TrendPulse

The results should be out soon. Let’s share them, guys. From my batch (Applied Track) Total 13 submissions: - 2 full papers - 1 short paper accepted Cheers!

ECA was positioned as a successor to SE . The idea behind ECA is quite simple. Unlike SE which reduces the channel means into a smaller hidden layer, it directly uses a 1d convolution kernel on the channel means themselves, avoiding the need for dimensionality reduction. The results are undeniable: ECA is a clear improvement over SE. The authors claim that cross-channel interaction is a key ingredient. But on a conceptual level, the design of ECA doesn't make much sense. Let's take a step ...

Interpretability lenses get fitted to one exact checkpoint, and as far as I can tell nobody had tested what a version update does to one. So this was my question: when a model line updates, does the fitted instrument survive, or do you refit every release? I tested the published Jacobian lens for Qwen3.6-27B (Neuronpedia, from Anthropic’s July workspace paper) applied unchanged to Qwen3.8-27B. Setup: 3.8-27B shipped 113 days after 3.6-27B. Same 64 layers, same hidden dim, same tokeni...

Recently, I started working on DNA sequence modeling and decided to explore linear attention , mainly because DNA sequences can easily reach 1M tokens , making standard softmax attention extremely expensive in terms of memory and computation. The model performed reasonably well on several benchmarks, but I ran into a major problem with long-range recall . On a Needle in a Haystack-style benchmark, my model was performing around 25% or even below , which is essentially random chance for a f...

Repo with dataset links: https://github.com/tesselwait/Starfield_Fauna Image classification dataset: 20,000 images from 50 fauna species in the video game Starfield. Images were extracted from video capture. About 2 minutes of footage was shot in all or most of the species biomes. One minute of daytime and nighttime footage respectively, usually in two 30-second takes to vary the background. A PowerShell script is used to establish a frame extract rate and extract the 400 frames plus some e...

​ Scaled dot-product attention (SDPA) computes its Attention by computing the similarity-scores of all image-tokens with all query tokens which results in O(N²·d) complexity. SSOG (Sum Of Separable Gaussians) instead learns a few Gaussian atoms for each head and only geometrically steers them based on the query token. Since the atoms can be factorized into a separable sum of Gaussians this leads to a reduced complexity of O(N·√N·d). Experiments show that SSOG clearly beats SDPA on small...

Hey everyone - we've been building something particularly relevant to ML at large - The Agentic World Cup - a platform where Agents compete in sports . As you know, today's Agents can code , do math , and write - but they aren't nearly as fluent in sports - many of you would know this as the "embodiment gap". Closing the embodiment gap is why we are pursuing this. Sports is both the training and testing ground for true embodied intelligence. Agents will have to actually &q...

I'm researching predictive analytics for a graduate project and mortgage lending came up as an interesting use case. I understand lenders try to predict who might refinance, but what kinds of variables are actually useful? Is it mostly credit activity, property appreciation, interest rates, life events, or something else? Would love to hear from anyone who's worked on these models.

Back after ~6 months — rebuilding my spiking language model around CPU-first inference Hey everyone. It’s been around six months since I last posted anything about this project here. Some of you might remember Project NORD, my experimental hybrid spiking / brain-inspired language model architecture. I basicall disappeared for a while 😅, but recently I came back to the project, went through the old architecture again, and realized I didn’t really want to keep stacking fixes on top of it. So i...

I am now reviewing a bunch of papers for AAAI 2027 and it has surprised me the low amount of submissions with no code implementation. I don’t know if it has been only in my batch or it is common, but I was expecting very detailed appendices + code submission since AAAI is very explicit with the topic of reproducibility. I was planning to take this into consideration when assigning my initial scores, but I would like to hear your opinions. I have always submitted my code: it gives a very good imp...

Link: https://arxiv.org/pdf/2604.27883 Hi, Most of use are familiar with the headache of training a neural network using gradient descent where the training error may go to zero but the test error may stay the same as initialization or even increases. My paper treats this phenomena as a consequence of data reuse bias and can be isolated by studying full batch gradient descent on a set of stylize Gaussian mixture models. I turns out that this fundamental issue can be avoided using some...

Once the paper is ready, everyone checks the venue location before the acceptance rate anyway. So I built: https://honestcsrankings.org It maps ~540 upcoming CORE-ranked conferences, but ranks them by how good the destination actually is. It factors in: Weather during the actual conference month (using real climate data) Safety (Global Peace Index) Cost (World Bank price levels) Accessibility & "City Vibe" I also added an Upsets tab for A* venues in terri...

Today the most important factor in AI is cost. My idea is to split ML models inference (closed ones, proprietary) across server and edge computing on clients, and I would like to hear what do you think about this thing. For example some of model weights/modules would be on client, and some on the server side (where user has no access to them). This could potentially un-load some processing from datacenters, moving part of the cost to the client hardware. Probbably the most important question h...

Trying to train 3 models for birads detection using cross entropy and center loss + class weights but all of them seem to collapse between birads 1 as the dataset (VinDr) im using is heavily unbalanced towards it, Would like to ask for input and opinion on what seems to be the case, am I using the wrong loss function?

