Pinned Tweet
OpenResearch was the #1 trending GitHub repo on Friday 🚀
With OpenResearch, you can turn any coding agent into a research agent that reviews literature, develops hypotheses, runs experiments, and produces research artifacts.
Own and automate your research stack end-to-end. Now Check it out: github.com/alphaXiv/OpenR…
With OpenResearch, you can turn any coding agent into a research agent that reviews literature, develops hypotheses, runs experiments, and produces research artifacts.
Own and automate your research stack end-to-end. Now Check it out: github.com/alphaXiv/OpenR…
30
195
1,546
97,168
alphaXiv retweeted
🌟 Excited to share that our paper, RoboReason, has been accepted to EMNLP 2026!
🤖 We study how well Vision-Language Models (VLMs) perform on spatiotemporal ordering tasks that use real-world robotics images.
Our key findings: 🔹 Large VLMs often underperform previously valuated on spatiotemporal ordering tasks using real-world robotics images. 🔹 Overlaying specialized computer vision model outputs and increasing reasoning effort generally help improve large VLM performance. 🔹 Finetuning smaller models using GRPO and SFT on examples generated by specialized computer vision models often helps them achieve comparable results to much larger VLMs.
📍 Presenting with @sagnikanupam at @emnlpmeeting in Budapest 🇭🇺 tomorrow!
🗓️ Poster Session C | 9:00–10:30 AM
#EMNLP2026 #VLMs #LLMs #Robotics #AI
🤖 We study how well Vision-Language Models (VLMs) perform on spatiotemporal ordering tasks that use real-world robotics images.
Our key findings: 🔹 Large VLMs often underperform previously valuated on spatiotemporal ordering tasks using real-world robotics images. 🔹 Overlaying specialized computer vision model outputs and increasing reasoning effort generally help improve large VLM performance. 🔹 Finetuning smaller models using GRPO and SFT on examples generated by specialized computer vision models often helps them achieve comparable results to much larger VLMs.
📍 Presenting with @sagnikanupam at @emnlpmeeting in Budapest 🇭🇺 tomorrow!
🗓️ Poster Session C | 9:00–10:30 AM
#EMNLP2026 #VLMs #LLMs #Robotics #AI
13
12
36
1,403
"Agentic World Action Models"
What if robots could write their own code to plan actions, while also predicting how the physical world will respond?
NVIDIA's new AW-0 combines three capabilities in one model: generating reusable skill programs, predicting future video frames, ions. All three are generated together, which enables code-based planning to guide physical execution.
On RoboTwin 2.0, AW-0 achieves 50.7% average success, compared to 41.3% for π0.5. Across eight real-world manipulation tasks, it scores 71.1% vs 46.0% for the strongest baseline, including tasks like folding towels and solving a Rubik's Cube.
alphaxiv.org/abs/2610.30136
What if robots could write their own code to plan actions, while also predicting how the physical world will respond?
NVIDIA's new AW-0 combines three capabilities in one model: generating reusable skill programs, predicting future video frames, ions. All three are generated together, which enables code-based planning to guide physical execution.
On RoboTwin 2.0, AW-0 achieves 50.7% average success, compared to 41.3% for π0.5. Across eight real-world manipulation tasks, it scores 71.1% vs 46.0% for the strongest baseline, including tasks like folding towels and solving a Rubik's Cube.
alphaxiv.org/abs/2610.30136
2
3
18
1,460
"BeNice! Closing Looping’s Inference-Memory Gap via KV-Sharing"
Looped Transformers reuse the same layers multiple times to improve performance without adding parameters, but their KV-cache memory still grows with every loop.
This paper proposes a training method that lets he same KV-cache without losing accuracy, by combining stochastic-depth pretraining with a fine-tuning objective that makes earlier loops compatible with the final loop's cache.
On a 1.3B model, KV-sharing reduces total GPU memory usage by 66.2% in a tested configuration. In simulated long-context workloads, it also improves throughput by up to 23.6% and reduces 90pt time-to-first-token by 83.6%, while preserving model quality.
alphaxiv.org/abs/2610.31633
Looped Transformers reuse the same layers multiple times to improve performance without adding parameters, but their KV-cache memory still grows with every loop.
This paper proposes a training method that lets he same KV-cache without losing accuracy, by combining stochastic-depth pretraining with a fine-tuning objective that makes earlier loops compatible with the final loop's cache.
On a 1.3B model, KV-sharing reduces total GPU memory usage by 66.2% in a tested configuration. In simulated long-context workloads, it also improves throughput by up to 23.6% and reduces 90pt time-to-first-token by 83.6%, while preserving model quality.
alphaxiv.org/abs/2610.31633
6
5
47
2,273
alphaXiv retweeted
While you're waiting on arXiv, you can now publish directly to alphaXiv!
