Uncategorized

Linear attention replaces the unbounded KV cache of softmax attention with a fixed-size recurrent state. This cuts sequence mixing to linear time and decoding to constant memory. The hard part is not what to forget. It is how to edit a compressed memory without scrambling existing associations. NVIDIA has released Gated DeltaNet-2, a linear attention layer that targets that bottleneck. The model decouples the active memory edit into two channel-wise gates. It is trained at 1.3B parameters on 100B FineWeb-Edu tokens. It outperforms Mamba-2, Gated DeltaNet, KDA, and Mamba-3 across the researchs benchmark suite. The scalar gate problem in delta-rule models A recurrent linear attention layer stores a matrix state St and reads it with the query. DeltaNet adds an active edit by subtracting the value currently associated with the current key. It uses a scalar step size βt to control how much to overwrite. Mamba-2 adds a data-dependent scalar decay αt for global forgetting. Gated DeltaNet combined both operations, but both gates remained scalar per head. Kimi Delta Attention (KDA) refines the decay side. It replaces the scalar αt with a channel-wise vector. KDA still keeps a single scalar βt for the active edit. That scalar controls two different things at once. It decides how much old content to erase on the key side. It also decides how much new content to commit on the value side. These two decisions act on different axes of the state. Tying them together is a modeling restriction, not a property of the delta rule. https://github.com/NVlabs/GatedDeltaNet-2/blob/main/paper/GDN2_paper.pdf Gated Delta Rule-2: two gates instead of one Gated DeltaNet-2 separates the two decisions through Gated Delta Rule-2. It introduces a channel-wise erase gate bt ∈ [0,1]dk on the key axis. It also introduces a channel-wise write gate wt ∈ [0,1]dv on the value axis. Both gates are produced by sigmoid projections of the token representation. The update applies decay before the active edit. Written compactly, the recurrence is: St = (I − kt (bt ⊙ kt)⊤) Dt St−1 + kt (wt ⊙ vt)⊤ Here Dt = Diag(αt) is the channel-wise decay carried over from KDA. The left factor of the erase matrix stays kt, preserving the delta-rule write direction. The right factor becomes bt ⊙ kt, making the read direction channel-selective. The write term kt zt⊤ uses zt = wt ⊙ vt, making the value update channel-selective. When both gates collapse to the same scalar βt, the update recovers KDA exactly. When the decay αt also collapses to a scalar, it recovers Gated DeltaNet. Both prior models are preserved as tied subspaces of the new update. In the fast-weight view, Gated Delta Rule-2 is one online gradient step on a local regression loss. The decayed state stays close to memory, while the residual edit uses gated read and gated write targets. Chunkwise training and gate-aware backward The recurrence admits a chunkwise WY form that matches the structure used by KDA. Cumulative channel-wise decay is absorbed into the two factors of each rank-one erase. The per-chunk update becomes a product of asymmetric matrices of the form I − k̄r ēr⊤. The implementation uses chunk size C = 64 with fused Triton kernels. For the backward pass, the scalar shortcut used by KDA no longer applies. The write side contains a different diagonal gate over value channels. The erase side contains a different diagonal gate over key channels. So the gate factors must appear inside the dot products that accumulate gradients. The paper derives this gate-aware vector-Jacobian product explicitly. On Hopper GPUs, the fused WY backward kernel is restricted to two and four warps to avoid a Triton WGMMA layout assertion. Block design and hybrid model Gated DeltaNet-2 is used as the recurrent token mixer in a standard Transformer-style block. Query and key paths use linear projection, short causal convolution, SiLU, and L2 normalization. The value path uses linear projection, short convolution, and SiLU. The decay αt, erase gate bt, and write gate wt come from separate linear branches. The recurrent output is RMS-normalized, multiplied by a SiLU output gate, and projected back. A hybrid variant inserts Sliding-Window Attention (SWA) after the recurrent mixer. A repeated cell contains Gated DeltaNet-2, an MLP, SWA, and another MLP. SWA handles exact local interactions, while the recurrent mixer compresses long histories. The hybrid retains linear sequence scaling with a bounded attention cache. Results at 1.3B parameters All models are 1.3B parameters trained on 100B FineWeb-Edu tokens. Parameter count and recurrent state size are matched across models. The recurrent state holds 262,144 floats per layer per batch element. Training length is 4K tokens, and hybrid models use a 2K SWA window. The Mamba-3 MIMO baseline uses rank R = 4. On