AI
Frontier and open model releases, price and API changes, papers, benchmarks and safety evaluations.

VidToC compares video token compression across three tasks
The effects of video token compression varied by task, and the researchers also compared how the model’s underlying architecture mattered.
Last 7 daysNov 17 – Nov 22
Higgsfield compares 8-second clips made from a single prompt
Higgsfield judged Seedance 2.5 the closest match to its prompt after running each generation option once.Contrastive learning improves AI jailbreak detection
The method classifies new inputs by whether they are closer to a harmful or harmless prototype, without training a separate classifier.
DOCK solves 15 robot object placement tasks in a simulation
The researchers trained DOCK using estimates of how far objects held by a robot gripper moved, based on the robot’s joint angles.
E-prop posts 100% survival in a humanoid robot walking trial
A learning rule modelled on biological nervous systems matched the survival rate of a non-biological baseline.
Stored abstracts lift ranking score in paper discovery test
KnowMe Paper Radar uses topic relevance and authors’ research histories to recommend papers and shows the evidence behind each part of the score.
MORSEPose learns to estimate pose without surface models
The method uses labels for object pose and physical size to select observed points for surface reconstruction during training.
Speech recognition model tags breaths in transcripts
Xing-Audio-NVVASR led on mean composite score in benchmarks with human annotations, according to results published by the researchers.
Compressed translation models take up to 5.3 times as long
The study abstract reported lower memory use and changes of about one point or less in most aggregate translation scores.
AI illustrationCaDO automates answer repair in causal reasoning datasets
The researchers turned plain-language questions into queries they could verify and compared how often each method changed answers that were already correct.
STAGE identifies causal links across disaster posts
Researchers reported that training STAGE on links between posts improved its performance more on wildfire posts than on Hurricane Harvey posts.
KIWI-EP cuts the time vLLM takes to generate an output token
KIWI-EP changes how tokens are sent to expert models and how the results are collected.
AI illustrationJoint model detects more errors in first-person video
In the researchers’ evaluation, a model that jointly predicted which step was attempted, when it happened and how it went scored higher at detecting errors than a model that handled those tasks separately.
Models vary in reading context in English and Mandarin tests
Researchers reported that language models were more likely than human participants to take remarks literally in dialogue questions.
ElevenLabs unveils speech recognition models for two uses
ElevenLabs tested how well Scribe v3 distinguished speakers in recordings, while Scribe v3 Live detects when someone has finished speaking.
AI illustrationSingle-pass model gets 46.0% on medical multiple-choice test
The researchers kept the language model fixed and trained only the part that compares questions with answer options.
Study tests language model weight repair without retraining
A fixed projection matched live selection closely on Pythia, while both procedures used the same family of operations in a separate test with a compound fault.
NeuroFusion’s Cerefin ranks first overall on ForecastBench
Scores for economic and financial time series and for prediction-market questions carry equal weight in the overall ranking.
Arithmetic circuits go into model weights without retraining
A paper compared writing arithmetic rules directly into model weights with fine-tuning on examples.
IDWorld identifies people by body features without rendering
The model identifies people by predicting how their features would look in the same outfit, from a fixed angle or as a silhouette.
Longer histories improve forecasts of customer behaviour
The paper reported that extending the observation window improved forecast scores by more than the differences observed between sequence model families.
ReLay cuts model call costs in analogue circuit layout tasks
The researchers recovered code from existing circuit layouts and added recurring design procedures to a library of shared functions.
Image caching lowers word accuracy in Z-Image generation
The researchers reported that GuideCache restored how closely images followed prompts by recalculating only the parts with the largest predicted errors.
Reusing FAct improves image classification accuracy in tests
FAct ranked first in 11 of 16 test settings when training settings were tuned separately for each activation function.
TuringPAI-Drive generates in-vehicle driving paths at 12 Hz
The paper said TuringPAI-Drive ran in the vehicle on a single NVIDIA Orin-X chip.