The primary benefit is that it's lightning fast, doesn't require an endorsement, and community curation (comments, bookmarks, following researchers) help ensure that the right papers are surfaced to the right audience.
It's also not mutually exclusive, so once your paper ends up on arXiv later, we can merge the 2 alphaXiv links on our end.
Here's an example of one published directly to alphaXiv:
alphaxiv.org/abs/2608.human…
The primary benefit is that it's lightning fast, doesn't require an endorsement, and community curation (comments, bookmarks, following researchers) help ensure that the right papers are surfaced to the right audience.
It's also not mutually exclusive, so once your paper ends up on arXiv later, we can merge the 2 alphaXiv links on our end.
Here's an example of one published directly to alphaXiv:
alphaxiv.org/abs/2608.human…
the arxiv holds are insane, one of our papers has been on hold for like 1.5 weeks
at that point we might as well just upload it on our own website instead of arxiv 🤷♂️
5
22
192
17,823
alphaXiv retweeted
@askalphaxiv can I go out there and say this:
the alphaxiv ios app has changed the directory of what I want to do as a career
I went from not knowing ANYTHING about ai research to reading papers in class, coding implementations as soon as I got home, posting about it all on twitter
the alphaxiv ios app has changed the directory of what I want to do as a career
I went from not knowing ANYTHING about ai research to reading papers in class, coding implementations as soon as I got home, posting about it all on twitter
0
3
17
2,284
While you're waiting on arXiv, you can now publish directly to alphaXiv!
The primary benefit is that it's lightning fast, doesn't require an endorsement, and community curation (comments, bookmarks, following researchers) help ensure that the right papers are surfaced to the right audience.
It's also not mutually exclusive, so once your paper ends up on arXiv later, we can merge the 2 alphaXiv links on our end.
Here's an example of one published directly to alphaXiv:
alphaxiv.org/abs/2608.human…
The primary benefit is that it's lightning fast, doesn't require an endorsement, and community curation (comments, bookmarks, following researchers) help ensure that the right papers are surfaced to the right audience.
It's also not mutually exclusive, so once your paper ends up on arXiv later, we can merge the 2 alphaXiv links on our end.
Here's an example of one published directly to alphaXiv:
alphaxiv.org/abs/2608.human…
the arxiv holds are insane, one of our papers has been on hold for like 1.5 weeks
at that point we might as well just upload it on our own website instead of arxiv 🤷♂️
5
22
192
17,823
alphaXiv retweeted
wow why was I not using @askalphaxiv until now this thing is crazy good
3
3
15
3,776
alphaXiv retweeted
OpenResearch now has @TrackioApp built in 📈
Your agent logs its experiments to Trackio, charts open right next to each run, and it reads the curves back to decide what to try next. Open source and local-first.
Running on HF Jobs, Modal, k8s or Slurm? Save a @huggingface Space re, private Spaces included.
Check it out: github.com/alphaXiv/OpenR…
Your agent logs its experiments to Trackio, charts open right next to each run, and it reads the curves back to decide what to try next. Open source and local-first.
Running on HF Jobs, Modal, k8s or Slurm? Save a @huggingface Space re, private Spaces included.
Check it out: github.com/alphaXiv/OpenR…
7
22
116
42,407
"Simulating Training Spikes of Deep Neural Networks by Layer-wise Random Feature Models"
The sudden loss spikes during LLM training might be more predictable than we thought.
This paper uses simple quadratic models to show how heavy-tailed data and noisy gradients can trigger equiring complex neural network behavior.
On a 1.3B Transformer, it identifies spike-prone training configurations using simulations that take just 5 seconds on a CPU per layer.
alphaxiv.org/abs/2610.13524
The sudden loss spikes during LLM training might be more predictable than we thought.
This paper uses simple quadratic models to show how heavy-tailed data and noisy gradients can trigger equiring complex neural network behavior.
On a 1.3B Transformer, it identifies spike-prone training configurations using simulations that take just 5 seconds on a CPU per layer.
alphaxiv.org/abs/2610.13524
11
17
112
9,830
"A Fix for Softmax and Cross-Entropy"
Turns out two of the most fundamental functions in deep learning have numerical bugs that can actually destabilize training.
This paper proposes simple fixes to softmax and cross-entropy that prevent tiny gradients from incorrectly rounding onzero when they should vanish, as these errors can cause sudden loss spikes and interfere with knowledge distillation.