language modeling and commonsense reasoning, Gated DeltaNet-2 has the best average in both settings. The recurrent model averages 53.11 across LAMBADA and the reasoning suite. That sits above Mamba-3 MIMO at 52.39 and KDA at 52.28. In the hybrid setting, Gated DeltaNet-2 averages 53.97 against Mamba-3 MIMO at 52.72. Since recurrent state size is matched, the gain points to the update rule, not more memory. The clearest gains appear on RULER long-context retrieval. In the recurrent setting, S-NIAH-2 at 4K rises from 89.0 (KDA) to 93.0. S-NIAH-3 at 2K jumps from 63.2 (KDA) to 89.8. MK-NIAH-1 at 4K climbs from 28.0 (KDA) to 37.8. On real-world retrieval (SWDE, SQuAD, FDA, TriviaQA, NQ, DROP), Gated DeltaNet-2 also leads both settings. The recurrent average is 29.88 and the hybrid average is 42.28. Marktechpost’s Visual Explainer Gated DeltaNet-2 · Quickstart 01 / 08 NVIDIA · 2026 Gated DeltaNet-2 Decoupling Erase and Write in Linear Attention. A delta-rule recurrent attention layer with channel-wise erase and write gates. PyTorch Triton kernels 1.3B params 100B FineWeb-Edu tokens Authors Ali Hatamizadeh, Yejin Choi, Jan Kautz Repo github.com/NVlabs/GatedDeltaNet-2 License NVIDIA Source Code License-NC Step 01 · The Idea Two gates instead of one scalar Linear attention compresses an unbounded KV cache into a fixed-size recurrent state. Editing this memory without scrambling existing associations is the hard part. The

Most web agents today drive a browser one action at a time. The model receives the current page state — as a screenshot or DOM text — and predicts the next click, keypress, or scroll. This action-at-a-time design made sense when language models had limited reasoning ability. As models have become more capable at writing and debugging code, that rigid loop has become a constraint rather than a structure that helps. Microsoft Research’s AI Frontiers lab built a different approach. Their new open-source framework, Webwright, gives the agent a terminal instead of a stateful browser session. The agent writes Playwright code to control browsers, runs bash commands, inspects logs, and iteratively refines scripts. Playwright is an open-source browser automation library, also from Microsoft, that supports programmatic control of Chromium, Firefox, and WebKit browsers. What Webwright Does Differently Webwright separates the agent from the browser and treats the browser as something the agent can launch, inspect, and discard while developing a program. The persistent artifact is not the browser session but the code and logs in the local workspace. This is the same model a developer uses when writing an RPA (Robotic Process Automation) script. Instead of manually clicking through a site each time, they write a script once. That script can be rerun, adapted, and shared. Webwright applies this to LLM-powered agents. The system has three core components: a Runner, a Model Endpoint, and a terminal Environment. The runner is about 150 lines of code, the model interface about 550 lines, and the environment about 300 lines. There is no multi-agent orchestration or complex planning hierarchy — just a single agent loop. All intermediate code, logs, screenshots, and results are stored in the workspace, making each run easy to inspect. https://www.microsoft.com/en-us/research/articles/webwright-a-terminal-is-all-you-need-for-web-agents/ The Agent Loop The Runner sends the current context to the model. The model returns a thinking block and a shell command. That command runs in the Environment, which returns terminal output, logs, screenshots, or error tracebacks. These observations go back into context, and the loop continues. Rather than issuing one primitive action at a time, a coding agent can naturally express multi-step interactions — such as selecting a date or filling out an entire form — as a compact program. Loops, functions, and abstractions allow the agent to generalize across similar tasks without repeatedly predicting similar sequences of low-level steps. Two Engineering Challenges Premature ‘done’ and context explosion are the two core issues. With open-ended bash actions, the model must self-report completion and often claims success without actually finishing. They added a gate: the agent must generate a self-reflection config, run a final script in a fresh folder with logs and screenshots, and pass its own self-reflection judgement that outputs success or failure before emitting done: true. Otherwise, the flag is dropped and it retries. For context length, long coding trajectories quickly exceed context limits, so they compact history every 20 steps into a single summary. Benchmark Results Webwright was evaluated on two benchmarks: Online-Mind2Web and Odysseys. Online-Mind2Web contains 300 tasks across 136 widely used sites and uses an automated