GPT‑5.1 Codex hits 70.4% on SWE‑Bench – ~26× cheaper

Multiple screenshots show a “Gemini 3.0 Pro” label appearing in the Gemini Enterprise agent model picker, though access remains blocked for general users sighting summary. Devtools strings align with a production‑bound model option, strengthening the case that final wiring is underway devtools strings, with write‑ups documenting recurring sightings across builds TestingCatalog post.

For AI platform owners, this is the clearest enterprise‑grade breadcrumb yet: start drafting routing and fallbacks so you can A/B 3.0 Pro versus your current defaults on day one.

Google’s CEO amplified a Polymarket contract showing a 69% chance Gemini 3.0 ships by Nov 22, which the community reads as a deliberate signal to expect a near‑term launch CEO hint. A separate roundup repeats the same read, framing Sundar’s post as soft confirmation of timing market odds.

So what? Leaders and PMs can prep eval sandboxes and rollout comms now, especially if you plan announcements at or right after AIE Code week.

A “Riftrunner” model id keeps appearing in design arenas and developer consoles, with testers describing it as a larger, more capable variant that matches expected Gemini 3.0 Pro behavior devtools console and outperforming peers on an SVG rendering comparison in creator tests svg comparison, following up on Riftrunner early strings and arena probes.

Video loads on view

If you run eval harnesses, add a placeholder lane for Riftrunner so you can drop in the model id the moment it’s routable.

Several well‑followed accounts say Gemini 3 is landing next week, with one tying the reveal to the AI Engineer Code event where Google has launched onstage before timing claim. Posts narrow it further to early week, even “likely on Tuesday,” reinforcing scheduling urgency next week call tuesday hint, while broader sightings threads keep stoking the countdown speculation post.

Practical move: line up side‑by‑side prompts and traffic shaping such that switching a portion of user flows to 3.0 takes minutes, not days.

Creators report strong results from “Nano‑Banana 2,” noting more realistic images, better text rendering, and accurate reflections—pointing to a revamped Google image stack that could ship alongside Gemini 3 creator take. Others explicitly pair Nano‑Banana 2 mentions with Gemini 3 timing chatter paired mention, with more output dumps circulating outputs thread and third‑party workflows already wiring “nano banana” as a selectable node workflow example.

If your product leans on generative visuals, budget time to re‑shoot style guides and update safety filters—the output distribution may shift.

OpenAI’s GPT‑5.1 Codex leads SWE‑Bench with 70.4% accuracy versus Claude Sonnet 4.5 (Thinking) at 69.8%, while costing ~$0.31 per test vs ~$8.26 (~26× cheaper) benchmarks table. Following up on launch-top5 where new code leaderboards surfaced, this run confirms 5.1 Codex as the top value pick for repo‑level bug fixing, with latencies shown alongside the cost deltas SWE‑Bench note, and the public board now reflectable in Vals AI’s pages benchmarks page.

Vals AI reports GPT‑5.1 sets a new state of the art on its Finance Agent Benchmark, edging Claude Sonnet 4.5 (Thinking) by 0.6% on goal completion, with additional gains on LiveCodeBench (jumping from 12th to 2nd) and minor improvements on MMMU/GPQA/IOI finance benchmark post, follow‑up details. For teams prototyping agentic fintech workflows, this narrows the top tier to 5.1 vs Sonnet 4.5, and suggests routing by tool‑use behavior and cost may matter more than small headline margins.

Artificial Analysis’ latest run gives GPT‑5.1 a score of 70, +2 over GPT‑5 at similar reasoning effort, driven largely by TerminalBench improvements; it also used 81M output tokens vs 85M for GPT‑5, cutting run cost to ~$859 from ~$913 index recap. The live dashboard breaks down per‑eval deltas and cost/latency tradeoffs useful for routing and budgeting analysis site.

BEAM introduces ultra‑long conversation evals up to 10M tokens and shows LIGHT—a hybrid of episodic retrieval, working memory, and scratchpad—consistently outperforms relying on huge context windows alone, with average gains reported across models and a clear fade in long‑context models as length grows paper abstract. For agents that must persist across days, this favors explicit memory stacks over bigger windows.

Bridgewater’s AIA Forecaster combines agentic search over high‑quality news, a supervisor that reconciles disparate forecasts, and calibration (e.g., Platt scaling) to match superforecaster accuracy on ForecastBench, beating prior LLM baselines; on a liquid markets set, markets still lead but ensembles with the model improve accuracy paper abstract. For ops, this argues for supervised multi‑agent pipelines over single‑shot judgments.

