Skip to content
AI NEWS JOURNAL
AI, IN PERSPECTIVE

Search AI news

Search headlines, summaries and articles to connect the dots.

Newest first

SELD: Detecting Silent Evidence Loss in Context Compression beyond the Answer

arXiv:2611.12903v1 Announce Type: new Abstract: Context compression is now routine for language models with long inputs, but a compressed context is useful only if it keeps the evidence for the answer. Existing checks compare the reader model's answer or its most likely words, and thus miss a frequent failure, silent evidence loss: key evidence is removed while the answer stays the same, leaving the answer unsupported and risky for other models that reuse the context. To address this challenge, we propose SELD (Silent Evidence Loss Detector), which needs no training and no labels. Its principle is that when evidence is lost, a reader model's preferences among the answers it leaves unsaid shift even if its answer does not, so comparing these preferences before and after compression reveals losses that answer checks cannot see. We also build SEL-Pairs, a benchmark of length-matched evidence and control deletions from three question-answering datasets. Quantitative and qualitative experiments with five reader models on it show that SELD consistently outperforms strong and recent baselines across compressors, compression rates and models, identifying the evidence-removing edit in 83% of matched pairs with unchanged answers, against 70% for the strongest baseline. These results show that evidence loss in compressed contexts can be detected even when the answer is unchanged.

Self-stabilization and Byzantine Tolerance for Maximal Independent Set

arXiv:2210.06116v4 Announce Type: replace Abstract: We analyze the impact of transient and Byzantine faults on the construction of a maximal independent set in a general network. We adapt the self-stabilizing algorithm presented by Turau for computing such a vertex set. Our algorithm is self-stabilizing and also works under the more difficult context of arbitrary Byzantine faults. Byzantine nodes can prevent nodes close to them from taking part in the independent set for an arbitrarily long time. We give boundaries to their impact by focusing on the set of all nodes excluding nodes at distance 1 or less from Byzantine nodes, and excluding some of the nodes at distance 2. As far as we know, we present the first algorithm tolerating both transient and Byzantine faults under the fair distributed daemon. We prove that this algorithm converges in $ \mathcal O(\Delta n)$ rounds w.h.p., where $n$ is the number of nodes and $\Delta$ is the maximum degree of the network. Additionally, we present a modified version of this algorithm for anonymous systems under the adversarial distributed daemon that converges in $\mathcal O(n^{2})$ moves both in expected value and with high probability.

SetPrint: Black-Box LLM Fingerprinting via Setwise Correspondence of Response Changes

arXiv:2611.12950v1 Announce Type: new Abstract: Black-box LLM fingerprinting aims to determine whether a target model was derived from a source model when only the output responses to input queries are observable. One way to reveal this relationship is to perturb each query and observe the resulting responses of both models, yielding two corresponding sets of response variations. These sets provide two complementary sources of information: the relations among response changes within each set, and the correspondence of response changes between the two sets. Existing approaches, however, have not fully exploited both. We introduce \textbf{SetPrint}, a training-free method that jointly captures both within-set relations and between-set correspondence. SetPrint expresses each response change relative to the mean change of its own set, and then compares changes from the two models that were induced by the same perturbation. On LeaFBench, SetPrint consistently outperforms existing training-free black-box fingerprinting methods. Further analyses attribute these gains to both within-set relations and between-set correspondence.

Show HN: Beam – Pairing machines with passkeys; replace SSH&VPN's with Tailcat

A concept I wanted to share that is built on TailScale's Tailcat https://tailscale.com/tailcatThe inspiration is that I was thinking of ways to allow my users to be able to pool their machines easily without needing to set up SSH or a VPN.How can you lower the barrier to entry without compromising on security.The end result is you get a remote shell connection and a Tailcat tunnel that only requires your user to know how to use passkeys to verify their machines.Any machine can join the pool without having connected to the others at any point before which is also a strength.The primary weakness is that even though all the passkey assertion and encryption/decryption happens on the browser side you are still trusting a Cloudflare worker to store it. So I think this is more of a solution for things like Herdr, T3 code, etc etc where you want to connect machines in a pool. Each machine also has a small SQLite mail inbox for those purposes, you can queue messages that your machine will receive next time it is reachable.For a more in depth explanation I created this artifact with Fable 5.2 and Astra 6 to clean up the prose:https://claude.ai/artifact/NVHSV1FNmpPaPsqRakskVW Comments URL: https://news.ycombinator.com/item?id=50453055 Points: 1 # Comments: 0