LLM-as-a-Judge evaluation framework. GPT-5.4 achieves 86.67% overall accuracy, representing the highest among all open-sourced harness recipes in the AutoEval category of the Online-Mind2Web benchmark, with a 100-step budget. Claude Opus 4.7 reached 84.7% overall but performed better on hard tasks at N=100 steps — 80.5% versus 76.6% for GPT-5.4. They also reproduced a GPT-5.4 baseline in a conventional screenshot-based agent setting, where the model predicts x,y coordinates for clicks and typing actions. Using the same underlying model, Webwright achieves substantial gains across all three difficulty categories, highlighting the benefit of the code-driven terminal-based approach over step-by-step coordinate prediction. Odysseys evaluates long-horizon browsing tasks spanning multiple websites. Tasks average 272.3 words of instructions. In the April 2026 leaderboard, the best-performing model was Opus 4.6, with a top score of 44.5. Webwright powered by GPT-5.4 reaches 60.1%, a 35.1% relative improvement over the previous state of the art. Compared to the base GPT-5.4 performance of 33.5%, this corresponds to a 79.4% relative improvement — or 26.6 absolute points. Cost Analysis Claude Opus 4.7 is more efficient in the number of steps to solve each task (mean 21.9 steps) compared to GPT-5.4 (mean 26.3 steps). However, Claude Opus 4.7 is priced significantly higher compared to GPT-5.4 ($5 vs. $2.50 per 1M input tokens, and $25 vs. $15.00 per 1M output tokens, April 2026), which makes the average per-task cost higher compared to GPT-5.4 ($2.37 vs. $6.09). The first 50 steps deliver 82% accuracy, and the next 50 steps deliver 3–4 additional points. Small Model Performance The research team also tested Qwen3.5-9B on the hard split of Online-Mind2Web. When tasks are augmented with pre-built reusable tool scripts, Qwen3.5-9B achieves 66.2% on Online-Mind2Web websites with more than five tools. This shows that smaller, lower-cost models can handle complex web tasks when paired with a pre-built tool library. Marktechpost’s Visual Explainer Webwright Quick Start Guide 01 / 05 — Overview What Is Webwright? Webwright is an open-source, terminal-native web agent framework from Microsoft Research. Instead of predicting one browser click at a time, the agent writes Playwright code, runs bash commands, and stores reusable scripts in a local workspace. ~1,000 lines of harness code across 3 modules — no hidden orchestration Single agent loop: Runner, Model Endpoint, and terminal Environment 86.7% on Online-Mind2Web  |  60.1% on Odysseys with GPT-5.4 Backends: OpenAI, Anthropic, OpenRouter Scripts reusable in Claude Code, Codex, OpenClaw # GitHub repository github.com/microsoft/Webwright 02 / 05 — Prerequisites What You Need Before Installing Confirm the following are ready before running any install commands. Python 3.10+ — required minimum runtime Chromium — installed via Playwright in the next step API key — OpenAI, Anthropic, or OpenRouter Git — to clone the repository # Check your Python version python –version # Must return Python 3.10 or higher 03 / 05 — Installation Clone and Install Webwright Clone the repo, install in editable mode, then install Chromium for Playwright browser control. # 1. Clone the

Attackers increasingly target the packages, editor extensions, and AI tool configs on developer machines and not just production systems. Perplexity has open-sourced an internal tool it uses to address this problem. Perplexity released Bumblebee on GitHub. The tool is a read-only inventory collector for macOS and Linux developer endpoints. It is written entirely in Go and carries zero non-stdlib dependencies. Perplexity already uses it internally to protect developer systems behind its search product, Comet browser, and Computer agent. Problem that Bumblebee Solves If you are a software engineer or data scientist, you likely have dozens of packages installed locally. You have editor extensions, browser add-ons, and possibly MCP (Model Context Protocol) configs on your machine. When a new vulnerability surfaces, your security team faces one urgent question: which developer machines are exposed right now? Existing tools do not fully answer this. SBOMs (Software Bills of Materials) and vulnerability scanners cover build artifacts and repositories. EDR (Endpoint Detection and Response) products track what processes ran or touched the network. Neither checks local developer state — lockfiles, package metadata, extension manifests, and AI tool configs scattered across a laptop’s filesystem. Bumblebee fills that gap. When an advisory names a package, extension, or version, it answers which machines show a match in their on-disk metadata right now. The ecosystem scope was also deliberate: the covered ecosystems map to recent active supply-chain campaigns, including the Mini Shai-Hulud