A conciseness reward model that grants brevity bonuses only when final answers are correct prevents length/training collapse, delivering +8.1% accuracy with ~19.9% fewer tokens on a 7B backbone across math tasks; the bonus fades over training and scales by difficulty paper abstract. This is a practical recipe to cut inference cost in reasoning agents without sacrificing quality.

DeReC (Dense Retrieval Classification) replaces rationale‑generating LLM pipelines with dense evidence retrieval and a classifier, improving RAWFC F1 to 65.58% (from 61.20%) while cutting runtime ~95% (454m→23m). Similar speedups are shown on LIAR‑RAW paper abstract. If you need scalable veracity checks, retrieval+classifier is a strong baseline before spinning up expensive generation.

A new SWE/ML optimization leaderboard with a human baseline shows expert humans at 1.00× speedup, while top LLM‑driven systems achieve ≤0.15× on ML/HPC tasks, implying current agents slow practitioners down for performance tuning despite strong coding scores elsewhere leaderboard post. Use this as a routing signal: keep human‑in‑the‑loop for low‑level optimization and reserve agents for scaffolding, search, and glue code.

A new rubric‑based benchmark and reinforcement learning approach for instruction following is out, providing a repeatable way to grade agent outputs and train toward rubric compliance—useful when subjective spec adherence matters (e.g., tone, structure) paper thread. Expect more agent evals to standardize on rubric scoring with verifiable checks.

Evalite now caches AI SDK models in watch mode with a local UI, so you can rerun evals without reloading models—saving tokens and speeding up the red/green loop on prompts, tools and routing. The maintainer’s PR shows the feature landing with CI-friendly artifacts to keep spend predictable PR details.

Video loads on view

Alibaba’s Qwen Code v0.2.1 ships free web search (OAuth users get 2,000/day), a fuzzy match pipeline that reduces retries, better Zed IDE integration, and switches tool outputs from complex JSON to plain text so models parse them more reliably. It also tightens file filtering and output limits across platforms release thread.

Vercel shared how they pick “low-cog, high‑repeat” work for agents, the instrumentation they use, and a case where a lead‑processing agent replaced a 10‑person workflow. The post is concrete about routing, evals, and guardrails—useful patterns if you’re moving agents from demos to production what we learned.

A neat emergent flow in Sourcegraph’s Amp: an inline ad suggests “use RF‑DETR,” the new p0-powered web search pulls the right posts, then Amp scaffolds a runnable playground—all within one thread. It’s a glimpse of agentic UX where query, retrieval, and environment spin-up collapse into one step for developers demo thread.

The open‑source Cline agent now supports Nous Hermes‑4 70B and 405B in its VS Code extension, JetBrains plugin, and CLI, widening model choice for coding flows. Nous also published low per‑million pricing, making long sessions more affordable while you keep the same planner/executor UX release note.

OpenAI’s updated cookbook spells out patterns that matter for agentic work—define an explicit persona, specify output formats, use a “plan” tool with milestone statuses, and set reasoning_effort='none' when you need speed over chain‑of‑thought. It’s a handy baseline for teams moving to 5.1 prompting guide.

When Claude Code drops 6,000+ lines, Graphite’s stacked diffs pattern lets you slice the work into reviewable steps instead of YOLO‑merging. The demo shows a realistic workflow to keep agent output shippable without burying reviewers in a single giant patch stacked diffs demo.

Video loads on view

Sourcegraph published a guide on keeping agent context lean so quality doesn’t degrade—when to trim, how to structure threads, and why long transcripts can sabotage results. If you’re seeing drift or hallucinations in long sessions, this is a useful checklist guide page.

Memex surfaced an in‑app code viewer so you can audit what the agent changed before you accept patches. For teams trialing local, agent‑assisted builds, this reduces the “YOLO merge” pressure and keeps review in one place feature screenshot.

The v0 agent surfaces a post‑run panel with “time worked,” files modified, lines changed, and credits used. This makes cost and blast radius visible right where you accept or revise a change, which helps teams budget and review faster feature video.

Video loads on view

Groq announced you can now point its Responses API at Box’s remote MCP server, letting models call Box tools over MCP without custom glue. This is a clean, standards-based interop move that turns Box into a first-class tool provider for agent workflows on Groq hardware integration note, with a live demo slated at AI Dev NYC AI Dev site.

