Papersby GlyphSalesforce research papers
A company publishing mostly on agents and evaluation, best known here for BLIP-2.34 papers · 21,614 citations · latest Oct 2026
Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior ForesightSelf-Retrospection Distillation improves agent performance by using hindsight from completed trajectories to supervise foresight predictions during training.
Salesforce1 citations197 upvotesRead →Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM AgentsKeeps an agent's raw past runs and writes a task-specific memory only when a new task arrives, rather than deciding up front what to keep.Salesforce2 citations56 upvotesRead →Flattening Every Memory Peak in Long-Context Mixture-of-Experts TrainingBounds the four memory peaks that crash long-context mixture-of-experts training, from expert dispatch to optimizer state, with fixed-size GPU schedules.Salesforce19 upvotesRead →RISE: Recursive Improvement via Self-Extrapolating Policy DistillationRISE improves language model post-training by recursively generating dense token-level supervision from the model's own reinforcement learning trajectory via self-extrapolation, avoiding external teachers.Salesforce5 citations17 upvotesRead →EVOHARNESSBENCH: Can Your Agents Keep Pace with an Evolving Harness?The study introduces a benchmark to evaluate LLM agents under evolving tool, skill, and agent harnesses, revealing persistent gaps in retention, adaptation, and harness-induced forgetting.Salesforce7 citations32 upvotesRead →Random Attention: Rethinking KV Cache Eviction for Efficient ReasoningRandom eviction of reasoning tokens matches selective KV cache compression because reasoning traces are self-protecting through redundancy, making scoring unnecessary once prompts are preserved.Salesforce5 citations187 upvotesRead →DarwinX: Evolving Agent Harnesses Through Natural SelectionDarwinX evolves agent harnesses via population selection with frozen models, improving verified performance across benchmarks without benchmark-specific patches.Salesforce16 citations116 upvotesRead →StateAct: Program State, before Pixels, for Long-Horizon Computer-Use AgentsStateAct improves computer-use agents by grounding actions, verification, and memory in direct program state rather than screenshots, boosting success rates while reducing cost.Salesforce6 citations62 upvotesRead →Evidence-Backed Video Question AnsweringEvidence-Backed Video Question Answering requires models to provide answers with precise spatio-temporal segmentation evidence, revealing gaps between reasoning and visual grounding that are improved by large-scale instruction tuning.Salesforce2 citations4 upvotesRead →Learning from Language Feedback via Variational Policy DistillationVariational Policy Distillation enables reinforcement learning from language feedback by co-evolving teacher and student policies through variational expectation-maximization, overcoming limitations of passive distillation in complex reasoning tasks.Salesforce7 citations12 upvotesRead →The Illusion of Certainty: Decoupling Capability and Calibration in On-Policy DistillationOn-policy distillation suffers from miscalibration due to information mismatch between training and deployment contexts, which is addressed through a calibration-aware framework that improves both performance and confidence reliability.Salesforce10 citations14 upvotesRead →Beyond the Assistant Turn: User Turn Generation as a Probe of Interaction Awareness in Language ModelsUser-turn generation serves as a probe to measure interaction awareness in large language models, revealing that this capability is distinct from task accuracy and can be influenced by training methods.Salesforce5 upvotesRead →GPA: Learning GUI Process Automation from DemonstrationsGUI Process Automation (GPA) offers robust, deterministic, and privacy-preserving vision-based robotic process automation with faster execution than current vision-language model approaches.Salesforce2 citations11 upvotesRead →Least-Loaded Expert Parallelism: Load Balancing An Imbalanced Mixture-of-ExpertsImbalanced expert routing in Mixture-of-Experts models leads to computational inefficiencies in expert parallelism, which are addressed by a dynamic rerouting algorithm that balances workload and reduces memory usage.Salesforce11 citations5 upvotesRead →Agentic Confidence CalibrationAgentic confidence calibration addresses limitations of static calibration methods by introducing a trajectory-based diagnostic framework that improves reliability across diverse AI agent systems.Salesforce19 citations6 upvotesRead →Agentic Uncertainty QuantificationA unified dual-process framework transforms verbalized uncertainty into active control signals for improved reasoning reliability in AI agents.Salesforce20 citations9 upvotesRead →Future Optical Flow Prediction Improves Robot Control & Video GenerationA novel language-conditioned optical flow forecasting model combines Vision-Language Model and Diffusion architecture to predict future motion from noisy web-scale video data, demonstrating versatility in robotic manipulation and video generation tasks.Salesforce10 citations19 upvotesRead →LoCoBench-Agent: An Interactive Benchmark for LLM Agents in Long-Context Software EngineeringLoCoBench-Agent evaluates large