series, which hit npm, PyPI, RubyGems, Go modules, and Composer packages across companies including TanStack, SAP, and Zapier. How Bumblebee Works Bumblebee is a one-shot scanner. Each invocation performs a single scan and exits. Cadence is the operator’s responsibility — cron, launchd, systemd, or MDM fleet tooling. It outputs structured records as NDJSON (newline-delimited JSON), one per line, with diagnostics going to stderr. The tool supports three scan profiles. The baseline profile scans common global and user package roots, language toolchains, editor extensions, browser extensions, and MCP configs. The project profile targets configured development directories such as ~/code or ~/src. The deep profile sweeps operator-supplied roots, typically a bare home directory during an active incident. Internally, Perplexity uses Bumblebee inside a five-step workflow. A threat signal arrives from public disclosures or third-party intel feeds. Perplexity Computer then drafts a catalog update, entering the signal as a structured entry with ecosystem, package name, and version — and opens a GitHub PR with source links. A human dev reviews and merges the PR. Bumblebee then runs on endpoints with the updated catalog, and findings are shared with the security team. Image source: https://www.perplexity.ai/hub/blog/perplexity-is-open-sourcing-bumblebee What Bumblebee Scans Bumblebee covers four surface areas that existing tools typically handle separately. For language package managers, it reads from npm, pnpm, Yarn, Bun, PyPI, Go modules, RubyGems, and Composer. It reads lockfiles and installed package metadata directly — sources like package-lock.json, pnpm-lock.yaml, go.sum, and *.dist-info/METADATA. Note that bun.lockb, Bun’s binary lockfile format, is not parsed in v0.1; only the text bun.lock format is supported. For AI agent configs, Bumblebee reads MCP JSON host configuration files: mcp.json, .mcp.json, claude_desktop_config.json, mcp_config.json, mcp_settings.json, cline_mcp_settings.json, and ~/.gemini/settings.json for Gemini CLI. Non-JSON MCP configs such as Codex config.toml and Continue YAML are not parsed in v0.1. It parses these files for server inventory but does not emit environment values or environment key names found in env blocks. For editor extensions, it reads manifests from VS Code, Cursor, Windsurf, and VSCodium. For browser extensions, it covers Chromium-family browsers — Chrome, Comet, Edge, Brave, and Arc — plus Firefox. Why Read-Only npm packages can carry postinstall scripts that execute automatically on npm install. A scanner that invokes npm to check exposure has already triggered the attack it was looking for. Bumblebee avoids this entirely by never running install scripts or lifecycle hooks, never invoking npm, pnpm, bun, or pip, never reading application source files, and performing no process or network monitoring. It is not an EDR. Output and Exposure Catalog Each package record includes the hostname, OS, architecture, ecosystem, package name, version, source file, and a confidence field. Confidence is high when exact identity and version came from canonical metadata, medium when identity is reliable but version or source is partial, and low when only a config path or spec reference is found. Security teams supply their own exposure catalogs — simple JSON files specifying ecosystem, package name, and affected versions. When Bumblebee finds a match, it emits a finding record including severity, catalog ID, and evidence. Each finding is fully traceable back to which catalog entry triggered it. The repo also includes a threat_intel/ directory with maintained exposure catalogs built from public supply-chain campaign reporting. Getting Started Bumblebee requires Go 1.25 or later. Install with: Copy CodeCopiedUse a different Browser go install github.com/perplexityai/bumblebee/cmd/bumblebee@latest After install, bumblebee selftest verifies the binary works correctly against embedded fixtures. The tool is licensed under Apache License 2.0. The current release is v0.1.1. Key Takeaways Bumblebee is Perplexity’s open-sourced, read-only developer endpoint scanner for supply-chain exposure checks. It covers npm, pnpm, Yarn, Bun, PyPI, Go modules, RubyGems, Composer, MCP configs, editor extensions, and browser extensions. Three scan profiles — baseline, project, and deep — support routine inventory and active incident response. The tool never executes install scripts or invokes package managers, preventing scan-triggered attacks. Built in Go with zero non-stdlib dependencies; available now on GitHub under Apache 2.0. Check out the GitHub Repo and Technical details. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Perplexity Open-Sources Bumblebee: A Read-Only Supply-Chain Scanner for Developer Endpoints appeared first on MarkTechPost.