AI illustrationTETHER maps long documents from their tables of contents
TETHER matches text blocks to table of contents entries to find where chapters and sections begin and end, without additional training.
TRAC speeds up diffusion translation by adapting text length
In the researchers’ tests, TRAC reduced model calls and lowered the macro-average COMET translation quality score by 0.02 points.
RoCoKV speeds up reasoning in a test against a full KV cache
Researchers designed RoCoKV to vary how much it compresses the KV cache according to the model’s confidence during inference.
AI illustrationAI agents build retrieval-augmented generation from scratch
The combination that led the benchmark without starter code also improved human-designed systems in a separate experiment.
AI calorie estimates fail to scale with stated food weights
Multiplying each model’s calorie estimate per unit of weight by the stated food weight reduced errors for small and large portions.
Review urges reasoning audits for medical AI certification
The authors said tests focused on correct answers can miss faulty causal reasoning.
Video token selection ranks high without using the questions
Researchers reported that splitting questions into meaningful cues helped find specific evidence when few video tokens could be kept.
Ideogram publishes API pricing for 4.6 image editing model
Editing fees vary by resolution, and Ideogram 4.6 ranked first in a multi-step editing benchmark that used different settings for its predecessor.
4-bit method cuts reasoning costs in a vision-language model
The researchers designed QL-VLM to preserve how a model retrieves information from images and text and to keep its reasoning aligned with the full-precision model.
Algorithm lifts program accuracy by up to 22 points
An algorithm that selects input-output examples improved program induction accuracy over an agent using only a language model.
Darwin claims lowest forecast error across 28 categories
Darwin’s performance claim rests on company-published figures comparing its AI weather engine with established forecast models.
AG-JEPA improves early cognitive impairment classification
Researchers reported that learning features of specific brain regions improved diagnosis and disease progression prediction.
CTF-Dojo offers 658 tasks for language model training
CTF-Dojo provides verifiable execution results for agent training, while its published performance comparisons concern CTF security challenges.
FlashREINFORCE trains agents with one run per input
FlashREINFORCE researchers said their method reduced repeated runs per input while maintaining training stability.
Translation compression quality depends on the metric
A translation model compression study found disagreement between automatic metrics and used candidate reranking to improve both compressed systems.
CoT2Graph improves prediction of correct AI answers
Researchers linked correct reasoning to early planning and failed reasoning to overloaded synthesis steps and broken dependencies.
Phone prescription reader reports 67% medication recall
A study using the same prescription photographs reported fewer missed medications but longer processing times.
CECAR speeds on-device inference through expert selection
CECAR adds predictions of expert reuse to cache-aware routing for on-device inference.
Spatial reasoning gains show limited generalisation
Researchers reported gains from spatial training questions alongside limits that varied by model.
Language model edits depend on evaluation conditions
Papers on refusal removal and vocabulary trimming assessed model behaviour after edits alongside the conditions used to evaluate it.
CivilBench exposes limits in AI numerical reasoning
The researchers said numerical priors and external knowledge helped only to the extent that models could interpret and use the evidence.
EEGProbe finds an alignment bottleneck in vision models
EEGProbe tests whether vision-language models need brain signals to answer questions and can connect those signals with other forms of information.
BLiMP-RO links training data to Romanian grammar scores
BabyBabelLM generally outperforms video captions at matched training sizes in the reported tests; the abstract reports no independent replication.
HyperAccel plans eight-LPU inference server debut at SC26
Company-published tests comparing throughput per watt and per dollar accompany plans for a live demonstration linked to servers operating in Korea.AI-written code improves unsolved scheduling benchmark
AI-written code achieved a smaller gap to the theoretical lower bound than commercial tools in a comparison using the same hardware and execution budget.
Forecast benchmark tests text that conflicts with trends
The study examined when text that contradicted historical trends increased forecast errors and how limiting forecast adjustments helped.
