AI News Today

arXiv:2607.15146v1 Announce Type: new Abstract: Online encyclopedias shape political opinion and, through it, democratic discourse. In late 2025, Grokipedia was released, an encyclopedia written entirely by the LLM Grok. One motivation behind the project was to provide an unbiased alternative to Wikipedia, which has faced accusations of "left-wing" and "liberal" bias. But does an encyclopedia written by an LLM deliver greater neutrality, or does it simply embed a different ideology? We conduct a

arXiv:2607.15175v1 Announce Type: new Abstract: Whether neural language models (NLMs) possess the ability to distinguish strings on the basis of their grammaticality remains a debated topic in the computational linguistics literature. Existing evidence has largely relied on probability-based measures, testing whether models assign higher probabilities to grammatical than ungrammatical strings. However, probability comparisons have been criticized as a measure for grammatical knowledge based on t

arXiv:2607.15178v1 Announce Type: new Abstract: Transformer reasoning is limited by autoregressive decoding, which repeat edly compresses rich hidden computation through token space and makes it difficult for intermediate reasoning states to persist across time. We in troduce Transformers with Temporal Middle-Layer Recurrence (T2MLR), a transformers-based latent reasoning architecture that fuses a cached middle layer representation from the previous token directly into an earlier layer of the cu

arXiv:2607.15200v1 Announce Type: new Abstract: Reinforcement learning has proven effective for improving reasoning in large language models, but extending it to Masked Diffusion Language Models (MDLMs) remains challenging due to the intractability of the log-likelihood estimation. Existing approaches approximate this log-likelihood by modeling only the token predictions, ignoring the order in which positions are unmasked during generation. We observe that MDLM generation involves two decisions

arXiv:2607.15209v1 Announce Type: new Abstract: Multilingual pre-trained language models (PLMs) exhibit degraded performance on low-resource, non-Latin-script languages, driven by high out-of-vocabulary (OOV) rates and excessive subword fragmentation that result from Latin-script-centric tokenizer training. We introduce VEXMLM, a vocabulary-extended variant of XLM-R targeting the two highest-resource Ge'ez-script languages, Amharic and Tigrinya, and further evaluated on 17 additional low-resourc

arXiv:2607.15232v1 Announce Type: new Abstract: A tokenizer fixed at the start of pre-training allocates vocabulary in proportion to the pre-training corpus, reflecting the deployment priorities at that time. When those priorities shift, languages added later are split into many more tokens per word, which can raise latency, compute, and energy consumption for users of those languages. Cloud models can afford a broad vocabulary because the embedding and LM-head matrices are a small fraction of t

arXiv:2607.15238v1 Announce Type: new Abstract: Language identification is commonly addressed using either neural architectures or statistical n-gram models. Neural approaches typically require substantial computational resources, whereas classical frequency-based methods offer efficient linear-time performance, but rely on distance metrics that are not always appropriate for compositional data. This work models character and bigram frequency distributions as compositional vectors constrained to

arXiv:2607.15241v1 Announce Type: new Abstract: Healthcare multimodal AI must combine visual and textual evidence while remaining reliable and interpretable. Using MediaEval Medico 2025 as a retrospective GI endoscopy case study, we analyze design choices across nine documented systems for question answering and explanation quality. Parameter-efficient adaptation of pretrained backbones provides strong challenge performance, but answer-level gains do not consistently translate into faithful and

arXiv:2607.15277v1 Announce Type: new Abstract: In-context learning is commonly interpreted as a form of conditional inference, in which the prompt specifies a context and the model's output is treated as an estimate of the corresponding conditional distribution. If this interpretation holds, then LLM estimates should satisfy basic probabilistic identities. In particular, the law of total probability asserts that prior-weighted conditional distributions aggregate into population-level marginals

arXiv:2607.14158v1 Announce Type: cross Abstract: This position paper explores how Agentic AI and Model Context Protocol (MCP) can support power-grid studies in a Transmission System Operator (TSO) context. We focus on integrating Large Language Models with numerical simulation tools, structured workflows, and human supervision. We identify key industrial requirements for agent assisted grid studies and introduce pypowsybl-mcp, an MCP-based interface exposing selected capabilities of our simulat

arXiv:2607.14159v1 Announce Type: cross Abstract: An agent harness is the external control layer that turns a base LLM into an executable agent by managing context, tools, orchestration, memory, decoding, and output handling. While harness design strongly affects agent behavior, most automatic improvement methods optimize narrower artifacts such as prompts, pipelines, or workflows, and deployed agents usually reuse a single global harness for all cases. We introduce MemoHarness, an adaptive harn