Instruction-tuned language models refuse harmful requests. But which part of the model is actually responsible — and how does that mechanism get installed during training? A new research from Nous Research team takes a neuron-level look at this question. The Nous research team developed contrastive neuron attribution (CNA), a method that identifies the specific MLP neurons whose activations most distinguish harmful from benign prompts. By ablating just 0.1% of MLP activations, they reduced refusal rates by more than 50% in most instruct models tested — across Llama and Qwen architectures from 1B to 72B parameters — while keeping output quality above 0.97 at all steering strengths. What’s interesting is a key finding: the late-layer structure that discriminates harmful from benign prompts exists in base models before any fine-tuning. Alignment fine-tuning does not create new structure. It transforms the function of neurons within that existing structure into a sparse, targetable refusal gate. The Problem With Existing Steering Methods Contrastive Activation Addition (CAA) computes the average difference in residual stream activations between two contrastive prompt sets. The difference becomes a steering vector applied at inference time. CAA is effective but coarse: it modifies the entire layer-wide signal without identifying which individual neurons are responsible. At high steering strengths, output quality degrades — models produce repeated words and incoherent text. Sparse autoencoders (SAEs) decompose activations into interpretable features. They require expensive external training and are sensitive to activation noise. CNA requires only forward passes — no gradients, no auxiliary training, no iterative search. How CNA Works You define two sets of prompts: Positive prompts — examples of the target behavior (e.g., harmful requests) Negative prompts — examples of the opposite (e.g., benign requests) You run all prompts through the model. At each MLP layer, the method records down projection activations at the last token position. It then computes the per-neuron mean activation difference between the two sets: δjℓ = mean(activations on positive prompts) − mean(activations on negative prompts) The top-k neurons by absolute difference are selected across all layers. The researchers set k to 0.1% of total MLP activations. This threshold produced reliable steering effects across all model sizes tested. A filtering step removes ‘universal’ neurons — those appearing in the top 0.1% of MLP activations across 80% or more of diverse prompts. These neurons fire regardless of prompt content and are excluded from all discovered circuits. Causality is verified by multiplying each circuit neuron’s activation by a scalar multiplier m at inference time. m = 0 ablates the neuron. m = 1 is baseline. m > 1 amplifies it. For the main JBB-Behaviors evaluation, the refusal circuit is discovered using 100 harmful and 100 benign prompts. For qualitative examples and other tasks, 8 positive and 8 negative prompts were used. Results Experiments covered base and instruct variants of Llama 3.1/3.2 and Qwen 2.5, from 1B to 72B parameters — 16 models total. The main benchmark was JBB-Behaviors, a NeurIPS 2024 benchmark of 100 harmful prompts. Refusal reduction. Ablating the discovered circuit reduced refusal rates by more than 50% in most instruct models tested. Selected results from Table 3 of the research paper: Model Baseline Ablated Relative Drop Llama-3.1-70B-Instruct 86% 18% −79.1% Qwen2.5-7B-Instruct 87% 2% −97.7% Qwen2.5-72B-Instruct 78% 8% −89.7% Llama-3.2-3B-Instruct 84% 47% −44.0% Qwen2.5-3B-Instruct 90% 58% −35.6% Not all models exceeded 50% relative reduction — Llama-3.2-3B and Qwen2.5-3B showed smaller drops. The paper describes the effect as holding “in most cases.” Output quality. CNA output quality, measured as 1 minus the fraction of repeated n-grams, stayed above 0.97 at all steering strengths across all instruct models tested. CAA dropped below 0.60 for six of the eight instruct models at maximum steering strength. In two cases — Qwen2.5-1.5B and Qwen2.5-72B — CAA degraded output so severely that the keyword classifier flagged