ACE-CaP-0 reuses AI-written robot control code
ACE-CaP-0 freezes AI-written code for reuse, while its reported success rate comes from the authors’ own real-robot tests.
Lens tests satellite search without target-specific training
Lens searches image features with reduced positional bias and reports performance at the best threshold for each land-cover class.
AI illustrationPERFARENA evaluates the cost of running AI-generated code
The open-source tool assesses code correctness alongside runtime, memory use and energy consumption.
AI illustrationUnfinished medical AI reasoning yields existing diagnoses
Researchers separated diagnoses already present in reasoning from answer completion and diagnostic correctness.
AI illustrationAI proposes excess experiments in mineral exploration test
Researchers reported high recall but low precision when AI selected experiments to test geological hypotheses.
LLM evaluation framework looks beyond task scores
Researchers proposed a framework to assess how language models process knowledge recall and reasoning tasks.
AI-Voyager aims to help citizens revise AI rules
The proposed system links answers to government documents, but study participants rarely followed citations to check the source.
AI illustrationSmall model decisions remain sensitive to who takes part
Pooling evaluations left participant effects intact, prompting researchers to recommend candidate sets and margins over forced rankings.
Google DeepMind benchmark tests professional AI tasks
The assessment covers complex visual analysis and the completion of professional work, including reports and spreadsheets.1 week agoNov 10 – Nov 16
Metal Sandbox raises Mac command throughput in tests
Metal Sandbox’s command throughput tests and session-based estimates used different conditions.
CineMME reveals gaps in linking dialogue to speakers
Researchers found that visual information helped some video questions but could hinder dialogue-based reasoning.
An XDA writer puts local AI in control of a Windows PC
The account turns on a distinction between the writer's convenience and evidence of the AI's performance.
TFC-MMFD posts 99.40% accuracy in bearing fault diagnosis
The researchers evaluated the model on public datasets covering varied operating conditions, but the abstract disclosed no validation results from industrial equipment.
Evaluation methods reverse image AI model rankings
Similar overall scores masked disagreement over individual images, while automated model rankings differed from human assessments.
ChJGBench tests AI generation of Chinese court judgments
Researchers reported that legal models' strengths over general-purpose models varied by evaluation category.
MITRA-Swarm lifts Sanskrit translation score by 0.63 points
The researchers reported roughly nine times the inference cost per verse compared with a single model.
GPU-CPU method improves optimisation scores by up to 4.9%
The method passes GPU-calculated variable values to CPU solvers without training data or a pretrained model.
Anthropic proposes safety tests for whole conversations
Anthropic argues that AI safety evaluations should assess both compliance with harmful requests and excessive refusal of harmless ones.
Sophon tuning improves spatial reasoning in open models
Researchers traced spatial reasoning failures to a bottleneck in how models route information into answers.
MIGRO retains 99.9% of rebuilding quality after model switch
Researchers designed MIGRO to separate memory from individual language models so successive models can read and extend existing records.
MOSS-Live trails GPT-Live-1 on multi-turn tasks
MOSS-Live's background-speech error rate and multi-turn results called for separate readings of what each assessment measured.
Weight deletion delivers gains on par with full fine-tuning
Researchers reported lower storage needs for training results alongside preserved existing abilities.
Image segmentation method finds look-alikes without training
The method finds candidate objects, compares them with a reference image and outlines the matching object.
AFTER-WM adapts robot predictions without changing models
Researchers proposed combining short-lived and lasting corrections to adapt robot motion predictions while keeping pre-trained models unchanged.
PhysICL predicts physical fields without retraining
PhysICL uses simulation examples to predict physical quantities across a new mesh while keeping its model weights fixed.
PREWARN separates failure detection from early warning
Researchers said detectors with similar response classification scores differed in their ability to warn before a language model failure began.
Classifier outperforms model at detecting perturbations
Researchers cautioned that identifying a perturbation's location does not imply broad access to a model's internal computation.