arXiv:2607.14252v1 Announce Type: cross Abstract: Long-horizon robot planning requires more than predicting what actions will do next; it also requires memory of the embodied experience that makes future goals interpretable. People do not plan from the present scene alone: they draw on remembered places, object-state changes, prior procedures, and regularities revealed through repeated action. We formulate Embodied Action Memory (EAM) as the capability to form, maintain, and use such experience

arXiv:2607.14301v1 Announce Type: cross Abstract: As generative AI (GenAI) becomes increasingly embedded in undergraduate academic writing, how students rely on these tools, rather than simply whether they use them, has become a central question for learning, academic integrity, and educational equity. Existing measures of reliance were developed inductively, focused on discrete problem-solving tasks, and validated mainly with homogeneous samples. This study developed and validated the GenAI Rel

arXiv:2607.14512v1 Announce Type: cross Abstract: Multi-step retrosynthesis planning seeks to decompose a target molecule into commercially available building blocks through a sequence of feasible reactions. The vast combinatorial search space makes this task challenging even for expert chemists. Traditional methods combine tree search with offline-trained value networks that score candidates in isolation, without reasoning about complete multi-step routes. Recent work leverages Large Language M

arXiv:2607.14524v1 Announce Type: cross Abstract: This study presents WrAFT, a Writing Assessment and Feedback Tool, that delivers both accurate and reliable scores and effective comprehensive feedback to argumentative essays. WrAFT adopts a modular design by dividing automated writing evaluation (AWE) tasks into scoring, surface-level feedback, and deep-level feedback. In building the system, various Large Language Models (LLMs) have been evaluated, including LLaMA-3.3-70B-Instruct, GPT-4o, and

arXiv:2607.14575v1 Announce Type: cross Abstract: Generative AI chatbots promise to transform English as a Foreign Language (EFL) writing by providing immediate, personalised feedback. However, their pedagogical value depends on how learners engage with them - a process often treated as a "black box." This study uses Transition Network Analysis to model the temporal dynamics of Japanese EFL learners using "Penny," an LLM-powered writing chatbot. Analysis of over 4,500 writing sessions and 21,000

arXiv:2607.14593v1 Announce Type: cross Abstract: As conversational AI systems are designed for repeated use, a central question is how a series of interactions becomes a relationship. We present a longitudinal multimodal study of a memory-augmented conversational agent (24 participants x 10 sessions), in which participants rated five relational constructs -- familiarity, self-disclosure, perceived memory, conversational quality, and enjoyment -- after each session. Two complementary dynamics em

DataRobot OpenCode: your coding agent, your model choice There are over 70 coding agents on the market: Claude Code, Codex, Cursor, Copilot, Devin, and a long tail of smaller entrants showing up in engineers’ newsfeeds every week. Someone runs a benchmark, posts the leaderboard, and by the end of the week, half the org wants... The post DataRobot OpenCode: your coding agent, your model choice appeared first on DataRobot .

<pre style='white-space:pre-wrap;width:81ex'>docs(sandbox): document multiple folders per agent (#109530) * docs(sandbox): document multiple agent folders * style(docs): format sandbox examples * docs: refresh generated map</pre>

Article URL: https://www.qolca.org/blog/stop-sending-giant-system-prompts Comments URL: https://news.ycombinator.com/item?id=48942802 Points: 3 # Comments: 0

<p>Back when I first delegated QA to an AI agent, it signed off on a tool with "all features working, pass." I opened the tool myself. The canvas was blank.</p> <p>The AI wasn't lying. In the environment it was looking at, the tool genuinely appeared to work.</p> <p>I run visual QA across a large fleet of web tools using Claude + Chrome MCP, and this class of false positive traced back to exactly two causes.</p> <h2> Cause 1: requestAnimationFrame stops in hidden tabs </h2> <p>Chrome MCP typical

Adapting existing local LLM project for security and sovereignty purposes and hopes to one day match Mythos

<p>For the past few months I've been running my personal AI assistant on my own hardware instead of a third-party cloud. Here's why — and what I learned.</p> <h2> The problem with cloud assistants </h2> <p>Every message, every preference, every bit of context you share with a hosted assistant lives on someone else's servers. You don't control retention, you can't inspect the memory, and you can't move it. For a tool that's supposed to know you, that's a lot of trust to outsource.</p> <h2> What s

<pre style='white-space:pre-wrap;width:81ex'>feat(llama-cpp): in-process local GGUF text inference provider (#109444) * feat(llama-cpp): add in-process text inference * test(llama-cpp): narrow setup provider fixture * fix(llama-cpp): trim public surface and refresh docs map * fix(llama-cpp): import Context type in inference test</pre>

<p>Worktrees isolate each AI agent's working directory, but two agents in separate worktrees can still write code that only collides when both PRs merge to main.<br> If you run several AI coding agents at once, you have probably reached for git worktrees. It is the standard answer, and it is a good one: give each agent its own worktree, and no agent overwrites another agent's files mid-run. Each agent gets a clean, isolated checkout on its own branch, off the same shared repo.<br> Use worktrees.

