llms.txt
Compare with other sites' llms.txt files →# Sugam Panthi
> Software engineer, researcher, and writer based in Hattiesburg, MS. Research on LLM memory, evaluation, and agents. Engineering and design essays that document real experiments and shipped work.
## Core Pages
- [Home](https://spanthi.com/): Portfolio of Sugam Panthi, software engineer, LLM memory and evaluation researcher, and writer. Bio, experience, projects, publications, and honors.
- [Blog](https://spanthi.com/blog/): Technical essays on LLM memory and evaluation, agents, engineering, and design. Each is a public companion to real experiments or shipped work.
- [Posters](https://spanthi.com/posters/): Research posters for each paper, readable in the browser and printable at A0.
- [Slides](https://spanthi.com/slides/): Talk slides, viewable full screen in the browser.
- [Resume (PDF)](https://spanthi.com/Resume.pdf): Sugam Panthi resume: experience, research, publications, and skills.
- [CV (PDF)](https://spanthi.com/CV.pdf): Sugam Panthi academic CV.
## Papers
- [LAPSE: Memory Consolidation Flattens the Temporal Shape of User Facts (arXiv:2609.36457)](https://arxiv.org/abs/2609.36457): Memory writers rewrite temporary statements ("I'm staying in Pasadena") as permanent facts ("lives in Pasadena"). LAPSE measures how consolidation drops temporal aspect.
- [Can LLMs Separate Pasted Artifacts from User Speech? Absorption at Unmarked Prompt Seams (SEAM, arXiv:2610.04210)](https://arxiv.org/abs/2610.04210): The SEAM benchmark: across 20 models, 7.7% to 66.7% absorb a trailing remark into the pasted artifact at a bare newline; explicit markers help in 19 of 20.
- [Outcome Monitors: Recovery Affordances for Silent Tool Failures (arXiv:2608.19303)](https://arxiv.org/abs/2608.19303): AI agents trust plausible but corrupted tool results; outcome-monitor receipts improve recovery across five models and two environments.
- [Compression Is Not Evaluation-Neutral: Fixed RAG Compression Can Distort Reader Comparisons (arXiv:2606.21807)](https://arxiv.org/abs/2606.21807): Giving every reader the same stored compressed evidence shrinks reader upgrades: on HotpotQA, a 31.8-point gap between readers falls to 7.8 points under one RECOMP output. Releases the ragscale audit toolkit.
- [Same Ranking, Different Winner: How Scoring Targets Shape LLM Memory Benchmarks (Findings of EMNLP 2026, arXiv:2605.24060)](https://arxiv.org/abs/2605.24060): Raw, Source, and Canonical scoring targets change LLM memory benchmark conclusions on LoCoMo and LongMemEval-S.
## Posters and Slides
- [LAPSE slides](https://spanthi.com/slides/lapse/): Talk slides for LAPSE: memory writers turn "I'm working at" into "works at" and lose the cue that a fact may not last.
- [Poster: Same Ranking, Different Winner: How Scoring Targets Shape LLM Memory Benchmarks](https://spanthi.com/poster/memory-targets/): Research poster, Sugam Panthi and Rabab Abdelfattah.
- [Poster: Outcome Monitors: Recovery Affordances for Silent Tool Failures](https://spanthi.com/poster/outcome-monitors/): Research poster, Sugam Panthi and Rabab Abdelfattah.
- [Poster: Compression Is Not Evaluation-Neutral: Fixed RAG Compression Can Distort Reader Comparisons](https://spanthi.com/poster/rag-compression/): Research poster, Sugam Panthi and Rabab Abdelfattah.
- [Poster: Can LLMs Separate Pasted Artifacts from User Speech? Absorption at Unmarked Prompt Seams](https://spanthi.com/poster/seam/): Research poster, Sugam Panthi, Muhaiminul Yeamin, and Rabab Abdelfattah.
- [Poster: Memory Consolidation Flattens the Temporal Shape of User Facts](https://spanthi.com/poster/aspect-persistence/): Research poster, Sugam Panthi, Muhaiminul Yeamin, Siyan Luo, and Rabab Abdelfattah.
- [Poster: The Stale Answer Problem: Where the Current Statement Drops Out](https://spanthi.com/poster/stale-policy/): Idea proposition poster, no paper yet, Sugam Panthi.
## Writing
- [What Is AI Memory Dreaming? ChatGPT, Claude, Devin Compared](https://spanthi.com/blog/ai-memory-dreaming/): What is dreaming in AI memory? A background process that rewrites stored memory between chats. How ChatGPT, Claude, Devin, Honcho, OpenClaw and Letta do it.
- [Thinking Gain Under Missing Evidence](https://spanthi.com/blog/thinking-gain-missing-passage/): When a multi-hop question is missing one passage, the thinking gain mostly records abstention and which passage was kept. Results on 600 comparison questions.
- [Building a Synthetic LLM Benchmark](https://spanthi.com/blog/building-a-synthetic-dataset-for-llm-memory/): How the LAPSE synthetic dataset was built: small probes first, minimal pairs, checked invented names, run-date-relative dates, built-in controls and a hand-read smoke test.
- [ARR Score Distribution: No Paper Averages 5.0 (69,781 Papers)](https://spanthi.com/blog/five-years-zero-fives/): ACL Rolling Review score distribution across 69,781 scored papers, 2021 to 2026: the mode is 2.5, no paper averaged 5.0, and meta scores swing wider.