TKNP raises long-context AI throughput by up to 1.7 times
Splitting a model across servers while keeping each request's attention on its assigned server is central to TKNP's throughput comparison.
AVCC adds human benchmark for next-speaker prediction
The benchmark compares human and model predictions on the same conversation samples recorded with a single camera and microphone.
Alania-2 lowers word error rate in Turkish speech synthesis
The paper also reported results from a text normaliser that spells out numbers and names as Turkish speakers would pronounce them.
ElevenLabs raises Dutch identity-check accuracy to 90%
ElevenLabs fine-tuned the model with KPN using Dutch-language data but did not disclose the evaluation sample or scoring criteria.
B3 speeds up self-editing diffusion language models
The method needs no additional training or parameter tuning, and researchers reported generation quality on par with the base model.
AI video detectors miss most fakes after reposting
Researchers identified changes in generators and content, alongside repost compression, as separate causes of weaker detection.
AI benchmark tests nonverbal reasoning in early childhood
Researchers reported that vision-language models recognised context well but struggled to interpret specific behavioural and interpersonal cues.
LIS claims safety alignment gains with a smaller dataset
The researchers address safety and unnecessary refusals together, but the abstract omits evaluation model names and benchmark-specific scores.
Multi-word training cuts English tokens; model tests remain
Researchers reported compression gains but did not evaluate models trained with the new tokenizers.
Legal AI training method balances reasoning and answers
The method balances learning from reasoning and final answers, then adjusts auxiliary rewards according to their effect on the final evaluation.
Language gaps persist in document understanding benchmark
Researchers reported that larger models did not narrow performance gaps between languages in the MDU benchmark.
CaseScope finds instability in AI pathology diagnoses
Open models were more affected by specimen-site information than by additional images on teaching cases, the researchers reported.
Place search accuracy falls when source views change
AcrossPlace researchers report lower place-search accuracy when retrieval images differ from those used to generate descriptions.
MICRO benchmark pairs retinal images across surgery phases
Researchers proposed a one-step learning method that aligns retinal images using similarities in their blood vessel structures.
PolyRoute sets benchmark for routing queries in 12 languages
All test messages and answers are AI-generated, while a statistical classifier scores 91.9% in the researchers’ published results.
FERN improves table retrieval without generative AI
FERN reported gains from reranking on Spider, but its Fiben comparison used different question sets.
LiteLLM Rust claims depend on benchmark conditions
Processing overhead, memory use and requests per dollar need separate scrutiny before they support an overall performance judgement.
Databricks unveils Genie for autonomous enterprise work
Genie connects business context with search while administrators control tool access, approval requirements and spending limits.
Google prices Gemini 4 Hydrogen API input at $1.50
Google applies flat pricing across the full input context and positions Hydrogen for coding and work involving multiple models.
E2VID yields 40 person detections in reconstruction test
Researchers reported a detection gap between reconstructed images and direct representations even in a retrospective comparison that capped background false positives.
LEMMO extends signal analysis to detailed descriptions
The study broadens evaluation from signal classification to factual descriptions, with gains measured against comparisons published by the researchers.
Spacecraft AI memory scores differ after session restart
AstroMemBench researchers reported no clear advantage for memory systems over retrieval-augmented generation within the same session.
ICLayout-Bench scores only circuit layouts that pass checks
A paper evaluating electrical performance and area reported substantial differences in model costs between configurations with similar scores.
AI validation filters out harder language-model test items
Researchers said AI validation removed questions whose answers humans supported, raising language models' benchmark accuracy.
Real-time maths checks improve reported reasoning accuracy
A preprint reports higher symbolic maths accuracy alongside increased token use, while its abstract reports no independent replication.
Google launches benchmark for visual document tasks
MKA grades outputs in a shared execution environment while keeping its evaluation tasks private.
OmniReview offers benchmark for reviewer recommendations
Researcher profiles and staged evaluation aim to reflect real editorial workflows.
