The paper introduces InfiLoop, a loop-native residual connection designed for looped Transformers to prevent performance degradation as iterations increase. InfiLoop learns which past computations to retain and how much to accept from each new update, combining content-based weighting with learned temporal decay to maintain a running summary of recurrent states. Across extensive reasoning tasks, a 7M-parameter InfiLoop model outperforms existing recursive architectures, achieving 97.9% exact accuracy on Sudoku-Extreme and 13.6% pass@2 on ARC-AGI-2, and continues to improve with test-time looping beyond 20,000 effective steps.
By Pengxiang Li, Dilxat Muhtar, Di He, Guinan Su, Lu Yin, Shiwei Liu
The paper proposes InfiLoop, a loop-native residual connection designed for looped Transformers to prevent performance degradation as iterations increase. InfiLoop learns which past computations to retain and how much to accept from each new update, using content-based weighting and learned temporal decay to maintain a running summary of recurrent states. The approach yields strong results on reasoning tasks, achieving 97.9% exact accuracy on Sudoku-Extreme and 13.6% pass@2 on ARC-AGI-2, and continues to improve with over 20,000 test-time iterations.
arXiv:2607. 17374v1 Announce Type: cross Abstract: Graph Neural Network (GNN) inference on billion-scale graphs is challenging due to the large memory footprint of features and embeddings and high disk I/O costs in out-of-core settings.
By Pranjal Naman, Yogesh Simmhan
The paper investigates how to allocate test‑time computation in recursive reasoning models, specifically the Tiny Recursive Model (TRM). Experiments show that resampling the answer state (y) is more effective than perturbing the reasoning state (z) for recovering failures, especially as compute budgets increase. The proposed Gated Answer Resampling (GAR) method, which restarts only rejected puzzles, boosts TRM performance from 85.9% to 99.4% on Sudoku‑Extreme and from 68.8% to 85.6% on the Pencil Puzzle Bench, outperforming the probabilistic variant that uses far more compute.
By Maxime Illouz
arXiv:2607. 05104v1 Announce Type: cross Abstract: Grokking -- the delayed onset of generalization long after a network has fit its training set - -is usually studied in models too large to read completely and reported from single training runs.
By Yoshiyuki Ootani
arXiv:2609.37066v1 Announce Type: cross
Abstract: Post-training is central to mathematical reasoning in modern large language models (LLMs), but endpoint pass@1 alone underidentifies what has changed...
By Hongyang Li, Yiming Zhu, Xiao Li, Caesar Wu, Said Mammar, Pascal Bouvry
arXiv:2607. 11116v1 Announce Type: cross Abstract: Deep equilibrium models promise input-adaptive implicit computation: harder problems should demand more solver iterations, and the solved equilibrium should encode the result of genuine iterative inference.
By Joyjeet Singh
G-ReAct is a reasoning framework that frames deep search as state evolution over a fixed-topology query graph, enabling explicit tracking of search progress and constraint preservation. It generates high-quality trajectories for fine-tuning and provides structured guidance during inference without extra fine-tuning. Experiments show that with only 1.9K generated trajectories, a Qwen3 model achieves strong accuracy on BrowseComp-ZH and XBench, outperforming larger open-source baselines, and consistently improves existing LLMs on deep-search tasks.
By Shaoxiong Yang, Mengyuan Zhang, Shaojun Lin, Chao Li, Wei Liu, Kun Shao, Jian Luan
The paper argues that pattern recognition and step‑by‑step reasoning lie on a spectrum, with large language models (LLMs) learning the latter when the next token depends on only a few preceding tokens. It formalises reasoning traces as paths on a De Bruijn graph, showing that the number of edges is far smaller than the number of possible traces, making step‑by‑step reasoning sample‑efficient. Experiments fine‑tuning Qwen2.5‑1.5B‑Instruct demonstrate that a moderate density of states balances accuracy and robustness, and that real‑world models like Qwen3 retain most of their performance even when attention is limited to a small sliding window.
By Amrut Nadgir, Pratik Chaudhari, Vijay Balasubramanian
arXiv:2606. 03938v1 Announce Type: cross Abstract: Multi-epoch training is becoming the standard now that compute is growing faster than the supply of high-quality text.
By Bishwas Mandal, Shmuel Berman, Akshay Vegesna, Samip Dahal
The paper introduces a self‑distillation method that leverages answer‑only data to improve large language models (LLMs) after training. By training an inverse reasoning LLM to generate latent chain‑of‑thought traces conditioned on the answer, the authors can perform offline distillation that significantly boosts performance on hard prompts for tasks like Countdown and GSM8K. Additionally, the inverse model serves as an accurate verifier, enhancing best‑of‑n selection accuracy.
By Eric O. Korman
The paper introduces a feedforward graph architecture that uses several frozen large language models as computational nodes connected through a shared continuous latent space via learned linear projections. By jointly optimizing projection matrices through backpropagation, the system combines the representations of three small frozen models with two larger ones, culminating in a lightweight cross‑attention output node. With only 17.6 M trainable parameters, the architecture attains state‑of‑the‑art results on ARC‑Challenge, OpenBookQA, and MMLU, surpassing both individual constituent models and parameter‑matched learned classifiers.
By Marcus Armstrong, Navid Ayoobi, Arjun Mukherjee