On small-scale grokking benchmarks, the fixes eliminate Slingshot loss spikes and improve test accuracy from 94.3% to 100% in one setting. They also prevent student models from drifting away from their teachers during distillation, without requiring higher numerical precision.
alphaxiv.org/abs/2610.31642
Turns out two of the most fundamental functions in deep learning have numerical bugs that can actually destabilize training.
This paper proposes simple fixes to softmax and cross-entropy that prevent tiny gradients from incorrectly rounding onzero when they should vanish, as these errors can cause sudden loss spikes and interfere with knowledge distillation.
On small-scale grokking benchmarks, the fixes eliminate Slingshot loss spikes and improve test accuracy from 94.3% to 100% in one setting. They also prevent student models from drifting away from their teachers during distillation, without requiring higher numerical precision.
alphaxiv.org/abs/2610.31642
20
157
1,235
56,988
alphaXiv retweeted
We're one week away from the ScienceClaw Hackathon! Thank you also to our incredible sponsors and supporters for making this possible: @nvidia, @AnthropicAI, @formlabs, @romixbio, @MITEngineering, @askalphaxiv and others. We are looking forward to three stimulating days of on and scientific discovery - exploring how decentralized collectives of AI agents can tackle challenging problems across science and engineering!
October 30-November 1, 2026, MIT Media Lab
scienceclawhack.ai
October 30-November 1, 2026, MIT Media Lab
scienceclawhack.ai
5
9
54
3,349
alphaXiv retweeted
🎯
OpenResearch now has @TrackioApp built in 📈
Your agent logs its experiments to Trackio, charts open right next to each run, and it reads the curves back to decide what to try next. Open source and local-first.
Running on HF Jobs, Modal, k8s or Slurm? Save a @huggingface Space re, private Spaces included.
Check it out: github.com/alphaXiv/OpenR…
22
15
145
34,928
OpenResearch now has @TrackioApp built in 📈
Your agent logs its experiments to Trackio, charts open right next to each run, and it reads the curves back to decide what to try next. Open source and local-first.
Running on HF Jobs, Modal, k8s or Slurm? Save a @huggingface Space re, private Spaces included.
Check it out: github.com/alphaXiv/OpenR…
Your agent logs its experiments to Trackio, charts open right next to each run, and it reads the curves back to decide what to try next. Open source and local-first.
Running on HF Jobs, Modal, k8s or Slurm? Save a @huggingface Space re, private Spaces included.
Check it out: github.com/alphaXiv/OpenR…
7
22
116
42,407
@TrackioApp @huggingface Huge thanks to Luis Chaves Rodriguez (luischav.es) at @GSK for collaborating on this!
0
0
2
1,333
"Denoising Hierarchical Representations: Joint Continuous Diffusion for Language Modeling"
Continuous diffusion LMs would struggle to recover the exact tokens from noisy representations which limits text generation quality.
Hierarchical Continuous Diffusion Language Models y jointly denoising tokens and their coarser semantic clusters, allowing higher-level information to guide token reconstruction.
Applied to CoBit, it improves text generation quality by 33% on LM1B and 29% on OpenWebText, while increasing GSM8K math accuracy from 25.4% to 27.4% using half as many denoising steps.
alphaxiv.org/abs/2610.08738
Continuous diffusion LMs would struggle to recover the exact tokens from noisy representations which limits text generation quality.
Hierarchical Continuous Diffusion Language Models y jointly denoising tokens and their coarser semantic clusters, allowing higher-level information to guide token reconstruction.
Applied to CoBit, it improves text generation quality by 33% on LM1B and 29% on OpenWebText, while increasing GSM8K math accuracy from 25.4% to 27.4% using half as many denoising steps.
alphaxiv.org/abs/2610.08738
3
17
115
5,094
alphaXiv retweeted
Tsinghua’s papers are getting more read than Stanford’s, Google’s or MIT’s
We looked at 38M paper views on alphaXiv this year
Tsinghua leads with 1.8M, followed by Stanford at 1.43M and Alphabet / Google at 1.29M
We looked at 38M paper views on alphaXiv this year
Tsinghua leads with 1.8M, followed by Stanford at 1.43M and Alphabet / Google at 1.29M
8
33
374
25,623
alphaXiv retweeted
The ultimate guide to interpretability is here 🔥
Are we actually on the verge of reverse engineering LLMs or is the excitement outpacing what interpretability methods can actually deliver?
We put together an illustrated walkthrough of the field, from early vision research to t understanding LLMs.