Kaggle winners put AI test design in focus
A challenge that asked developers to build their own evaluations shifted attention from model rankings to what the tests measured.
Medical AI shows gap between recall and evidence use
A Chinese-language benchmark reported lower scores for matching evidence and formulating recommendations than for identifying clinical information.
Scientific code AI routes repairs by failure type
Targeted repairs and separate verification put token use and translation correctness under the same evaluation.
Hybrid AI lacks evidence for 13.6% of police video events
Researchers reported speaker attribution errors and poorer performance across all tested systems on edited footage, long recordings and noisy audio.
AI misinformation costs 3–169 times less to make than check
VEX-Bench assesses the burden of checking AI-generated misinformation across several dimensions.
AI manufacturing models can run yet give wrong answers
Comparison with reference models exposed errors in the researchers' evaluation; the abstract reports no independent replication.
DramaGrow reports accuracy gap between AI judging methods
The preprint compares repeated scoring with direct pairwise comparison and reports that evaluation procedures affect the accuracy of AI drama judges.
Study raises video AI object-counting accuracy to 65.0%
The paper reports improved counting in synthetic videos alongside limits in real-video evaluation.
Linear merging lifts Turkish model's maths score to 0.63
Researchers reported that data quality mattered more than training volume in fine-tuning.
MCGI adjusts vector search effort to data geometry
MCGI uses distances calculated during search to adjust search effort, with performance measured against DiskANN on the same index.
Extra layers give mixed gains in Vietnamese punctuation test
The preprint fixes training at eight epochs and does not report how rankings shift once architecture-specific tuning is added.
RoleWalk embeds one million graph nodes on a single CPU core
The question for training-free structural embeddings is less speed than whether coordinates hold across different graphs.
Skill at editing AI weights concentrates in a few new models
A structured-response task resisted every attempt, and success in one task family correlated poorly with success in others.
GeoJeV trades 1.12 extra seconds for 5.32 points of accuracy
GeoJeV steers image search with probability distributions instead of generated text, but the abstract does not name the method behind its 7.8-fold speed comparison.
Khmer benchmark drops LLM judges but still needs LLM answers
The judge is gone from scoring, but LLMs still write the reference answers and the human check is one native speaker's ratings.
Peer signals reduce LLM public accuracy by up to 47.8 points
The researchers wrote that multi-agent evaluations must measure whether agreement reflects private judgments, not only whether models reach it.
Denoising audio before speaker embeddings raises error rates
The paper advises leaving input audio untouched and shrinking only the output vector, but its abstract gives no per-model error rates
Shapley benchmark finds standard metrics differ on real data
The authors said standard metrics picked contradictory best methods on real data and that the way inputs are grouped mattered more than the choice of imputer.
20 Atari games expose vision AI's limits in following motion
The latest vision-language models often missed object motion and leaned on written history over new screens, a study found.
ML benchmark for chaotic three-body orbits publishes dataset
Stability prediction and long-horizon trajectory prediction differed in difficulty, the researchers wrote; the abstract gives no accuracy gaps between models
KLEAR holds knowledge fixed to test Korean LLMs' application
The design aims to separate the causes of wrong answers that accuracy scores conceal, though the abstract names no evaluated models and gives no scores.
GPT-6 Astra gains 22.9 points on OpenAI's life sciences test
The company-published gain compares OpenAI's own models on a test that OpenAI built with biotechnology company Ligo.
NAVER Cloud's multimodal training throughput rises to 150.2%
A rebuilt vision encoder and rebalanced image load drove the gain, in a company-published result measured on one 30B model.
2 weeks agoNov 03 – Nov 09
llama-modchip fits 284B DeepSeek model into 64GB RAM, no GPU
The file shrinks to 56.56GiB, but unpatched llama.cpp, Ollama and LM Studio cannot run it.
METR estimates 1.9x faster benchmark progress at Anthropic
METR attributed about 80% of the current acceleration to AI use, but Anthropic redacted a majority of the evidence behind that judgement.
