WRITER · Research Lab
Frontier research for enterprise AI.
Long context, self-improving models, agent evaluation and reliability in finance and medicine. Every paper here comes with an explanation of how it works and an interactive demo you can use.
- Publications
- 27
- Interactive
- 27
- Since
- 2020
Latest research
All 27 publications →- Multi-layer Indexer+14.5%decode speed over IndexCache on DeepSeek-V3.2 (retention 1/2)LCFM Workshop, NeurIPS 2026 →
- Denoising blocks, not tokens>80×fewer generation FLOPs than the ELF-L diffusion baselineInteractive →
- Palmyra x60.785BFCL Core, the highest in its comparison cohortInteractive →
- The Harness Effect−41%blended cost per task ($0.21 → $0.12), quality at parityInteractive →
FIG. 01: Mechanism
Move through the papers. The tokens rearrange into each one’s mechanism.
Featured research
Inside the research
Eight papers to start with. Every paper in the archive has its own interactive demo.
- 01Reflect, Retry, RewardWhen a model fails, it writes a reflection on what went wrong and tries again. If the retry succeeds, only the reflection tokens get rewarded. With nothing more than a binary pass/fail signal, small models learn to self-correct well enough to beat models ten times their size.+34.7%math equation writing (Countdown)
- 02Writing in the MarginsLong documents are processed in chunks, and the model writes short notes in the margin as it reads each one. The relevant notes are carried into the final answer. Off-the-shelf models handle long context better, with no fine-tuning.+7.5%average accuracy on multi-hop reasoning (HotpotQA, MultiHop-RAG)
- 03FailSafeQAReal users misspell, ask half a question and upload the wrong document. FailSafeQA stress-tests LLMs on 10-K filings under these conditions and measures two things: whether a model stays accurate, and whether it admits when it can't answer.0.80best Context Grounding (Palmyra-Fin-128k-Instruct)
- 04Sycophancy & memoryMemory systems are supposed to make assistants better. We found they can make them agree with you even when you're wrong. When a memory system extracts facts, it often drops the correction the assistant made earlier, and the model then sides with the user's mistaken belief.1.6% → 40.2%Sonnet 4.6 sycophancy on MIST-Moral: chat history vs. Mem0 memory
- 05Prediction ≠ preventionA critic that predicts agent failures with 0.94 AUROC sounds like a safety net. Put it in charge of intervening, though, and it can cut success by 26 points. Intervening rescues some failing runs and breaks some that would have succeeded. We show how to measure that trade-off before you deploy.0.94critic AUROC (an accurate failure predictor)
- 06Outcome-oriented agent evaluationLatency and time-to-first-token say how fast an agent is. They say nothing about whether it did the job. This white paper proposes eleven outcome-based metrics, including goal completion, autonomy, resilience and business impact, and uses them to compare four agent architectures across five industries.11outcome-based, task-agnostic metrics
- 07OmniACTCan an AI agent look at a screen and write the script that completes a task, from "play the next song" to sending an email? OmniACT is a benchmark of desktop and web tasks for answering that. The strongest baseline, GPT-4, reached only 15% of human proficiency.15%of human proficiency reached by the strongest baseline (GPT-4)
- 08The Harness EffectAgent bills grow with every reasoning trace, tool schema and sub-agent prompt. Holding six models fixed and swapping only the orchestration layer, the Writer Agent Harness cut cost per task 41%, wall-clock time 44% and tokens 38%, with quality at parity.−41%blended cost per task ($0.21 → $0.12), quality at parity
FIG. 02: Findings
Findings that changed how we build.
- 0.94AUROCA failure critic that accurate still cut agent success by 26 points once it was allowed to intervene.Read the paper →
- 41%fabricationThe most robust model on FailSafeQA invented answers in 41% of cases where it should have declined.Read the paper →
- 25×sycophancyAdding a memory system raised one frontier model’s sycophancy from 1.6% to 40.2%.Read the paper →
- 10×smallerModels trained to reflect on their own failures beat same-family models ten times their size.Read the paper →
FIG. 03: Program
Four directions, one standard: does it hold up in production?
Enterprise-optimized models
Models built for reliability, long context, tool use and cost at enterprise scale.
Practical evaluations
Evaluation methods that reflect real deployments, real risks and real failures.
Domain-specific specialization
Specialized models and benchmarks for finance, medicine and other high-stakes domains.
Retrieval & knowledge
Retrieval and knowledge integration that holds up on real enterprise data.
FIG. 04: Archive
All research
2026
- Multi-layer Indexer: Overlapping Selection and Attention in DeepSeek Sparse AttentionInteractivePaperEnterprise-optimized models
- Denoising Blocks, Not Tokens: Efficient Compressed Continuous Diffusion with Branching Token RealizationInteractivePaperEnterprise-optimized models
- Palmyra x6 Technical Report: An Agentic, Tool-Use Model Post-Trained via Anchored Supervised Fine-TuningInteractiveModelEnterprise-optimized models
- The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AIInteractivePaperEnterprise-optimized models
- Recalling Too Well: Sycophancy Evaluation and Mitigation in Memory-Augmented ModelsInteractivePaperPractical evaluations
- Shorthand for Thought: Compressing LLM Reasoning via Entropy-Guided SupertokensInteractivePaperEnterprise-optimized models
- The Price of Agreement: Measuring LLM Sycophancy in Agentic Financial ApplicationsInteractivePaperDomain-specific specialization
- Accurate Failure Prediction in Agents Does Not Imply Effective Failure PreventionInteractivePaperPractical evaluations
2025
- Towards Outcome-Oriented, Task-Agnostic Evaluation of AI AgentsInteractivePaperPractical evaluations
- Palmyra-mini: Small models, big throughput, powerful reasoningInteractiveModelEnterprise-optimized models
- Reflect, Retry, Reward: Self-Improving LLMs via Reinforcement LearningInteractivePaperPractical evaluations
- Palmyra X5: The 1M-token enterprise LLMInteractiveModelEnterprise-optimized models
- Expect the Unexpected: FailSafe Long Context QA for FinanceInteractiveBenchmarkDomain-specific specialization
2024
- Palmyra Creative: Unlocking creativity with AIInteractiveModelDomain-specific specialization
- Introducing self-evolving modelsInteractiveResearch postEnterprise-optimized models
- Introducing actions with Palmyra X4InteractiveModelEnterprise-optimized models
- Writing in the Margins: Better Inference Pattern for Long Context RetrievalInteractivePaperEnterprise-optimized models
- Palmyra Fin: A domain-specialized LLM for financeInteractiveModelDomain-specific specialization
- Comparative Analysis of Retrieval Systems in the Real WorldInteractivePaperRetrieval & knowledge
- Grammatical Error Correction for Code-Switched Sentences by Learners of EnglishInteractivePaperEnterprise-optimized models
- OmniACT: A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and WebInteractiveBenchmarkPractical evaluations
- Prompting open-source and commercial language models for grammatical error correction of English learner textInteractivePaperPractical evaluations
2023
- Achieving State-of-the-Art Open-Domain QA Performance through Fusion-in-DecoderInteractivePaperRetrieval & knowledge
- On the application of Large Language Models for language teaching and assessment technologyInteractivePaperDomain-specific specialization
- Becoming self-instruct: introducing early stopping criteria for minimal instruct tuningInteractivePaperEnterprise-optimized models
- Palmyra-Med: Instruction-Based Fine-Tuning of LLMs Enhancing Medical Domain PerformanceInteractivePaperDomain-specific specialization
- Grammatical Error Correction: A Survey of the State of the ArtInteractiveSurveyPractical evaluations
No research in this area yet.