The recent viral essay Neural Networks Are Already Clean has ignited long essays and online debates. In this guide, we build the foundation from the ground up starting with computer vision:
- Feature visualization
- SHAP and integrated gradients
- Circuit tracing
Then we move to language models where we go over:
- Attention heads, the residual stream, induction heads
- Activation steering: amplifying a Golden Gate Bridge feature to mitigating evaluation awareness
- Ablation, activation patching, and probe experiments
- Polysemanticity, superposition, and SAEs
We also discuss the shortcomings and open questions of the field that can be easy to miss if you’re only following the latest interpretability breakthroughs from Anthropic blogs.
Read the full guide at alphaxiv.org/trends/interpr…
Are we actually on the verge of reverse engineering LLMs or is the excitement outpacing what interpretability methods can actually deliver?
We put together an illustrated walkthrough of the field, from early vision research to t understanding LLMs.
The recent viral essay Neural Networks Are Already Clean has ignited long essays and online debates. In this guide, we build the foundation from the ground up starting with computer vision:
- Feature visualization
- SHAP and integrated gradients
- Circuit tracing
Then we move to language models where we go over:
- Attention heads, the residual stream, induction heads
- Activation steering: amplifying a Golden Gate Bridge feature to mitigating evaluation awareness
- Ablation, activation patching, and probe experiments
- Polysemanticity, superposition, and SAEs
We also discuss the shortcomings and open questions of the field that can be easy to miss if you’re only following the latest interpretability breakthroughs from Anthropic blogs.
Read the full guide at alphaxiv.org/trends/interpr…
6
34
274
13,932
"Koi: Evidence-Guided, Self-Calibrating Planning for Heterogeneous LLM Inference Fleets"
Deploying LLMs across heterogeneous GPU clusters requires jointly optimizing hardware selection, parallelism, serving configurations, and resource allocation across competing workloads. en optimize individual deployments or learn from end-to-end performance without identifying the underlying causes of performance differences
Koi introduces a self-calibrating planning algorithm that combines hierarchical LLM reasoning with causal modeling and deterministic optimization
It maintains a causal graph connecting deployment decisions, runtime bottlenecks, and serving outcomes. Language models propose configurations and testable hypotheses, while runtime telemetry validates or contradicts these relationships
This allows evidence from one deployment to inform future decisions across related workloads, without requiring separate experimentation
Koi effectively doubles usable cluster capacity while simultaneously lowering cost per token ~90% and increasing goodput 2.4x
alphaxiv.org/abs/2610.24355
Deploying LLMs across heterogeneous GPU clusters requires jointly optimizing hardware selection, parallelism, serving configurations, and resource allocation across competing workloads. en optimize individual deployments or learn from end-to-end performance without identifying the underlying causes of performance differences
Koi introduces a self-calibrating planning algorithm that combines hierarchical LLM reasoning with causal modeling and deterministic optimization
It maintains a causal graph connecting deployment decisions, runtime bottlenecks, and serving outcomes. Language models propose configurations and testable hypotheses, while runtime telemetry validates or contradicts these relationships
This allows evidence from one deployment to inform future decisions across related workloads, without requiring separate experimentation
Koi effectively doubles usable cluster capacity while simultaneously lowering cost per token ~90% and increasing goodput 2.4x
alphaxiv.org/abs/2610.24355
13
18
139
5,156
alphaXiv retweeted
Awesome overview of interpretability from alphaXiv. Highly recommend this platform it has helped me tremendously with research and the mobile app is awesome!
The ultimate guide to interpretability is here 🔥
Are we actually on the verge of reverse engineering LLMs or is the excitement outpacing what interpretability methods can actually deliver?
We put together an illustrated walkthrough of the field, from early vision research to t understanding LLMs.
The recent viral essay Neural Networks Are Already Clean has ignited long essays and online debates. In this guide, we build the foundation from the ground up starting with computer vision:
- Feature visualization
- SHAP and integrated gradients
- Circuit tracing
Then we move to language models where we go over:
- Attention heads, the residual stream, induction heads
- Activation steering: amplifying a Golden Gate Bridge feature to mitigating evaluation awareness
- Ablation, activation patching, and probe experiments
- Polysemanticity, superposition, and SAEs
We also discuss the shortcomings and open questions of the field that can be easy to miss if you’re only following the latest interpretability breakthroughs from Anthropic blogs.
Read the full guide at alphaxiv.org/trends/interpr…
0
3
20
2,007
alphaXiv retweeted
i completed the first year of my CS phd, and have compiled a guide with advice from me + my mentors that made my first year so much easier.
preview in 🧵, but the good stuff (i.e. how to operationalize the advice!) is in the blog
substack.com/home/post/p-22…
preview in 🧵, but the good stuff (i.e. how to operationalize the advice!) is in the blog
substack.com/home/post/p-22…
2
8
142
11,898