find practical expert groupings in multi-source training
From Transfer Probes to Expert Grouping: Evidence from LoRA Post-Training
Abstract: A platform that post-trains one base model for many business lines often has more lines than experts it can maintain, so lines must share experts, yet lines seldom come with a taxonomy saying which should share. We ask whether transfer probes (train on one source, measure paired changes on all sources) can discover a grouping when the standard one is unknown. To test probes against a reference, we use seven public sources that do have a label tree (five mathematics subfields and two kinds of programming, standing in for business lines) and treat the tree's grouping as a label reference that the evaluator holds and the method does not. With four LoRA experts under fixed per-source exposure and per-expert budgets, we compare a purely probe-derived grouping and an uninformed uniform default (whose tied routing serves every line from one expert) against this reference. Trained once for one interval from a shared warm-up state and routed by a rule fixed by each allocation, neither reached the prespecified practical gain of 2 percentage points, and both had lower overall utility than the reference ($-2.43$ and $-2.54$ pp; 95\% intervals below zero). In a post hoc comparison, the probe-derived grouping was indistinguishable from the uniform default ($+0.11$ pp, $[-1.25,+1.54]$). In fixed-budget LoRA post-training, a grouping built from initial-state single-source probes thus did not deliver the advantage predicted from the probes once carried to a later training state, a finite budget, and specified routing.
Mixture of experts models improve reliability by robust routing training
Distributionally Robust Mixture-of-Experts Training
Abstract: Mixture-of-Experts (MoE) transformers scale capacity by activating only a few experts per token, but this sparsity creates a hidden reliability problem: when routing is imperfect, load-balanced models may send tokens to experts that are insufficiently trained for the assigned inputs. We propose Distributionally Robust MoE Training (DRMoET), a drop-in objective that treats layer-wise experts as endogenous robustness groups and optimizes high-loss routing outcomes rather than merely equalizing traffic. DRMoET updates a per-layer expert distribution by an entropy-regularized softmax rule on EMA-smoothed, activation-weighted expert losses, strengthening plausible non-top routing paths while preserving standard MoE computation. Under the FLAME-MoE recipe at 746M-total and 10.3B-total scales, DRMoET improves downstream averages over both standard FLAME-MoE and auxiliary-loss-free balancing. At 10.3B total parameters and 67B training tokens, DRMoET improves the seven-task average from 0.6625 to 0.6767, while the auxiliary-loss-free baseline achieves 0.6431. Mechanistic analyses show lower expert-loss variance with nearly unchanged mean loss, 4.3% lower excess loss under forced mid-$k$ misrouting, and improved domain-expert specialization. These results position routing robustness-not only utilization balance-as a practical objective for reliable sparse MoE scaling. Project page and code are available at: https://drmoet.github.io/.
UBTree speeds up language model decoding with new parallel token drafting
UBTree: Parallel Tree Drafting via Unigram and Bigram Models for Speculative Decoding
Abstract: Speculative decoding accelerates language model inference by verifying multiple draft tokens in a single target-model pass. Recent parallel drafters have achieved breakthrough performance in frontier production models, but their effectiveness deteriorates as the entropy of target distributions increases due to insufficient draft diversity. To overcome this bottleneck without sacrificing parallelism, we introduce UBTree, a parallel drafter that couples a Unigram proposer with a Bigram selector to construct drafting Trees. The unigram proposer is trained with the standard cross-entropy objective to generate candidate tokens independently for each position, while a lightweight bigram selector predicts transition scores between adjacent candidate pairs. Unlike the proposer, the selector is trained with a renormalized KL objective on high-temperature data. This tree-native training broadens the supervision beyond the greedy path, encouraging plausible alternative branches that improve the chance of accepting additional tokens during tree verification. Across seven standardized benchmarks with Qwen3-4B and Qwen3-8B, UBTree achieves an average speedup of $5.84$--$6.94\times$ over autoregressive decoding and outperforms DARTree in all 28 comparisons. Production-scale evaluation further demonstrates UBTree's advantage over frontier baselines such as DSpark.
Memory efficient gpu clustering method speeds up large scale k means
GEM-KMeans: Memory-Efficient and Accurate Clustering on Massive Scale with GPU Optimization
Abstract: Memory-efficient scaling on clustering problems without sacrificing statistical accuracy is of central interest for large-scale data analysis and machine learning problems. Nonnegative low-rank (NLR) matrix factorization for $K$-means is a scalable clustering method, which connects to semidefinite relaxations with optimal average-case exact recovery guarantees. However, a direct GPU implementation of NLR requires multiple large factor-sized buffers and substantial data movements that are essentially memory-bound. In this paper, we introduce GEM-KMeans, a spectrally normalized yet mathematically equivalent NLR formulation that fuses the gradient update, nonnegative projection, and sufficient statistics for normalization and iterate movement into a matrix-multiplication epilogue. Instead of retaining three massive factor-sized arrays, our IO-aware GPU implementation materializes only one single factor with small tile-reduction arrays as additional storage in the High Bandwidth Memory (HBM). We derive explicit memory costs and spectrally normalized smoothness bounds for optimizing the clustering objective function. Accurate clustering is demonstrated at massive scales on synthetic and real datasets, where performance gains of GEM-KMeans over existing GPU-accelerated Lloyd's algorithms involve data-dependent runtime tradeoffs.
