A 35B Agent Team Still Needs a Hard Stop
Apodex 1.1 shows when a 35B agent team earns its overhead, and when it only creates more plausible debris. Here is the safe local test path.
TOPIC_INDEX
research
47 published entries in this topic.
RESEARCH_ENTRIES
RSS← ALL_TOPICSAlways-On Agents Need a Budget and a Sandbox
Headlong makes AI agents think continuously. Here is the safe way to test one without an uncapped bill, exposed host, or shared-memory surprise.
The $8 AI Paper Pipeline Still Needs a Human
Spark-to-Paper turns a research idea into a checked draft for $8.1, but its gates catch format failures, not scientific truth.
Long videos still drown VLMs in irrelevant frames
ReToken adds one learned token to visual-language models and improves long-context image and video retrieval without retraining the whole backbone.
Production oncall is still a human job at 10% accuracy
ORCA-bench puts coding agents in a noisy six-day microservice testbed. The best hard-task RCA accuracy was just 10%.
32 Hz Robotics Just Lost Its Billion Parameter Habit
TurboVLA removes the LLM from the robot control loop, reaching 32 Hz and 97.7% LIBERO success with 0.2B parameters and 0.9 GB VRAM.
Long context just stopped being a memory problem
HiLS-Attention learns which distant chunks deserve attention, reaching million-token extrapolation without dense attention costs.
Small Models Stop Digging Their Own Reasoning Holes
Relay-OPD lets a teacher briefly rescue a small model when its reasoning goes off course, improving math accuracy while cutting training trajectories by more than half.
Models Still Answer Nonsense. BullshitBench Shows Which Ones Don't
BullshitBench v2 tests whether AI models reject plausible nonsense instead of confidently answering it. The results are uncomfortable.
Agents Still Fail Half Their Real World Jobs
A new 400-task benchmark finds leading agents pass less than half of realistic jobs, with memory, vision, and cross-app coordination still breaking most systems.
Training Agents Still Breaks at the Harness Layer
OpenForgeRL trains agents inside the real harnesses they use in deployment, posting strong tool-use and GUI results while exposing why error recovery remains hard.
Russian prompts expose a hidden coding agent gap
RuBench tests coding agents on fresh repository fixes written in native Russian, then audits what each product actually read and ran.
Deep Research Got Better When AREX Learned to Doubt Itself
AREX improves deep research by verifying partial answers, preserving useful evidence, and sending unresolved claims into a targeted second pass.
Drop 39% of agent tokens by reading the model's hidden state
ByteDance researchers show the keep-or-prune signal for coding agent context is already inside the model's own hidden states. SWE-Pruner Pro prunes tool outputs without a separate scoring model.
Gold at IMO and IPhO with just 2B active parameters
Loopie-20B-A2B, a looped Transformer with only 2B active parameters, won gold at IMO 2025 and IPhO 2025 by reusing layers instead of stacking more, challenging the assumption that frontier reasoning requires massive scale.
An 87-year-old math problem just fell to Claude Fable 5
Claude Fable 5 produced a hand-checkable counterexample to the 87-year-old Jacobian conjecture. The polynomial map fits in one tweet and can be verified in minutes with Wolfram Alpha.
A 10-page prompt closed a 30-year math gap in 148 minutes
A Berkeley professor spent a year on an open optimization problem. GPT-5.6 solved it in 148 minutes with a 10-page prompt. The proof is Lean-verified.
975 billion parameters and none of them are locked
Thinking Machines Lab released Inkling: a 975B open-weight MoE model with Apache 2.0 license, controllable thinking effort, and native multimodality across text, images, and audio.
64 Agents Just Proved What Mathematicians Couldn't in 50 Years
GPT-5.6 Sol Ultra proved a 50-year-old graph theory conjecture using 64 parallel subagents in under an hour. The model became publicly available the same day.
5B parameters play Rocket League without a physics engine
MIRA runs four-player Rocket League at 20 FPS using a 5B diffusion transformer with no physics engine, trained entirely on bot gameplay. Open-source.
Real quantum hardware just recovered a 10-bit encryption key
Researchers recovered a 10-bit crypto key on real IBM quantum hardware, doubling the previous record. Your AES-128 is still safe.
84.7% accuracy at 1/14th cost. Bridgewater just proved fine-tuning beats frontier
Bridgewater fine-tuned Qwen3-235B on expert-labeled financial data and beat every frontier model at 1/14th the cost. The era of renting the biggest generalist may be ending.
When does scaling actually stop working?
Lilian Weng traces the full history of scaling law research, showing where Kaplan and Chinchilla disagree and why the exponents matter more than anyone admits.
A $475M startup just proved oscillators can draw
Unconventional AI released Un-0, an image generator running on coupled oscillators instead of neural networks. FID 6.74 on ImageNet 64x64. The hardware to make it 1000x more efficient does not exist yet.
Open image models finally got a training pipeline that works
Krea 2 drops open weights for a 12B image model with a two-checkpoint workflow: train on Raw, generate fast on Turbo. Zero synthetic training data, 2-second inference.