language models as autonomous software development agents using interactive scenarios, specialized tools, and multi-turn conversations to assess their long-context performance, comprehension, and efficiency.Salesforce20 citations3 upvotesRead →MMPersuade: A Dataset and Evaluation Framework for Multimodal PersuasionMMPersuade is a framework for studying multimodal persuasion in Large Vision-Language Models, revealing insights into their susceptibility and the effectiveness of various persuasive strategies across different contexts.Salesforce1 citations8 upvotesRead →Enterprise Deep Research: Steerable Multi-Agent Deep Research for Enterprise AnalyticsEnterprise Deep Research (EDR) is a multi-agent system that automates report generation and real-time data analysis by integrating specialized agents and tools, outperforming existing agentic systems on open benchmarks.Salesforce21 citations11 upvotesRead →Foundational Automatic Evaluators: Scaling Multi-Task Generative Evaluator Training for Reasoning-Centric DomainsFARE, a family of large-scale parameter evaluators, surpasses specialized RL-trained evaluators in both static benchmarks and real-world tasks through data-driven development and iterative rejection-sampling supervised finetuning.Salesforce8 citations4 upvotesRead →LiveResearchBench: A Live Benchmark for User-Centric Deep Research in the WildLiveResearchBench and DeepEval provide a comprehensive framework for evaluating deep research systems across various domains, focusing on real-time web search, synthesis, and citation-grounded long-form reports.Salesforce39 citations13 upvotesRead →Hard2Verify: A Step-Level Verification Benchmark for Open-Ended Frontier MathHard2Verify, a human-annotated benchmark, evaluates step-level verifiers for LLM-based mathematical reasoning systems, highlighting the challenges and performance gaps between open-source and closed-source models.Salesforce23 citations6 upvotesRead →Webscale-RL: Automated Data Pipeline for Scaling RL Data to Pretraining LevelsA scalable data engine converts large-scale pre-training documents into diverse question-answer pairs for reinforcement learning, significantly improving model performance and efficiency.Salesforce9 citations33 upvotesRead →UNIDOC-BENCH: A Unified Benchmark for Document-Centric Multimodal RAGUniDoc-Bench is a large-scale benchmark for multimodal retrieval-augmented generation, evaluating systems across text, images, and their fusion in real-world document-centric scenarios.Salesforce27 citations17 upvotesRead →CoDA: Coding LM via Diffusion AdaptationCoDA, a 1.7B-parameter diffusion coder, achieves competitive performance with smaller models through confidence-guided sampling and is released with open-source tools.Salesforce5 citations43 upvotesRead →MCP-Universe: Benchmarking Large Language Models with Real-World Model Context Protocol ServersMCP-Universe is a comprehensive benchmark designed to evaluate large language models in realistic tasks through interaction with real-world MCP servers, addressing challenges like long-horizon reasoning and unfamiliar tool spaces.Salesforce128 citations43 upvotesRead →GTA1: GUI Test-time Scaling AgentA GUI Test-time Scaling Agent addresses task planning ambiguity and visual grounding accuracy in GUI interactions using reinforcement learning and test-time scaling.Salesforce120 citations27 upvotesRead →Demystifying Domain-adaptive Post-training for Financial LLMsFINDAP, a systematic investigation, improves LLMs for finance through continual pretraining, instruction tuning, and preference alignment, achieving state-of-the-art performance using a novel preference data distillation method.Salesforce30 citations11 upvotesRead →xLAM: A Family of Large Action Models to Empower AI Agent SystemsxLAM, a series of large action models, enhances agent performance by utilizing diverse datasets and achieving top results on benchmarks like the Berkeley Function-Calling Leaderboard.Salesforce131 citations6 upvotesRead →CodeT5+: Open Code Large Language Models for Code Understanding and GenerationCodeT5+, a family of flexible encoder-decoder LLMs initialized with frozen LLMs and trained with a mixture of pretraining objectives, achieves state-of-the-art performance across various code-related tasks.Salesforce836 citations7 upvotesRead →BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language ModelsBLIP-2 efficiently pre-trains vision-language models using off-the-shelf frozen encoders and decoders, achieving state-of-the-art performance with fewer parameters.Salesforce9,805 citations3 upvotesRead →BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and GenerationBLIP, a Vision-Language Pre-training framework, improves performance across both understanding and generation tasks by bootstrapping captions from noisy web data, achieving state-of-the-art results on image-text retrieval, image captioning, and VQA, and showing strong generalization to video-language tasks.Salesforce7,655 citations5 upvotesRead →CodeT5: Identifier-aware Unified Pre-trained Encoder-Decoder Models for Code Understanding and GenerationCodeT5, a unified encoder-decoder Transformer model, improves code understanding and generation by leveraging semantic information from identifiers and user comments.Salesforce2,633 citations3 upvotesRead →