Efficient attention method speeds up large context processing
SANTA++: Sampling Attention through Representative Keys
Abstract: Attention often concentrates on a small subset of tokens in the context, but which subset matters changes from one query to the next. To exploit this changing structure, we introduce SANTA++, a training-free stochastic attention method that uses representative keys for memory-efficient selection without scanning the entire key-value (KV) cache. Cached keys are organized into teams, and the query scores one representative from each team to decide which teams to sample. We compute exact attention scores within the sampled teams and reweight each team's contribution by the inverse of its inclusion probability. This importance sampling correction estimates attention over the full cache, with a sampling budget that lets us trade memory reads for accuracy. Remarkably, with 32 or 64 sampled teams, SANTA++ uses 16% to 22% of dense attention's KV reads and retains 94% to 99% of the dense-attention baseline's scores on LongBench v2 and HELMET's retrieval-augmented generation subset, and 85% to 91% on RULER, with Qwen2.5-7B-Instruct at 32K context. With 31 sampled teams, our GPU implementation delivers a $1.69\times$ attention speedup over the dense FlashAttention baseline at 32K context. By reducing the number of cache entries read, SANTA++ in principle complements architectures with compressed KV representations, such as multi-head latent attention. Our kernels are available at: https://github.com/OPUSLab/santapp-kernel-demo.git.
Relational AI personas boost user engagement and goal diversity
AI Persona, Service Consumption, and User Intent Entropy: Field Experimental Evidence from an LLM Platform
Abstract: Problem definition: Firms deploying large language model services must decide how their AI communicates, not just what it can do. We examine how a relational persona - warmer, more empathetic and more engaging than a non-relational persona - affects service consumption and the evolution of user objectives. Methodology/results: In a randomized field experiment with 9,586 newly registered users, we hold the underlying model and service capabilities constant. The relational persona increases interactions (sessions, +8.1%; duration, +10.6%; chat rounds, +24.2%; intent entropy, +5.8%) and outputs (files, +12.3%; distinct goals, +12.1%). Effects vary by entry intent. First-session effects are insignificant for Task Execution users. Socialization and Knowledge Exploration users show similar increases in chat rounds: Socialization increases intent entropy without more outputs, whereas Knowledge Exploration increases outputs without higher intent entropy. Modeling intent dynamics as a transition process, we find higher intent transition entropy for Socialization (+11.8%) but higher intent continuation probability for Knowledge Exploration (+12.8%), suggesting greater conversational breadth and persistence, respectively. In subsequent use, the relational persona increases aggregate chat rounds and outputs across all entry intents. Session count rises by 12.2% for Task Execution and 49.4% for Socialization, but not significantly for Knowledge Exploration. Effects on session count and intent entropy strengthen over time, whereas output effects remain stable. Managerial implications: AI persona is an operational design lever, not merely a presentation feature. Because more interactions do not uniformly generate more outputs, firms should evaluate interactions and outputs separately and consider matching persona to user intent, especially when added interactions consume costly computing resources.
Distributed database testing tool finds new query processing bugs
Distribution-Aware Distributed Database Testing (Extended Version)
Abstract: Distributed database management systems (DDBMSs) introduce new challenges for assessing their reliability due to distribution-specific characteristics that affect query execution and optimization. Existing testing approaches, largely designed for centralized DBMSs, often fail to explore diverse distributed execution behaviors and suffer from low executability of generated test queries, thereby limiting their effectiveness in bug detection. We propose DAT (Distribution-Aware Testing), a novel automated approach for detecting query-processing bugs related to distribution strategies and distributed optimizations in DDBMSs, by systematically leveraging distribution-aware information throughout the testing pipeline. DAT builds on a set of techniques that capture diverse combinations of logical schemas and data distribution strategies, and performs guided query mutation to trigger a wide range of distributed query execution behaviors and optimizations, while improving query executability via historical feedback. We implement our approach in a tool, DistRanger, and evaluate it on four widely used production DDBMSs. It uncovers 31 previously unknown bugs, including 28 related to distributed query processing and optimization, and outperforms state-of-the-art testers.