Simulated agent training now beats the real thing
Alibaba's Qwen-AgentWorld is the first language model built to simulate agent environments, and training in its fake worlds beats training in real ones.
That 3B model matched Claude Opus at math. Benchmarks broken?
A 3B parameter model from Sina Weibo claims to match Claude Opus 4.5 on math benchmarks. The AI community is split on whether this is a breakthrough or benchmark gaming.
Your 11.9B inpainting model just got outperformed by something 50x smaller
A new 0.22B parameter model matches or beats 10B-level inpainting models like FLUX and SD3.5 using less than 2% of the parameters and 15x faster inference.
59% SWE-Bench score from a model costing $0.30 per million tokens
MiniMax M3 is an open-weight model scoring 59% on SWE-Bench Pro at $0.30/M tokens. The architecture is novel, the price is aggressive, and the weights are real. But vendor-run benchmarks and Chinese jurisdiction concerns complicate the story.
Robots need three brains to think and move
Alibaba's Tongyi Lab launched three specialized robotics foundation models that cover navigation, manipulation, and world prediction. The complete stack runs on NVIDIA Jetson Thor and is already in enterprise pilot testing.
The AI Boom Just Outpaced Every Safeguard We Built
Stanford's 2026 AI Index: $581.7B invested, a 2.7% U.S.-China gap, junior dev hiring down 20%, and transparency scores falling. The systems can't keep up.
Your LLM's KV Cache Is Eating 75% of Your GPU Memory
A new lossless compression technique uses a smaller model's predictions to shrink LLM KV caches by 4x on top of existing methods, with zero quality loss and no retraining required.
Anthropic warns AI could soon build itself without human help
Anthropic admitted 80% of its own code is now written by Claude. The company says recursive self-improvement is approaching and wants a global pause.
AI Broke Berkeley CS With 35% Student Failure Rates
Berkeley CS failure rates hit 35% in Spring 2026 as AI cheating and dwindling math skills converge. Professors name LLM overreliance as the primary driver.
MiniMax M3: Open-Weight Model With 1M Context and Frontier Coding
MiniMax M3 is the first open-weight model to combine a 1M-token context window, native multimodality, and frontier coding — beating GPT-5.5 on SWE-Bench Pro while cutting compute by 20x.
Law Professors Picked AI Answers Over Their Own Colleagues
A blind study of nearly 3,000 comparisons found law professors preferred AI-generated answers to student questions 75% of the time — and flagged human answers as harmful 3x more often.
Your LLM works harder after a short nap
A new paper proposes letting LLMs 'sleep' between processing chunks, consolidating context into fast weights and clearing the cache. The results on multi-hop reasoning are 3x better than baseline.
Vision models finally broke free from the token-by-token cage
NVIDIA's LocateAnything uses Parallel Box Decoding to make vision-language models 10x faster at visual grounding while improving accuracy on LVIS and COCO. The 3B model hits 12.7 boxes per second and handles dense detection, GUI, and document grounding from a single checkpoint.
Long Context Still Breaks Memory. Gated DeltaNet-2 Fixes It
Gated DeltaNet-2 gives linear attention separate erase and write gates, targeting the memory interference that breaks long-context retrieval.
Scaling Laws Need Fewer Questions Than Anyone Thought
Item Response Scaling Laws estimate model scaling with about 50 questions per benchmark after calibration. That cuts evaluation cost, not model-training cost.
An 80-Year-Old Math Conjecture Just Fell to AI
An OpenAI model has disproved an 80-year-old mathematical conjecture by Paul Erdős, marking the first time AI has autonomously solved a prominent open problem in mathematics.
A Benchmark Caught AI Faking Answers to Broken Problems
A consortium of 64 mathematicians built Soohak, a 439-problem benchmark revealing that frontier AI models confidently produce wrong answers for unsolvable math problems. No model exceeded 50% on the refusal subset.
CoT Boosts Agent Performance 10x But RL Still Wins Planning
Agentick puts RL agents, LLMs, VLMs, and hybrid systems on equal ground across 37 tasks. GPT-5 mini leads overall but PPO dominates planning. Chain-of-thought multiplies LLM performance by up to 10x.
Your LLM generates one word at a time. This one doesn't.
ByteDance's Cola DLM is a 2.3B parameter language model that ditches autoregressive token prediction for continuous latent diffusion. It beats matched autoregressive baselines on reasoning benchmarks and suggests a path beyond the left-to-right token parade that's dominated NLP for a decade.
ExoActor: When Video Generation Becomes Robot Imagination
BAAI's ExoActor framework uses video generation models as a robot's imagination — generating third-person videos of task execution and translating them into physical humanoid robot behaviors on Unitree G1 hardware.
Tuna-2: Meta's Pixel Embeddings Beat Vision Encoders
Meta's Tuna-2 proves pretrained vision encoders are unnecessary. Direct pixel embeddings achieve SOTA on OCR, counting, and perception benchmarks—no CLIP, no VAE.
Talkie: The 13B Language Model That Thinks It's 1930
Alec Radford and team built a 13B LM trained only on pre-1931 data. It writes Python without ever seeing computers, and enables clean experiments on LLM generalization vs memorization.