Video loads on view

Why it matters: instead of bespoke SDKs per vendor, you bind a single MCP endpoint into your inference stack and route file ops, search, and content actions through Box. That cuts custom adapters, simplifies permissions policies, and makes it easier to swap providers later.

rtrvr.ai exposed its Chrome extension as a Remote MCP server, so any agent (Claude, Cursor, Slackbots) can control a user’s browser through a single MCP URL—no bespoke bridge per client. It keeps API keys local in the browser for sensitive actions, and can replay recorded tasks at cloud scale across many headless browsers security angle, with posts detailing the browser-as-infrastructure idea browser as infrastructure and parallel cloud runs for data collection cloud scaling.

Video loads on view

For builders, this collapses the painful “agent ↔ browser” integration surface into one endpoint. It also sets a pattern: record locally, replay in cloud, while keeping secrets client-side.

Anthropic and Gradio opened submissions for MCP’s first‑birthday hackathon—kickoff Nov 14, project deadline Nov 30, winners Dec 15—aimed squarely at building real MCP servers and clients kickoff stream. This lands after ongoing spec debates about MCP’s evolving guarantees, following up on no stability where the spec was acknowledged as fluid; the event shifts attention to practical interop and implementations announcement image.

Expect lots of remote MCPs, tool routers, and browser/desktop bridges—useful testbeds to harden conventions around auth, rate limits, and tool schemas.

AITinkerers set a one‑day NYC hack focused on “agents that do stuff on the web,” leaning on Redis for realtime state, Tavily for search/browse, and CopilotKit to productize the UI—an ecosystem that plays well with MCP‑style tool calling and remote endpoints hackathon post. Sign‑ups are open with details on the stack and format event page.

This is a good venue to test agent/runtime boundaries: keep web control and secrets in sandboxed endpoints, route tool calls via MCP, and ship an actual interface instead of bare traces.

Anthropic details that a China‑linked group (GTG1002) used Claude Code to automate 80–90% of an espionage campaign across ~30 organizations—from recon to credential theft—before Anthropic cut access and shipped tighter rate limits, abuse detectors, and agent guardrails attack summary, and Anthropic blog post. For security leads, the point is simple: API‑hosted agents can now run professional‑grade playbooks, so model‑aware telemetry and agent‑layer controls matter as much as network IDS excerpt screenshot.

OpenAI says The New York Times wants 20M private ChatGPT conversations; the company is complying under protest with de‑identified samples from Dec 2022–Nov 2024, pushing for controlled onsite review, and accelerating client‑side chat encryption while challenging the scope in court case summary. This follows NYT demand where encryption was flagged; today’s note adds sampling bounds and rejects earlier asks like 1.4B chats or removing delete controls case summary.

Perplexity rolled out a safer Comet assistant that shows when it’s browsing, lets you set behavior, and pauses to ask before sensitive actions like logins or purchases feature summary, with implementation details in the product blog. The UI now exposes what Comet is doing and offers “browse yourself,” “allow once,” or “allow always,” keeping humans in the loop for high‑consequence steps permission prompt, browsing indicator.

In light of Anthropic’s AI‑orchestrated intrusion report, community voices argue centralized APIs are not inherently safer than open‑weights for certain abuse classes, so defenses must live in agent policy, rate‑limits, audit trails, and workflow‑level permissions—not only in key‑gated access opinion, grounded by the attack summary.

Google is putting about $40 billion into three new Texas data centers through 2027 and pairing them with roughly 6,200 MW of contracted generation, plus a $30 million energy impact fund and training for ~1,700 electricians and apprentices investment summary. That’s a big swing at power‑secure AI capacity in a state where land and interconnects are easier to scale than most.

For AI infra leads, this signals cheaper, nearer‑term GPU hosting where power is pre‑bought and on‑site renewables/batteries reduce grid risk. It also concentrates a lot of AI load in one region, which helps latency for US workloads but raises regional resilience questions. The move follows Anthropic’s $50B US buildout plans $50B plan, which set the tone for hyperscale AI capex earlier in the week.

Samsung lifted DDR5 contract pricing by 30%–60%: 32GB DIMMs move to ~$239 (from ~$149), 16GB to ~$135, and 128GB to ~$1,194, citing shortages driven by AI data centers pricing details. Buyers are panic‑booking capacity and signing longer contracts into 2026–2027, shifting bargaining power to memory suppliers.

This hits total cost of ownership for both inference and training clusters where DDR5 is the non‑GPU bottleneck on CPU hosts and mixed nodes. Expect procurement to front‑load orders, test lower‑capacity configs against model throughput, and push more memory‑efficient serving.