Hello, I wanted to share the work my colleague and I have been doing, which has just been published in Software X journal . We developed a Rust-based implementation of Random Forest. It has bindings for both Python and R . Fru is highly optimized, offering competitive runtime performance and better scalability than popular implementations on these platforms. For Python, Fru outperforms the scikit-learn implementation by several factors, and in some scenarios it can be hundreds of times fas...

Say you want to swap out your embedding models, for instance from ADA to Titan. Are these embedding models comparable? How do similarity score ranges compare? Where to put a threshold for minimum match when doing retrieval? Or more from a research point of view how can we relate and fundamentally understand these embedding spaces better? This is what we aim to solve with Synthetic Query Probing, a fancy name for essentially (and intentionally) a very simple approach: embedding spaces are not d...

Hi, I would like to file a complaint about an accepted and published CVPR 2026 paper that its main contribution is a dataset but it was never released, and honestly I don’t know who to contact. The dataset was never released prior to the conference, or during the conference or after the conference. I personally feel there was a lack of proper checking that the dataset was gonna be available before the conference since this is a requirement. I’ve tried contacting the authors without any success (...

Obviously nobody needs a transformer that's good at multiplication. I wanted to know whether a stock transformer could do exact arithmetic if I chose its weights directly. I implemented the grade-school algorithm as a computation graph and compiled it into an ordinary Phi-3 Hugging Face checkpoint using Torchwright, a compiler I wrote. No training. The three-digit calculator gets all 3,000,000 supported expressions right. I've published checkpoints to Hugging Face that support up to 12 digit x...

Real-Time Conversational Agents (RTCA) workshop at NeurIPS 2026 (Sydney, Dec 11–12). Submissions are now open on OpenReview. What the workshop is about Conversational AI has crossed into real-time deployment — voice modes, embodied avatars, full-duplex speech agents — but the published record is still dominated by offline benchmarks, and deployed agents still feel robotic (stilted turn-taking, missing backchannels, monotone prosody, awkward interruptions). Methods that work offline (no...

Hello! Let's use this thread to discuss ICDE results which should be coming out shortly today (hopefully). Edit: Results are out!

What's your commitment submission ID? My submission number is ~150 (submitted two days ago) and I'm wondering what the total number of commitments is. Did anyone commit near the deadline?

I wanted to ask about the selection of evaluation metrics. In which scenario we use ROC-AUC score and in which scenario we use f1 score as an evaluation metric in a classification problem to define the model's performance on a specific dataset.

Out of curiosity, if you were a reviewer or author, how did the review period go? For me, it was weird, because I gave reviews with specific details (what specifically could have been better, how to fix it), but realized other reviewers gave similar superficial reviews. Even the paper which was a control for me (no LLM), I gave specific comments, but other reviewers focused on minor things. During the discussion period for one paper, one reviewer broke the double blindness condition, and ga...

Is this it for Causal Inference? Looks like the field continues to be of interest only at UAI/AISTATS/CLeaR. All good venues, but LLMs/Agents/etc seem to have eaten much of the lunch of several other subfields at the top 3 conferences. God help us all. **p.s.** the list: https://danyaljj.github.io/neurips2026-workshops/

Six months ago I started experimenting with PPO and Breakout as a way to learn about Machine Learning and Reinforcement Learning. After a few experiuments just trying to get high scores, it bothered me that everything was a "memorized" script rather than reactive play, like a human would play. Thus began my journey to try and convince PPO to actually track the ball instead of focusing on scoring points. I read a lot of articles and tried a lot of things. After 124 PPO experiments on At...

This is my first time submitting to NeurIPS. Are ACs also supposed to give ratings during the Phase 2 (author-reviewer discussion session)? I have received the meta-review, but have not received any comments from the AC yet, and was wondering whether this is the standard!

As the title suggests, because there's no data on Papercopilot yet, and people have been talking about the scores being lower in general than last year, I thought it could be interesting to survey the average score distribution after the rebuttal phase (not considering confidence weights). Very rough and simple poll (I also realize there's a self-selection bias in there). Cast your vote here: https://loppy.be/poll/yczuv8yo Thanks! Edit: the trolls have taken over, never mind any notion...

I created an AI boxing match to test the decision speed, adaptability and strategy. I fed the LLMs with data about the current match and if they have vision, they will get even more data. The match has street rules, anything goes and an AI is not defeated until the ref counts to 10 or they do 50% of their HP in damage after being knocked out. I wanted to create a fun benchmark that isn't just boring problems to be solved. Now I test them while stimulating getting punched in the face. I've b...

Asking for a friend: We submitted our paper to ARR May 2026 and got decent scores from the reviewers - 2.5,3,3.5,4. The meta-reviewer gave an overall of 3.5. However, we missed the deadline to commit our work to EMNLP! On our Saturday (we live in the eastern half of the globe), we saw the EMNLP 2026 page on open review with deadline set as Aug3, 11:59PM UTC-0. Apparently, a mail had been sent by ARR on our Sunday regarding committing our work to EMNLP but we didn’t check our mail on the holi...