From Attacks to Curricula: Learnability-Guided Adversarial Training for Safe Autonomous Driving
Yuewen Mei, Tong Nie, Jie Sun, Haotian Shi, Wei Ma, Jian Sun
arxiv.org/abs/2606.14032 [𝚌𝚜.𝚁𝙾]
ALT Closed-loop adversarial training improves autonomous driving safety by exposing policies to rare safety-critical scenarios. Standard pipelines first generate adversarial scenarios and then sample them for policy optimization. However, most existing frameworks remain attack-oriented: collision-driven generators often synthesize unsolvable extreme situations, which can degrade learning, while heuristic samplers ignore the evolving capability of the driving policy, causing sample inefficiency and delayed convergence. We propose AlignADV, a learnability-guided closed-loop adversarial training framework that converts adversarial scenarios into resolvable and capability-aligned curricula. First, we reformulate adversarial scenario generation as a preference alignment problem and employ direct preference optimization to guide the generator toward critical yet resolvable scenarios. Second, we introduce behavioral fingerprints to capture the intrinsic characteristics of the evolving policy and
47
Causally Evaluating the Learnability of Formal Language Tasks
Vésteinn Snæbjarnarson, Anej Svete, Josef Valvoda, Reda Boumasmoud, Brian DuSell, Ryan Cotterell
arxiv.org/abs/2606.09822 [𝚌𝚜.𝙲𝙻 𝚌𝚜.𝙵𝙻]
60
Jun 13
Maybe not in the infinite-compute limit. But under real constraints, architecture matters.
Even if two systems can eventually reach similar representations, they may differ dramatically in sample efficiency, number of units, energy, robustness, and learnability.
So the point is not that point networks cannot do this. It’s that biology may get some of this computation “inside” a single neuron, which changes the relevant unit of computation.
9
OpenURMA now supports integration with unmodified, official openEuler UMDK / URMA stack. I used Pine Copilot to instruct Claude Code w/ Opus 4.8 which autonomously work for 3 days to complete the full RTL, SystemC two-node, and gem5 full-system simulation with real EulerOS kernel and perftest/KV/RPC applications.
I manage 5~10 agents as digital co-workers, one of them taking the OpenURMA project. It reports overnight progress every morning, then I discuss with it for a while. In the daytime it iterates using test cases to fix every detailed discrepancy from the spec and the official UMDK stack. Every evening I ask the agent to report progress again and discuss about next steps before going to sleep. Then the agent goes into an autonomous loop overnight, and resolved the Gem5 compatibility problem with EulerOS 6.6 kernel fully on its own.
The tasks that agents can do really well are mostly in the learnable regime (the learnability in machine learning theory). If you can work step by step toward easily verifiable objective, then the agent can can climb the mountain autonomously over hundreds or even thousands of steps. This is how OpenURMA was built.
However, what agents do not do well is anything requiring a clean architecture where every single component must be correct for the entire system to function (e.g., our realtime voice agents). You cannot use a rough design and evolve it bit by bit by adding patches.
github.com/bojieli/OpenURMA
1
223
@RichDiviney noted “Learnability” as one of the mental acuity attributes. I think when people use the word intelligence, they normally mean learnability.
People with high learnability who are unhappy need are missing something else.
4
To prepare students for the workforce today, colleges should teach students emotional intelligence, resilience, empathy, integrity, learnability, and leadership skills. s.hbr.org/3R3lOhg
4
26
6,536
Jun 9
Researchers use formal languages to isolate how data frequency affects task learning in language models. Controlled experiments reveal systematic patterns in learnability that are obscured in natural language settings. #ml #ai #nlp #distillai
distillai.ai/p/t1mcuuks9vsj7…
8
Jun 9
Today I chatted with a Business Analyst who’s been job-hunting for 6 months. He hadn’t used AI tools at work in 2024–25 and felt “behind” now that many roles ask for AI skills. I told him: build small personal projects, play with AI tools, and show your curiosity — employers hire learnability.
Do you know someone in the same spot? What helped you catch up?
2
78
shivya bhede retweeted
May 26
Can AI help unlock India’s vast rural talent pool?
@svembu believes geography will matter less when curiosity, learnability and access to technology become the real differentiators.
A conversation on talent, opportunity and India’s future.
20
189
1,326
1,462,259
Tensor Logic provides a framework where neural networks gain the reliability of symbolic AI, and symbolic systems gain the learnability of deep learning. #AI #ArtificialIntelligence #DataScience hubs.li/Q04kp8sJ0
1
298
1
2
25
Jun 6
It is a misperception to think that human-like LLM behavior implies that neural networks possess human-like attributes.
The paper’s Age of Empires II example makes this point sharply: similar input-output behavior can arise from very different computational substrates, so behavior alone does not justify anthropomorphic claims.
From the Deep Manifold view, neural networks are propertyless learnable numerical systems. They do not inherently contain morality, intent, anxiety, understanding, or consciousness; their strength lies in learnability and discoverability, the ability to absorb patterns, form useful pathways, and stabilize behavior under data and boundary conditions.
Human-like behavior is therefore learned manifold alignment, not intrinsic human quality. Everyone is anxious about AI, but as long as we keep the 1st and 2nd Amendments intact, we will be fine as humans.