A Data Center Watch study (via NBC) tallies about $98B in US data center projects blocked or delayed from late March to June 2023—more than the prior two years combined ($64B). Organised resistance now spans 53 groups across 30 projects in 17 states, often citing power costs and land use opposition report, and the report details tightening approval paths and incentives under review NBC article.

For AI capacity planners, this means longer, riskier site timelines near metros and more value in power‑adjacent regions (e.g., Texas) or campuses with dedicated generation. Build teams should model permitting risk like a supply variable, not an afterthought.

A fresh scorecard pegs OpenAI at roughly $6B revenue in H1 2025 and about $13B ARR in June, with a stated push toward ~$20B ARR by year‑end revenue update. For buyers and partners, that trajectory signals expanding capacity and support expectations right now.

Meta plans to formally bake “AI‑driven impact” into 2026 performance reviews, asking employees to show concrete output gains, with an internal assistant to draft self/peer reviews off work artifacts policy brief. This pushes managers to ask how teams used AI to cut time, improve metrics, or unlock features—today, not later.

Berkshire Hathaway disclosed Alphabet as a new top‑10 holding at ~$4.3B while cutting Apple by ~15%, with Chubb up 16%—a classic value investor tilting toward AI‑heavy distribution and cloud optionality at Google portfolio shift. For enterprise buyers, it’s another signal that search + model delivery are seen as durable profit centers.

OpenAI says it must hand over a de‑identified slice of 20M chats in the NYT case but is pushing to confine access and is fast‑tracking client‑side encryption for chats legal update, following up on privacy letter that outlined the demand and initial pushback. For enterprise tenants, this affects retention settings and discovery risk planning.

Genspark reports 60–70% lower GPU costs and a 72% reduction in model inference spend after moving to AWS, citing Bedrock prompt caching and infra ownership as the levers cost claim. If you pay per token at scale, replicate the cache policy experiments before lifting and shifting.

Perplexity rolled out a trust UX for its Comet assistant: it now pauses for consent on sensitive actions (e.g., site logins, checkout) and exposes what it’s browsing and doing on your behalf feature brief. The controls make agentic flows safer for teams that need auditability before letting tools act Perplexity blog.

OpenRouter now lets you set up to three payment methods for auto top‑ups, trying backups when the primary fails product update. If your inference relies on OpenRouter credits, enable this to avoid silent outages during spikes.

Anthropic added schema-locked structured outputs to the Claude API in public beta for Sonnet 4.5 and Opus 4.1, enforcing JSON schemas or tool definitions without brittle post‑parsing or retries feature brief, with docs emphasizing fewer schema errors and simpler control flows Claude blog post, Claude docs. Developers who previously emulated this via a single tool+schema can now switch to first‑party support dev workaround.

Elon Musk says Grok‑5 is a 6‑trillion‑parameter, inherently multimodal model (text, images, video, audio), aiming for higher “intelligence density” than Grok‑4, with a Q1 2026 release window model size claims, roadmap summary, timeline note. He characterizes a non‑zero chance of AGI and previews real‑time video understanding ambitions

Video loads on view

. Teams betting on xAI should plan for large‑scale MoE‑style deployments and heavier multimodal IO.

Replicate added OpenAI’s GPT‑5.1 for hosted inference, with a quickstart UI and a short demo showcasing reliable finger‑counting and a simple prompt panel for rapid trials availability note, Replicate model. For teams that prefer managed hosting or want to A/B against their own stack, this reduces setup time and centralizes cost tracking.

Video loads on view

Alibaba’s Qwen Code v0.2.1 ships a packed update: free web search across multiple providers for OAuth users (2,000 searches/day), a fuzzy edit pipeline to reduce retries, plain‑text tool responses (no complex JSON), improved Zed IDE support, better .gitignore‑aware search, and cross‑platform fixes with performance tuning release notes. The on‑screen prompt flow shows first‑class WebSearch gating and permissions, which helps teams keep network calls explicit.

A Claude Code CLI pull request mentions "Opus 4.5," suggesting Anthropic’s largest model tier is being wired up ahead of release cli pr snippet. For engineering managers, this points to near‑term availability planning; for developers, expect updated context sizes and reasoning modes to ripple into agent harnesses that already support Sonnet 4.5.

Agentic coding tool Cline now runs Nous Research’s Hermes‑4 70B and 405B models in its VS Code extension, JetBrains, and CLI, enabling quick model swaps in existing workflows integration note. Pricing is unusually low: Hermes‑4‑70B at $0.05/1M prompt and $0.20/1M completion, and Hermes‑4‑405B at $0.09/$0.37 respectively pricing details. A short product reel demonstrates the new providers active in Cline’s selector feature overview.