Aging starts at birth. ScienceAlert stories are written, fact-checked, and edited by humans, never generated by AI. Don't miss a story, subscribe here.

<p>I built a Cloudflare Worker so you can give the public free AI without ever getting a surprise bill</p> <p>Every time I wanted to add an AI feature to one of my apps and let anyone try it no login, no "enter your card," just here, use it the same thing stopped me: the bill.</p> <p>The moment your API key is reachable by the public, you're one traffic spike, one bored person with a for loop, or one leaked key away from waking up to a four-figure invoice. So most of us do one of two things: bol

<h1> corpus-scrub 0.1.0: detecta y redacta PII y secretos en corpus de entrenamiento antes del entrenamiento </h1> <blockquote> <p>Los modelos memorizan y filtran lo que hay en los datos de entrenamiento. <code>corpus-scrub</code> los limpia antes de entrenar: detecta PII (email, teléfono, IBAN, tarjeta, nombres) y secretos (API keys, tokens) y los redacta localmente, sin enviar nada a ningún servicio.</p> </blockquote> <h2> El problema </h2> <p>Los corpus raspados de la web contienen rutinariam

Article URL: https://github.com/orbsh/wiki/blob/main/tail-prompt-optimization-en.md Comments URL: https://news.ycombinator.com/item?id=48942361 Points: 1 # Comments: 0

Hi HN friends, it's july 2026 and I am assuming many of you use fully agentic building now. Securing secrets and credentials from being committed to GitHub repo has been even more critical. What's your best practice nowaday? Thx Comments URL: https://news.ycombinator.com/item?id=48942352 Points: 1 # Comments: 1

Moonshot AI’s Kimi K3 is very good — but the hype may be getting ahead of reality. (For now.)

I'm not a developer; MBA, cyber project manager by day. Years ago I taught myself enough SQL and API glue to automate the risk levels on my own trades. That was my ceiling, and every idea that needed real software died there. AI coding tools moved the ceiling, so I spent months of nights rebuilding it into something real and launched it last week. Six days later: three posts, two followers, essentially zero visitors. The best thing I have ever posted got four likes. 1 like was from me lol... The

Eg. do you update your docs/AGENTS.md/CLAUDE.md? Do you have other markdown files you write to? Did maintaining them start mattering at some point? For example with a bigger team or codebase, parallel agents etc. or has it always been the same? Do you maintain your product context somewhere else? Comments URL: https://news.ycombinator.com/item?id=48942090 Points: 1 # Comments: 1

Comments

GLM 5.2 is 35% off on AI Gateway through July 24 when routed through Novita. To get the discounted rate, set the model to zai/glm-5.2 in the AI SDK and route requests through Novita: After July 24, the model stays available at standard provider rates with no markup. Try GLM 5.2 in the model playground . Read more

You can now build native Slack agents with Chat SDK's Slack adapter . The adapter supports the full Slack agent messaging experience, from agent conversations in the Messages tab to suggested prompts, rotating status messages, token-by-token streamed replies, and native feedback buttons. Here's what the adapter gives you: Suggested prompts, per thread: Pass a static payload or an async resolver that receives the thread context, including what the user is currently viewing under agent_view . Prom

As concerns around data privacy in machine learning grow, the ability to unlearn—or remove—specific data points from trained models becomes increasingly important. While state-of-the-art unlearning methods have emerged in response, they typically treat all points in the forget set equally. In this work, we challenge this approach by asking: do points that have a negligible impact on the model’s learning need to be removed? Through a comparative analysis of influence functions across language and

The Vercel Plugin is now available in the Kimi Code CLI . Kimi Code can now draw on Vercel platform knowledge on demand, with skills for Next.js, AI SDK, Vercel Functions, and more. The Vercel Plugin also helps Kimi Code stay up to date with the latest Vercel APIs and recommended patterns. To install it, upgrade to the latest Kimi Code CLI and select the Vercel Plugin from the Third-party section of the /plugin menu. Learn more in the Vercel Plugin documentation . Read more

Vision-language models (VLMs) can follow complex textual instructions, yet they struggle to reason from purely visual context. In particular, current models fail to infer shared concepts from sets of example images and apply them to new inputs. We introduce Visual Concept Inference from Sets (VICIS), a task that evaluates this capability. Given a small context set of images sharing a concept and a query image, the model must generate new images that preserve the context-defined concept while rem

In the wake of a massive hours-long gridlock event, San Francisco mayor Daniel Lurie has told state regulators it's time to put more requirements on robotaxi operators like Waymo.