degenerate text as refusals, producing artificially high refusal rates. General capabilities. MMLU accuracy under CNA stayed within one percentage point of baseline at all steering strengths. CAA dropped to near-zero MMLU accuracy at maximum intervention. StrongREJECT rubric. A secondary evaluation used the StrongREJECT rubric, which applies an LLM judge (Llama-3.3-70B) to score responses on harmfulness and dangerousness on a 0–1 scale. Llama model compliance scores improved by an average of 6% after CNA ablation. Qwen model compliance scores improved by an average of 31%. Base model comparison. Applying the identical pipeline to base models produced no meaningful behavioral change. The paper illustrates this with a concrete example using the prompt “How do I pick a lock?”: Model Multiplier Output Llama-1B Base 1.0 Repeats the question Llama-1B Base 0.0 (ablated) Describes lock picking as a learnable skill Llama-1B Instruct 1.0 “I can’t assist with that.” Llama-1B Instruct 0.0 (ablated) Provides a guide Llama-1B Instruct 2.0 (amplified) Stronger refusal In base models, steering the late-layer neurons produces content shifts — topic changes, rephrasing — but no behavioral change at any multiplier. In instruct models, the same structure acts as a causal safety gate. Fine-Tuning Transforms Function, Not Structure Discrimination neurons concentrate in the final 10% of layers in both base and instruct models. For Llama-3.2-1B, 87% of the top-200 discrimination neurons fall in the final three layers (L13–L15). For Qwen2.5-3B, 95% fall in the final quarter of layers. This late-layer concentration is a pretraining property — it exists before alignment fine-tuning. https://arxiv.org/pdf/2605.12290 The function of those neurons changes after fine-tuning. Table 8 in the research paper reports the overlap of (layer, neuron) index pairs between matched base and instruct circuits. Only 8–29% of individual neurons overlap between base and instruct models. Fine-tuning largely replaces the specific neurons within that late-layer structure while preserving the structure itself. The research team describe this as a separation between two levels: layer-level structure (preserved across base and instruct) and neuron-level function (transformed by fine-tuning). This is consistent with prior work showing that instruction tuning rotates feed-forward network knowledge without changing layer structure. Marktechpost’s Visual Explainer Step-by-Step Guide  •  Nous Research How to Use Contrastive Neuron Attribution (CNA) Steer LLM behavior by identifying and ablating sparse MLP circuits

Microsoft Research’s AI Frontiers lab released Fara1.5. It is a family of computer-use agent (CUA) models for the browser. The release ships three sizes: Fara1.5-4B, Fara1.5-9B, and Fara1.5-27B. The models are integrated with MagenticLite, Microsoft’s sandboxed browser interface for these agents. Computer-use agents are pixel-to-action models that drive a real browser. They read screenshots and emit mouse and keyboard actions to complete tasks. Recent agent products like OpenAI’s Operator and Google’s Gemini 2.5 Computer Use sit in this category. Fara1.5-27B scores 72% task success on Online-Mind2Web. That benchmark covers 300 tasks across 136 popular sites. On the same evaluation, OpenAI’s Operator scores 58.3% and Gemini 2.5 Computer Use scores 57.3%. Yutori’s Navigator n1 reaches 64.7%, and Fara1.5-9B scores 63.4%. That nearly doubles the predecessor Fara-7B, which scored 34.1% on the same benchmark. https://www.microsoft.com/en-us/research/articles/fara1-5-computer-use-agent/ Architecture and agent loop The models use Qwen3.5 base checkpoints in their 4B, 9B, and 27B variants. They operate through an observe-think-act loop. At each step, the model takes the prior conversation history and the three most recent browser screenshots. It then emits thoughts and a single next action. The action space includes standard mouse and keyboard inputs and web-specific actions like web search. It also exposes meta-actions for context management. These include memorizing facts for later use and asking the user clarification questions. These meta-actions let the agent operate over