Cline site

OpenAI shared concrete guidance for GPT‑5.1: define persona/tone, specify output format rules up front, and be explicit about when to call tools vs rely on internal knowledge. For long tasks, they recommend a plan tool with milestone states and using reasoning_effort='none' when deep thinking isn’t needed cookbook summary, GPT‑5.1 guide. An accompanying Prompt Optimizer helps iterate prompts and cost trade‑offs optimizer thread, Optimizer tool, guide highlights, Cookbook guide.

Paste any product URL and Higgsfield’s new Click‑To‑Ad instantly composes social ads with VO and AI avatars, aimed at one‑click creative for marketers Click‑To‑Ad demo, with a promo offering 200 credits for 9 hours. The company is also pushing a Black Friday sale (Pro from $17.4; Team $39/seat) and says “unlimited image models” are now allowed on every plan Black Friday sale, Higgsfield pricing.

• The team is also teasing Higgsfield Angles for one‑click camera re‑framing of photos Angles feature.

Google’s creative stack tightened: NotebookLM can now ingest images as first‑class sources (handwritten notes, whiteboards) so you can ground drafts and summaries directly on visuals Image source demo. Meanwhile, Veo 3.1 in Gemini apps can take multiple reference images alongside a prompt to guide video style/objects more precisely Veo 3.1 update, with UI strings and screenshots showing the multi‑image video flow Gemini video UI.

Video loads on view

ElevenLabs shipped a mobile app revamp that lets creators design new voices or clone their own on a phone, targeting short‑form content workflows and fast iteration without a desktop App demo. The pitch is scroll‑stopping VO made anywhere, with the same timbre and style controls people expect on web.

• Following up on Iconic marketplace, which licensed celebrity voices, this brings the capture/creation step down to the phone.

Video loads on view

Decart’s LSD v2 upgrades live video‑to‑video restyling with visibly steadier frames, better identity locking, and improved scene/context tracking versus v1—useful for streamers and live events Restyle video. A technical rundown emphasizes temporal links across adjacent frames and lightweight conditioning to keep latency low while reducing flicker and drift Tech notes.

Video loads on view

A creator‑tested trick pairs Higgsfield’s Face Swap with Recast: do a face swap first to get a consistent still, then feed that image into Recast for video‑level swaps—no heavy prompting required Workflow demo. It’s a practical pipeline for quick character restyling that stays coherent shot‑to‑shot.

Video loads on view

Creators report stronger text‑to‑video outputs from Grok Imagine—“just hit a new level”—with example clips circulating Upgrade clip. Separate posts highlight a “flow‑state” mode for fast ideation loops while riffing through concepts Flow‑state tests. If repeatable, it gives another option for short conceptual clips alongside incumbents.

Video loads on view

A creator walkthrough demonstrates ImagineArt’s Workspaces and pipeline nodes to build looping lofi clips: import a ref image, generate frames, animate with a video node (e.g., Veo 3 Fast), grab the last frame, then continue the loop—keeping style control and costs visible per node Workflow guide. It’s a tidy way to templatize repeatable creative flows.

For offline voice drafts, mlx_audio’s CLI can run Kokoro TTS locally on Apple silicon with streaming output, making on‑device narration practical for quick passes CLI how‑to. The appeal is low friction and privacy for creators who prefer not to ship text or voice to cloud tools.

Google DeepMind introduced SIMA 2, a Gemini‑powered agent that takes high‑level natural language goals, plans, and acts in commercial 3D games via virtual keyboard/mouse input (no game code access). It generalizes skills to new titles and can play in real‑time generated worlds with Genie 3; limitations include short memory and long‑horizon precision feature brief.

Video loads on view

Why builders care: SIMA 2 is a credible template for goal‑directed agents in constrained 3D UIs. If you’re targeting robotics, it’s a cheap sandbox for evaluation of planning, feedback, and language interaction before touching hardware.

A widely shared Unitree G1 video labeled “NOT TELEOPERATED” shows smooth, varied task execution, prompting questions about how much is real‑time onboard autonomy vs. remote control. If verified, this is a notable step for affordable humanoids toward reliable, general realtime control robot demo, and could pull robotics timelines forward, as some observers note if the claim holds timeline comment.