Show HN: Black Ops II statically recompiled to native ARM64 macOS

I grew up with Black Ops 2 and wanted to see if I could make it work on my 2020 MacBook Pro. So I used Claude Opus 5.5 to create a tool that converts your copy of the game (from Steam) to a native ARM64 Mac app, with almost 60fps in all areas. The whole game is statically recompilated ahead-of-time (AOT) and does not use any emulator whatsoever, be it Wine, Crossover or Rosetta. The best thing is the performance, although some heat issues are still to fix. You get to play the whole Campaign, local Multiplayer (with bots) and local Zombies (all DLCs maps; you need to have them in your Steam library obviously) at about 60fps. Split-screen was not in the PC version, but I've been working on it lately; it's coming soon! LAN mode is on the roadmap for later too :) Comments URL: https://news.ycombinator.com/item?id=50449891 Points: 1 # Comments: 0

Show HN: Buy the whole stock market in one cheap fund and get on with your life

I have spent countless years learning and reading about personal finance. This website felt like a good way to redirect people to one place for the repeated questions like "But tech stocks", "Waiting for a crash", "What about crypto, real estate, or gold?", etc.Would appreciate some thoughts and feedback. Comments URL: https://news.ycombinator.com/item?id=50450104 Points: 2 # Comments: 1

Show HN: CodeLens AI-Augmented Code Review Tool

As AI writes more code, the bottleneck has become human-driven code reviews. Code Lens serves to aid that task by augmenting the review outside of line-level AI feedback. Check it out and let me know what you think. Thanks.Try it on a public PR: npx @blong8334/codelens pr https://github.com/expressjs/express/pull/7574 Comments URL: https://news.ycombinator.com/item?id=50446651 Points: 1 # Comments: 0

Show HN: EarthData – find places to visit or live that match what you care about

Hi HN,A few years ago, I was trying to find the place on Earth that had the optimal combination of attributes to vacation to or live. I couldn't find anything that wasn't a marketing funnel or was limited in geography. So I built EarthData essentially for myself, but others might find it useful.There are two modes: "visit" for a trip (travel month, length of stay, things to do) and "live" for relocation or retirement (housing, healthcare, day-to-day fit).The way it operates is to allow the user to sculpt the attributes they care about in where they visit on vacation or live. Rate and rank what matters to you. And it will returns 10 initial places, down to specific towns, with a written explanation of why each one fits. From there, the user can expand on places that look interesting, and ultimately pull a TRIP KIT with full itinerary, web links to venues, packing list and other information.A few design choices:- It's anonymous. No account, no email required, and preferences live in your browser's local storage. Nothing is tied to a named profile, your preferences are sent to xAI's API to generate matches, but without any PII.- No ads and no advertising cookies.- I went with a mid-century travel graphic design. To me, it evokes nostalgia and romance for international travel.- It supports 10 different languages.- The initial top 10 Matching is free. The only paid part is an optional $3.75 trip kit for a single place (itinerary, 10 matched activities, checklist, PDF). It covers the Google Places API costs for venues and links, which seem oddly 'expensive'.- Matching uses an LLM (currently xAI's Grok) to generate and reason about candidates, which get shaped into place cards, maps, short videos and activities.One area I struggled with was how to group and refine the attribute selections. Higher slider levels reveal finer-grained choices and give that topic more weight. So if it matters a lot, you can really dial in what you're looking for. If it doesn't matter to you, you can skip entirely. If you see any that seem out of line, please point it out. It was a lot of work to make sense with the slider settings, and I'm not convinced it's dialed in.Finally, if you're looking for something that doesn't have a selection, you can free-form a must-have before submitting.Feedback welcome! Comments URL: https://news.ycombinator.com/item?id=50454344 Points: 1 # Comments: 0