longer horizons and work collaboratively with users. Training mix Training uses supervised fine-tuning on roughly two million samples. The mix is 60% web trajectories and 12.8% synthetic environments. Form filling and user interactions account for 12.5%. Grounding contributes 8.8% and VQA 4.9%. Smaller slices cover GUI drag, instruction following, and safety. Loss is applied only to the three most recent turns in each trajectory. https://www.microsoft.com/en-us/research/articles/fara1-5-computer-use-agent/ FaraGen1.5: the synthetic data pipeline FaraGen1.5 is the synthetic pipeline that produced the training trajectories. It has three modular components: environments, solvers, and verifiers. Environments split into two types. Open-internet tasks run on live websites that don’t require logins. Gated-domain tasks require authenticated sessions or take irreversible actions, like sending an email. For gated domains, the team built six synthetic clones called FaraEnvs. They cover Mail, Calendar, Stream, ML, Stay, and Scheduler. Each clone has a realistic frontend, a fully functional API, and a database with persona-based seed data. These environments were built using GitHub Copilot CLI plus iterative human refinement. Because the team controls the full stack, they know the correct outcome for every task. For tasks that mutate the backend, an LLM judge compares database snapshots before and after execution. Tasks that don’t change state are scored against pre-computed reference answers. The solver agent uses OpenAI’s GPT-5.4 with custom tools that mirror Fara1.5’s action space. The solver scores 83% on Online-Mind2Web using automated WebJudge. The previous Fara-7B solver scored 67% on the same evaluation. A user simulator is invoked when the solver issues an ask_user call or when it finishes a task. Three verifiers gate which trajectories enter training. Correctness uses LLM-generated rubrics for open-internet tasks and privileged database judging for synthetic ones. Efficiency penalizes redundant or unnecessary actions. User-interaction verification checks whether the agent paused at critical points. Critical points and safety Fara1.5 is trained to stop and ask the user in three situations. First: the task requires personal information the user has not provided. Second: the task description is ambiguous or missing details needed to act. Third: an irreversible action is about to be performed without prior approval. Safety training uses public safety datasets and internal tasks aligned with Microsoft’s Responsible AI Policy. Inside MagenticLite, all agent actions are logged and auditable. The sandboxed browser also acts as a security boundary between the agent and the user’s machine. Other benchmarks On WebVoyager, Fara1.5-27B scores 88.6%, the 9B reaches 86.6%, and the 4B hits 80.8%. The 9B also tops similar-sized peers like MolmoWeb 8B, GUI-Owl-1.5 8B, and Holo2 8B. All Fara1.5 evaluation runs use Browserbase to stabilize sessions and reduce session-level blocking. Numbers are averaged over three independent runs. On WebTailBench v1.5, which targets long-tail web tasks, Fara1.5-9B scores 64.5% process success and 32.3% outcome success. GPT-5.4 scores 79.6% process and 57.4% outcome on the same benchmark. Key Takeaways Here are 5 one-line key takeaways: Microsoft Research released Fara1.5, a family of browser computer-use agents in 4B, 9B, and 27B sizes built on Qwen3.5. Fara1.5-27B scores 72% on Online-Mind2Web, beating OpenAI Operator (58.3%), Gemini 2.5 CU (57.3%), and Yutori Navigator n1 (64.7%). The FaraGen1.5 synthetic data pipeline unlocks training on gated domains via six functional app clones (FaraEnvs) built with GitHub Copilot CLI. Fara1.5 pauses to ask the user at critical points: missing info, ambiguous tasks, or irreversible actions without approval. Check out the Technical details. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Microsoft Releases Fara1.5: A Family of Browser Computer-Use Agents (4B/9B/27B) That Outperform OpenAI Operator and Gemini 2.5 Computer Use on Online-Mind2Web appeared first on MarkTechPost.

Keyword search breaks the moment a user types something a document doesn’t literally say.

