frontpage.
newsnewestaskshowjobs

Made with ♥ by @iamnishanth

Open Source @Github

fp.

Open in hackernews

Resetting RAG-based LLMs doesn't reset behavior

https://github.com/VeritasAdmin/audit-grade-ai-workstation
1•URS_Adherent•2h ago

Comments

URS_Adherent•2h ago
I’ve been working on a small, independent evaluation framework to test a simple question:

Do common “reset” procedures in retrieval-augmented LLM systems (thread isolation, context flushing, cooldowns, re-initialization) actually return the system to a clean behavioral state?

Rather than testing prompts or jailbreaks, I treated this as a measurement problem.

The approach: - define clean vs. contaminated runs - apply standard reset/isolation procedures - analyze output statistically, not semantically - look for short lexical signatures that persist across resets

What I found is not instructions, payloads, or exploits — but consistent lexical residue that appears only in contaminated runs and survives resets that should have neutralized prior influence.

I’m sharing: - a short methodology appendix (PDF) - a design rationale explaining why laptop-class hardware invalidates deterministic evaluation for this workload

I am deliberately not sharing prompts, payloads, reproduction steps, or vendor-specific claims.

I’m posting this to get feedback on the measurement approach itself: - Does this seem like a reasonable way to test reset robustness? - What controls would you add or remove? - Have others seen similar residue in RAG or tool-augmented systems?

Methodology appendix (PDF): https://github.com/VeritasAdmin/audit-grade-ai-workstation/b...

Star – Your Digital Songbook

https://starapp.io
1•mianala•1m ago•0 comments

74% of European firms would fail without access to U.S. technology

https://europeancorrespondent.com/en/r/trumps-power-switch
1•speckx•2m ago•0 comments

Freeimageai.org is your go-to hub for free AI image generation and editing

https://freeimageai.org
1•zhouhua•2m ago•1 comments

Cecli AI Coding Assistant

https://cecli.dev
1•tomjuggler•2m ago•1 comments

Epsteinomatic: Turn Your Memories into Crimes

https://epsteinomatic.com/
1•dezmou•3m ago•0 comments

Show HN: Mouse-tracker – A tiny physics-based mouse tracker (Web Component)

https://gimli.app/mouse-tracker
1•gimliapp•3m ago•0 comments

Show HN: Radiant – Radial Menu Launcher for macOS Inspired by Blender's Pie Menu

1•sagawafumiya•4m ago•0 comments

The Left has a Hyperpolitics problem

https://newrepublic.com/article/205820/left-protests-hyperpolitics-building-political-power
1•RickJWagner•4m ago•0 comments

How to Migrate Your Custom GPTs to Claude

https://aiforcontentmarketing.ai/what-to-do-with-your-custom-gpts-when-you-switch-to-claude/
1•pakostina•5m ago•0 comments

The Screening Machine

https://www.tabletmag.com/sections/science/articles/screening-machine
1•RickJWagner•5m ago•0 comments

LingBot – open weights world model

https://huggingface.co/robbyant/lingbot-world-base-cam
1•nikhizzle•6m ago•0 comments

Show HN: AppControl – A Modern Windows Task Manager with History

https://www.appcontrol.com/
1•suprnurd•7m ago•1 comments

Show HN: Shuffled - Daily word puzzle game

https://shuffled.app
1•wmora•7m ago•0 comments

Show HN: The Control and Memory Layer for AI Agents

1•dan_lupashku•7m ago•0 comments

Show HN: Vela – Modern programming language compiling to native code via LLVM

https://github.com/MigMarGil/Vela_lang
1•MMG_dev•8m ago•0 comments

Show HN: Early detection of LLM hallucinations via structural dissonance

https://github.com/yubainu/SL-CRF
1•yubainu•8m ago•1 comments

America's $1T AI Gamble

https://www.apricitas.io/p/americas-1t-ai-gamble
1•m-hodges•9m ago•0 comments

Show HN: Octrafic – AI agent for API testing from your terminal

https://github.com/Octrafic/octrafic-cli
1•mbadyl•10m ago•0 comments

Accelerando, but Janky

https://taoofmac.com/space/blog/2026/02/06/1245
1•rcarmo•11m ago•0 comments

Show HN: Model Tools Protocol (MTP) – Forget MCP, bash is all you need

https://github.com/modeltoolsprotocol/modeltoolsprotocol
5•nr378•11m ago•2 comments

Thoughts on AI-Assisted Software Development in 2026

https://taoofmac.com/space/notes/2026/02/01/2130
1•rcarmo•11m ago•0 comments

Show HN: Sign Any PDF Free – No account, no watermarks, no limits

https://signanypdffree.com/
1•detroitwebsites•13m ago•0 comments

AgentVault: Security Wrapper for OpenClaw (built in a couple hours))

https://github.com/hugoventures1-glitch/agentvault
1•hugoventures1•14m ago•1 comments

UK justice ministry orders deletion of largest court archive court

https://twitter.com/europa/status/2020851497437708297
1•nomdep•14m ago•0 comments

SAIR: Terence Tao's Foundation Uniting Nobel, Turing, Fields Laureates and AI

https://sair.foundation/
1•badr_elmazaz•15m ago•0 comments

Show HN: GPU ROI simulator based on token usage and model architecture

https://axiomos.ai/decide/
1•pierreseck•16m ago•1 comments

Hex Fiend – Simple hexadecimal math game

https://do-say-go.github.io/hexfiend/
1•keepamovin•18m ago•0 comments

Show HN: Darna – Atomic commit validator for Go

https://github.com/darccio/darna
1•darccio•19m ago•0 comments

What we can learn from tiny traces of ancient blood chemicals

https://theconversation.com/life-in-fossil-bones-what-we-can-learn-from-tiny-traces-of-ancient-bl...
1•PaulHoule•19m ago•0 comments

"The things I am good at"

https://cnicodeme.com/the-things-i-am-good-at
3•cx42net•20m ago•0 comments