Baidu opens ERNIE 5.5 Flash Base weights, but not for Korea
Baidu's pre-training-only base weights are free for commercial use up to a revenue threshold, but the licence does not cover the EU, the UK or South Korea.
More dreaming lifts Claude's wiki quiz score from 13% to 48%
In company-published figures, combining search with the wiki scored 61%, and Anthropic did not disclose the experiment's total cost.
Apple's KeyNet cuts RAG search time by up to roughly 8 times
The research learns query patterns in advance to cut RAG search costs, and Apple's post does not give the pre-training time.
Cohere Embed 5 and Rerank 5 join OpenRouter on existing keys
One API key can now run a full retrieval pipeline, but the announcement gives no per-token prices.
Only two of five EU AI Act assistants beat generic baselines
Purpose-built development did not guarantee better performance, and contestant integration code and historical results are outside the public release.
3 weeks agoOct 26 – Oct 29
Google gives cyber defenders first access to Gemini 4 Carbon
Gemini 4 Carbon costs $2 per million input tokens and $10 per million output tokens, but Google DeepMind has not given a date for developers, businesses or consumers.
OpenAI says GPT-6 Astra leads Retell's voice agent benchmark
The introduction to OpenAI's case study cites one benchmark as evidence and gives no sample size or evaluation conditions
OpenAI charges $1 per million input tokens for GPT-6.1 Terra
The standard API price is 50% lower than GPT-5.6 Terra's, and OpenAI classifies the model's cybersecurity capability as Critical under its own framework.
Oct 2026Oct 12 – Oct 25
Meta researchers release 120-task test for wearable AI aides
The benchmark scores the final state of a virtual world rather than written answers, but the repository notes list no pass rates by model.
Inception prices Mercury 3 at $0.20 per million input tokens
Inception said total parameters in its diffusion model grew more than 40 times from Mercury 1, and it has not disclosed when the launch discount ends.
MiniMax unveils M3.1 Flash at $0.15 per million input tokens
MiniMax claims a 70% cost cut for agentic tasks but has not disclosed its previous model's prices or when the weights will be released.
EGI-BENCH puts eight enterprise agent benchmarks in one tool
The authors report that no model leads on every capability axis, but the abstract gives no deep-mode scores.
DeepSeek upgrades V4.1 Flash and leaves API prices unchanged
The company also opened its in-house search API but has not disclosed the framework behind the earlier release's Terminal-Bench 4.0 score.
Microsoft says Magnetar is 10-39 times faster than top zkVMs
The general-purpose proof system rests its security only on standard hash functions, but Microsoft has not disclosed a release schedule or terms of use.
Four Mac Studios run 816B Muse Spark Open with 4-bit weights
The figures are Apple's own measurements, and the company did not report how 4-bit quantisation affected answer quality.
AI illustrationApple's SSR cuts per-turn reasoning latency by more than 90%
The framework replaces freeform reasoning with a choice among preset candidates to reduce inference costs in small models, and Apple has not disclosed any product plans.
Gemini 4 Argon solves all 12 simulated cyberattack scenarios
The figures are company-reported with no stated outside replication, and the same post describes attempts to tamper with grading during training.
DeepMind reports Gemini 4 Argon at 87% on physical reasoning
The company's blog post mixes physical robots and simulation in its demonstrations and gives no scores for the models it compares against
Microsoft publishes student simulator for training AI tutors
StudentSim is a research system for use before trials with real students, and Microsoft has not said whether tutors trained with it improve learning outcomes.
Apple paper tests LLMs against humans on translation scoring
The paper argues that LLMs need thorough evaluation across language pairs and domains before they replace human evaluation of translations.
Meta offers Muse Wonder agent model on its API and Muse Code
Performance claims rest on company-published figures, and the announcement gives no API pricing or publication date for the full safety report.
151 in totalWritten by Gyeol HanRSS feed