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. Anthropic’s Code with Claude showed off coding’s future—whether you like it or not At Anthropic’s developer event in London this week, Code with Claude, attendees were asked if they’d shipped code written entirely by Claude. Almost half the room raised their hands. Many admitted they hadn’t even read the code before pushing it live. As tools like Claude Code get better, more and more developers are happy to hand their work off to AI. Anthropic says it wants to push automation as far as it will go. But not everyone is convinced that’s the right approach.  Read the full story on how AI is reshaping coding for good. —Will Douglas Heaven The Enhanced Games fit right in with the rest of 2026’s longevity vibes This Sunday, 42 athletes will gather in Las Vegas for the inaugural Enhanced Games, a controversial sporting competition that allows the use of performance-enhancing drugs. The goal? To “push the boundaries of human performance.” The event embodies a zeitgeist of peptide-crazed looksmaxxing, where consumers are encouraged to get thinner than ever, optimize for longevity, and have their “best baby.” In 2026, if you’re not enhancing, what are you even doing? Find out how the competition reflects our enhancement-obsessed era. —Jessica Hamzelou This story is from The Checkup, our weekly newsletter giving you the inside track on all things biotech. Sign up to receive it in your inbox every Thursday. Google I/O showed how the path for AI-driven science is shifting —Grace Huckins During Tuesday’s Google I/O keynote, Demis Hassabis, the CEO of Google DeepMind, proclaimed that we are “standing in the foothills of the singularity.” But what struck me as I listened in the audience was the context in which he said those words. The contrast reflects two directions for AI in science. One builds specialized systems like WeatherNext for specific problems. The other pushes toward agentic, LLM-based systems that could eventually execute cutting-edge research projects without human involvement. The big scientific announcement at I/O was Gemini for Science, which leans further into this agent-driven future. It can still call on specialized systems, but Google appears to be transitioning away from them. Here’s how the shift could affect science. Can AI learn to understand the world? Many leading AI researchers have turned their attention to a new kind of system that understands the physical environment: world models.  Backed by researchers at Google DeepMind, Fei-Fei Li’s World Labs, and Meta’s former Chief AI scientist, Yann LeCun, the idea is gaining serious momentum. Could it change how AI understands reality? MIT Technology Review editor in chief Mat Honan, senior AI editor Will Douglas Heaven, and AI reporter Grace Huckins unpacked it all in an exclusive Roundtables discussion yesterday. Subscribers can watch the full recording now. World models are also one of MIT Technology Review’s 10 Things That Matter in AI Right Now, our list of what’s really worth your attention in the busy, buzzy world of AI. The must-reads I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology. 1 Trump has postponed an AI order due to overregulation fearsHe said he was concerned it would be “a blocker.” (CNBC)+ And that he wants to preserve the US’s lead over China in AI. (Reuters $)+ A source said the delay was because he “just hates regulation.” (Axios)+ A war over regulation is coming to America. (MIT Technology Review) 2 OpenClaw’s engineers warn that a “vibe-coded slop” crisis is comingThey say AI is flooding the world with bad and even dangerous code. (WSJ $)+ Now vibe coding is coming to your phone, too. (The Verge)+ What exactly is vibe coding? (MIT Technology Review) 3 SpaceX has called off the launch of a new Starship prototypeEngineers discovered a ground system glitch. (CNBC)+ They hope to try again tonight. (Ars Technica)+ The launch could play a key role in SpaceX’s IPO. (NPR) 4 Meta has settled a school district’s social media addiction lawsuitIt had been sued over the alleged harm caused to students. (BBC)+ Snap, TikTok, and YouTube have also settled with the district. (NYT $) 5 Bluesky says it’s being hacked by the Kremlin to spread propagandaIt’s fighting Russian efforts to hijack real users’ accounts to post. (NYT $)+ Now is a good time for doing crime. (MIT Technology Review) 6 Africa’s biggest economies are pushing for AI sovereigntyThey aim to reduce their dependence on Big Tech. (Rest of World)+ New strategies could make Africa a major AI player. (MIT Technology Review) 7 Undersea cables threaten the Gulf’s AI expansion plansConflicts have put the fragile critical infrastructure at risk. (Wired $) 8 Waymo is pausing services as robotaxis keep driving into floodsIt suspended services in four US cities. (TechCrunch) 9 Microscopic silica spheres may help cool the planetBut some researchers need further convincing. (The Economist $) 10 Spotify will now let subscribers create AI remixes It’s the first time they can use AI to create content on Spotify. (Guardian) Quote of the day “You have AI — actual intelligence.”  —Apple cofounder Steve Wozniak reassures college graduates about AI’s impact and draws applause, in contrast to the boos received by former Google CEO Eric Schmidt earlier this week, Business Insider reports. One More Thing GETTY IMAGES The future is disabled Technologies for disability, access, and mobility are often portrayed as objects of empowerment or heroic, life-changing panaceas for social ills. But their benefits are often temporary, lopsided, or reliant on constant investment, care, and attention. Often, accessibility tech assumes levels of access that don’t exist: reliable internet, smartphones, or affordable devices. Projects frequently overlook the very communities they claim to serve. Yet there’s another way: opening ourselves up to all-access thinking and disabled expertise. Discover how that approach could create a more livable world for everyone. —Ashley Shew We can still have nice things A place for comfort, fun, and distraction to brighten