- [Research Provenance: Rules That Keep Numbers Traceable](https://spanthi.com/blog/a-number-you-cant-trace-is-a-rumor/): The bookkeeping rules behind running concurrent paper projects: trace every number to its scorer version and run, gate paid spend, keep append-only logs.
- [LLMs Absorb Text Typed After a Paste](https://spanthi.com/blog/where-does-the-paste-end/): Across 20 LLMs, the share of comments typed after a paste that end up inside the edited text ranges from 7.7% to 66.7%. Boundary markers cut it in 19 of 20.
- [Silent Tool Failures in AI Agents](https://spanthi.com/blog/agents-believe-tools-that-lie/): Agents use corrupted tool results as fact. Receipts naming recovery tools raised ToolMaze completion from 10.9% to 28.1% across four models in two families.
- [LLM Memory Consolidation Drops Temporal Aspect](https://spanthi.com/blog/aspect-persistence-eternal-present-memory/): Memory writers store "I am driving a Peugeot" as "drives a Peugeot." On LAPSE, three writers flattened the progressive in 244 of 381 pairs, never the reverse.
- [GSAP product demos part 2: pan-zoom camera, SEO, and accessibility](https://spanthi.com/blog/gsap-choreography-part-2/): How to add cinematic zoom and pan to GSAP product demo animations: zoom wrapper architecture, pan math, easing philosophy, SEO indexing, accessibility, and the AI agent workflow.
- [LongMemEval and LoCoMo Scores Depend on the Target](https://spanthi.com/blog/your-memory-benchmark-is-lying-to-you/): Rescoring the same ranked list against a different target changed nDCG on 83% to 93% of LoCoMo and LongMemEval-S queries. Companion to an EMNLP Findings paper.
- [Scripted product demos with GSAP instead of video](https://spanthi.com/blog/gsap-choreography/): How to build a scripted product walkthrough animation with GSAP: timeline architecture, cursor choreography, click feedback, scene transitions, and responsive scaling without video files.
- [Voice Agents Need a Record of What the User Heard](https://spanthi.com/blog/every-voice-ai-forgets-what-you-heard/): Why voice-agent interruption needs more than cancellation: generated audio runs ahead of playback, and agents need heard cursors, interruption dispositions, and resumable threads.
- [RAG Compression Can Hide an LLM Upgrade](https://spanthi.com/blog/compression-is-a-coin-flip/): One stored RECOMP output shrank a 31.8-point gap between RAG readers to 7.8 on HotpotQA. Why fixed compression distorts LLM comparisons and how to audit it.
- [Technical SEO for playimposter.xyz](https://spanthi.com/blog/seo-optimization/): How we optimized playimposter.xyz for search: page intent metadata, canonical tags, hreflang, FAQ/WebApplication schema, sitemap, robots, and infrastructure-level SEO.
- [Building deepseek-go, Part 1: Design Constraints](https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-1/): A first-person technical story of how deepseek-go started: architecture goals, design constraints, and practical SDK decisions in Go.
- [Building deepseek-go, Part 2: Architecture and API Design](https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-2/): Technical breakdown of deepseek-go architecture: stream handling, modular request/response flow, provider support, and design tradeoffs.
- [Building deepseek-go, Part 3: How Adoption Changed Maintenance](https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-3/): A data-backed look at deepseek-go growth using GitHub API snapshots, star trends, release cadence, and maintenance lessons.
- [Building deepseek-go, Part 4: Maintainer Lessons and Roadmap](https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-4/): Final part of the deepseek-go series: contributor acknowledgments, maintenance lessons, and practical roadmap priorities for the SDK.
- [Building deepseek-go, Part 5: Testing the Client](https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-5-testing/): How I test deepseek-go in Go: unit layers, stream correctness checks, error consistency, and contributor-safe quality gates.
- [Building deepseek-go, Part 6: The Ollama Adapter](https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-6-ollama/): Technical deep-dive into deepseek-go's Ollama integration: adapter design, streaming behavior, compatibility boundaries, and lessons learned.
- [What Edge Deployment Changed About My ML Work](https://spanthi.com/blog/reflections-on-ai-internship/): At Prediction3d, edge deployment made data quality, memory use, latency, quantization, pruning, and exported-runtime checks part of the model-selection decision.
- [Go Web Services Guide (Part 3): Testing, Deployment, Scaling](https://spanthi.com/blog/getting-started-with-go-testing-deploy-scaling/): End-to-end release and scale guide for Go APIs including quality gates, canary rollout, rollback strategy, and production hardening.
- [Go Web Services Guide (Part 2): Data, Middleware, Auth](https://spanthi.com/blog/getting-started-with-go-data-middleware-auth/): Production-oriented Go API patterns for middleware stacks, auth validation, request IDs, database boundaries, and observability.
- [Getting Started with Go for Web Services (Part 1)](https://spanthi.com/blog/getting-started-with-go/): Detailed guide to Go web service foundations: graceful shutdown, timeouts, config strategy, API error envelopes, and practical architecture.
- [Building My Portfolio Aesthetic](https://spanthi.com/blog/building-my-portfolio-aesthetic/): How I designed a paper-aesthetic texture and integrated it into a modern React portfolio using Tailwind CSS.