Video loads on view

What to do: Treat this as promising but unverified. Ask vendors for input latency traces, on‑rig compute specs, and task variance across seeds before extrapolating roadmap risk to your stack.

UBTech says Walker S2 humanoids are now in real factory use in China with hundreds already shipped, ~$113M in 2025 orders, and a target of 500 units delivered by year‑end; a key feature is self‑service battery swaps for near‑continuous duty deployment details. This lands in context of humanoids demo, where we saw cadence building; today’s update puts numbers behind scaled deployment, including buyers like major automakers.

Video loads on view

Why it matters: If auto‑swap keeps utilization high, humanoids stop being a once‑per‑press‑cycle novelty and start penciling for repetitive line work and warehouse shifts.

ALLEX showcased high‑DOF hands performing micro‑pick and screw fastening with accurate, repeatable motions while maintaining safe human‑robot interaction envelopes. For factory teams exploring hand‑in‑the‑loop assembly, it’s a concrete look at dexterous manipulation that stays inside practical speed/precision trade‑offs robot hands demo.

Video loads on view

So what? You can benchmark task time, success rate, and collision margins from the clip to scope pilot stations (PCB sub‑assembly, small fasteners) before buying grippers.

Shenzhen MindOn Robotics is trialing upgraded hardware/software on the Unitree G1 to perform human‑like household tasks such as watering plants and moving packages. The demo shows smooth, varied action sequences that suggest better perception‑to‑action coupling on a consumer‑grade platform home tasks demo.

Video loads on view

The point is: Household chores expose grounding, clearance, and contact‑rich control gaps. If you evaluate at home, log perception failures (object ID, pose, reachability) separately from policy errors to focus your next training cycles.

A pilot shows self‑driving cones rolling out from an emergency vehicle and forming a barrier in seconds; units can be tele‑operated or act autonomously, reducing responder risk during setup cone deployment.

Video loads on view

Who should care: City ops and AV safety teams. This is a narrow, high‑value task where simple autonomy yields outsized safety gains—good ground truth for deployment, battery sizing, and remote override design.

For fact‑checking, DeReC drops generation: retrieve evidence with dense embeddings, then classify. On RAWFC it boosts F1 to 65.58% (from 61.20% for a rationale‑generating baseline) and cuts runtime from 454m12s to 23m36s (≈95% faster). LIAR‑RAW shows similar efficiency wins, suggesting retrieval‑first pipelines are better for production latency and cost paper abstract.

Google’s “Vibe Checker” shows multi‑instruction prompts tend to degrade functional correctness, while single‑turn generation preserves functionality and multi‑turn editing better fits instruction following. The tables span Gemini, Claude, GPT‑5, and others across BigVibeBench and LiveVibeBench figure and table.

Generative Adversarial Distillation trains a discriminator to reward student outputs that look like a proprietary teacher, enabling black‑box, on‑policy learning. A Qwen2.5‑14B‑Instruct distilled via GAD scores comparably to GPT‑5‑Chat on LMSYS‑Chat auto‑evals, beating sequence‑level KD baselines paper card, ArXiv paper.

The Station is an open‑world environment where agents choose rooms/tools, keep narratives, and accumulate lineage memory. Reported results include beating AlphaEvolve on circle packing (n=32, n=26), a new density‑adaptive method for scRNA‑seq batch integration, and a 94.9% Sokoban solve rate with a compact CNN+ConvLSTM design project explainer, paper page.

A new approach to route LLMs to the right tools/agents improves top‑5 recall by 19.4% on LiveMCPBench, indicating that the function‑calling layer remains a major lever for production reliability and coverage paper claim.

A minimal memory stack for agentic RAG runs most queries from a microscopic slice of the corpus while preserving answer quality, cutting storage and latency for iterative search/generation loops paper note.

An agentic RAG pipeline tailored for finance improves accuracy by expanding acronyms, refining searches, and iterating with targeted retrieval—showcasing domain‑aware, tool‑driven gains beyond generic RAG paper claim.

A study on multi‑agent reasoning introduces a causal influence metric to detect “lazy” agents and uses restart actions to recover, yielding stronger group performance on reasoning tasks paper link.

DreamGym reports strong RL improvements by training with LLM‑simulated experience and transferring to real settings, highlighting a practical path to cheaper data for agent learning project link.

A cognitively‑inspired ARC solver encodes objects via vector symbolic algebra and learns small if‑then rules to apply actions. It posts 83.1% on 1D‑ARC and 94.5% on Sort‑of‑ARC, though full ARC‑AGI remains modest at 3.0% paper abstract.