#DeepManifoldInterpretation
1
16
1,059
Jun 5
🔥 Thrilled to share our new work!
What does it really mean for a model to "think with images"?
1⃣ We found that generated thinking images are often incidental rather than causal, with the model reasoning around them rather than through them.
2⃣ In this work, we study how to make the generated thinking image causally necessary for the model's reasoning. We also ask which kinds of thinking images work best through the lens of informativeness and learnability, and find that panorama comes out on top, with the strongest out-of-distribution generalization.
🧠 New paper: "How and What to Imagine? Visual Thinking in Unified Multimodal Models for Cross-View Spatial Reasoning"
Cross-view spatial reasoning is hard for VLMs. Language-only reasoning loses geometry.🧵👇
3
57
Jun 4
Having full faith in bright future at the depth of adversity and learnability are two most important attitude everyone should have for a good life. Optimism over self pity leads you to keep exploring way to get better and come out bright on the other side. Universe offers everyone a chance. Such stories contextualises these qualities and serves as inspiration!
Million Salute to you Rinki Rathod.
#optimism #positivity
@carnelian_asset
“What did you even bring from your parents’ house?”
Once, these dowry taunts shattered Rinki Rathore’s confidence. Today, she walks the ramp with pride.
Married at 16, she endured years of humiliation, raised her son alone, lived in a rented room, and stitched clothes to survive. But instead of giving up, she chose herself.
Between sewing machines and sleepless nights, she dared to dream of becoming a model. Learning through social media videos, she practised every day until life finally changed.
Today, her story reminds countless women that one brave decision can rewrite an entire future. >>
#WomenEmpowerment #InspiringJourney #SelfBelief #OvercomingAdversity #DreamsToReality
Rinki Rathore Story, Women Empowerment Stories, Overcoming Dowry Challenges, Inspirational Women, Single Mother Success, Twisha Sharma Case]
2
1
48
7,659
On the Learnability of Test-Time Adaptation: A Recovery Complexity Perspective
Zhi Zhou, Ming Yang, Shi-Yu Tian, Kun-Yang Yu, Lan-Zhe Guo, Yu-Feng Li
arxiv.org/abs/2605.28057 [𝚌𝚜.𝙻𝙶 𝚌𝚜.𝙰𝙸]
💬Accepted by ICML 2026
3
19
912
🐾 @saturdayrobotic Robotics & World Models Reading Club 10 Recap: From Platform → Instincts → Real-World Learning: Roadmap to 🐱Cat-Level Humanoid Intelligence
Keynote: Bringing Robots to Life — Learning Humanoid Instincts from the Body Up, by @HaochenShi74 (@Stanford PhD, adv. Karen Liu & Shuran Song), presents a full-stack humanoid loco-manipulation program spanning hardware, learning, and real-world deployment. Hosts @junfanzhu98, @aurorafeng_01.
🤖 Stage 0: ToddlerBot Platform (ML-compatible embodiment)
Open-source humanoid designed for learnability, not just capability. 30 DoF full-body design (arms/legs/torso/head), dual grippers, 2× fisheye cams, IMU, mic/speaker, Jetson Orin NX, 2–5h battery. Spur/bevel/linkage transmissions. Core idea: hardware is sufficient; bottleneck is learning. Key enablers: exact URDF digital twin, zero-point calibration, motor system ID capturing friction/backlash/controller response, and full actuation model (torque–velocity limits). Teleop via joystick VR. Sim2real depends on physics-accurate sysID, not kinematics alone.
🧠 Stage 1: Instincts (survival layer)
Locomotion: keyframes → RL w/ domain randomization → vision skill planner (depth IMU) 3.1Hz. Policy: 3-layer MLP 50Hz (low natural freq).
Motor Current-based Compliance (MCC): no force sensors. External wrench inferred from motor current/voltage Jacobians motor model → spring-damper correction. Works across whole body, any contact. Diffusion policy (200 demos, ~80% success), OCHS servoing (21° vs 2–3°), LEAP hand VLM skills, heart-drawing ablation shows wrench estimator key. Generalizes across robots (Unitree G1 etc.), framing compliance as embodiment-level safety primitive.
Energy autonomy: self-charging docking for continuous operation.
🌍 Stage 2: Real-world learning (RTR)
Robot Trains Robot replaces humans with a robot-arm teacher: reward via F/T sensing, XY compliant support, Z-axis curriculum withdrawal, perturbation, failure detection, and auto-reset. Enables safe real-world RL without humans. Key method: latent dynamics gap z optimized from real rollouts FiLM-conditioned actor/critic. Demonstrated on walking & swing-up from scratch.
📊 Key insight: highest-value data = real exploration (expensive, scarce). Must survive to collect it → instincts are prerequisite for data flywheel.
🧩 Big picture
Platform = learnable embodiment (URDF sysID teleop)
Instincts = survival (locomotion MCC compliance autonomy)
RTR = self-scaling real-world data engine
⚠️ Reality gap: sim2real still hard for manipulation; locomotion works better. No established robotics scaling law; foundation model form remains unclear.
🐱 Summary: Simulation starts robots. Instincts keep them alive. Real-world experience makes them intelligent.
3
7
35
4,254