Show HN: Editmamei – Giving AI agents access to desktop Photoshop and GIMP

Hi HN!I built Editmamei, a local MCP server for Adobe Photoshop and GIMP.I built it because I wanted to connect Claude to a powerful photo editor, since Claude can't generate images itself. Apparently I wasn't the only one, Adobe previewed their own local Photoshop MCP yesterday, announcing at Adobe MAX a private beta in early 2027.Editmamei brings the AI agent of your choice into Photoshop and lets it do as much or as little of the editing as you want. It has tools to measure and inspect, apply filters, transform the image, insert other images and create objects, and it can use some of Photoshop's more advanced features like Select Subject and Sky Replacement.The agent "sees" each edit with downscaled previews, layer trees, histogram calculations and other tools, so someone who has never used Photoshop can ask for a recommended edit, and a professional can hand Editmamei the specific directions for tedious actions and batch automation, keeping the finishing and favorite parts for themselves. Everything is done layer by layer in a fully editable PSD, just as a normal Photoshop user would work.Photoshop has more features overall, but GIMP is just as capable for normal photo processing and design. The main difference is the Adobe-specific tools; GIMP doesn't have heal/clone or AI subject and sky selection. Photoshop runs on Windows and Mac; GIMP runs on Windows, Mac and Linux.Source available: https://github.com/editmamei/editmameiI'd love feedback, and suggestions for other editors this could support! Comments URL: https://news.ycombinator.com/item?id=50453034 Points: 1 # Comments: 0

Show HN: Vaultbase – a self-hosted back end in a single binary

One binary to run and you have a full backend at your fingertips: a database with a REST API, sign-in, realtime, files, server-side hooks and jobs, on SQLite or Postgres. It's MIT licensed. You can try the demo at https://demo.vaultbase.sh (the login is filled in, and it resets every hour). I built it to ship fast. I always tend to write the same logic for any new project, whether it's a side project or a small to medium project for a client. I wanted something my friends and I could work with fast, try new things with, and that's easy to use. Then it became something I really wanted to work on and make as complete as I can for any type of backend. As of writing this post it has 80 releases since May this year and is backed by more than 2,800 tests. I also run security reviews with Claude and fix what they find. It hasn't had an outside audit yet. It was built with the help of Claude, from the codebase to the documentation and fact-checking it against the code. Claude helped speed up the pace, find bugs, and write tests. Vaultbase is pre-1.0, so expect some bugs and issues here and there. I'm working hard to fix each and every one of them. Thank you for taking the time to read this. I'd really appreciate any feedback, even if you didn't like Vaultbase. The demo is yours to try, and it's two commands to run locally. Comments URL: https://news.ycombinator.com/item?id=50448562 Points: 1 # Comments: 0

UAEF: A framework-agnostic system for closed-loop evaluation of AI agents

The Universal Agent Evaluation Framework (UAEF) is a framework-agnostic system for evaluating AI agents. UAEF converts each framework's traces into one canonical format, scores them with 36 built-in metrics across 7 dimensions, and stores results so teams can compare agent versions and catch regressions.

A Model Organism for Studying Unverbalized Evaluation Awareness

arXiv:2611.09945v1 Announce Type: new Abstract: Large language models can recognize when they are being evaluated and change their behavior accordingly, undermining the reliability of safety benchmarks. Evaluation awareness (EA) is usually measured by having an LLM judge scan reasoning traces for explicit statements of recognition, but this signal is increasingly unreliable. Recent frontier-model system cards report a drop in verbalized rates between releases, which is indistinguishable from awareness that has gone unverbalized. Interpretability methods could detect such unverbalized evaluation awareness (UEA) in individual interactions, but lack instance-level ground truth. Asking the model alters the input, and dataset labels do not reveal whether it recognized a particular prompt as an evaluation. We address this gap with a model organism. Using synthetic document fine-tuning, SFT, and DPO, we train Qwen3.5-27B to emit a marker (a designated token) in its final response when it recognizes an evaluation, while suppressing the marker and verbalized awareness in its reasoning. The marker largely tracks the evaluation concept as it can be strongly modulated by evaluation framing in the system prompt. When not told whether it is being evaluated, the organism recognizes evaluations from test-question cues, mirroring what its base model often verbalizes. Using the marker as a label, we benchmark linear probes, the Jacobian Lens, and Activation Oracles, built on the base model and blind to the marker. None recovers the label reliably. These results suggest that current interpretability methods are not yet reliable UEA detectors, even when recognition relies on simple cues. Our model organism offers a testbed for improving them.

Adobe expands AI assistant into creative workflows

Adobe Inc. today announced an expansion of its CX Enterprise Coworker agentic artificial intelligence platform into its GenStudio content supply chain platform. The company said the integration move allows creative professionals to bring AI agents into production workflows to develop, manage and deliver personalized marketing content at large scale. The announcement, made at its Adobe […] The post Adobe expands AI assistant into creative workflows appeared first on SiliconANGLE.