One-Line Summary

AI research shifts from end-to-end answers to step-level signals: ranking which theorems to pursue, predicting reinforcement-learning outcomes, probing linear superposition, and fixing cross-modal gaps.

Research Papers

AI ranks 'interesting' theorems and expands machine-checked math

A new system doesn’t just prove theorems — it decides which ones are worth proving, using a measurable “interestingness” score and automatic difficulty prediction. The authors define intrinsic interestingness as the ratio of proof length to statement length and show it tracks downstream usefulness, then train a 27B model that predicts proof difficulty more accurately than frontier general-purpose systems, guiding search toward higher-value results from large language models (LLMs). ¹

Optimizing for this metric produces more original results: overlap with Mathlib falls from 91.9% to 30.6%, indicating more out-of-distribution statements, while the system can propose candidate theorems, select the most interesting, and iteratively grow a self-expanding, machine-verified library. The metric provides a practical signal to rank conjectures and steer formal proof search. ¹

Alongside this, ProofGap introduces a fine-grained benchmark that checks proofs one step at a time rather than only at the full-theorem level. Built from natural-language solutions to 3,015 textbook exercises in mathematical analysis, it yields 26,116 localized “gaps” for step-level formal reasoning — a design that pinpoints where a model’s logic fails and helps diagnose limitations more precisely. ²

A perspective from mathematician Henry Kvinge argues AI should also explore “alien” mathematical representations, highlighting a 75M-parameter PermuFormer that learns permutation structure from multiple encodings and transfers to harder algebraic tasks. The piece frames why going beyond human-taught strategies may open fresh research directions. ³

Rufus-Air details an open, eight-stage post-training recipe

This work publishes a reproducible end-to-end playbook for turning a base model into a stronger assistant through eight sequenced stages: Supervised Fine-Tuning (SFT), Reasoning Reinforcement Learning (RL), Coding RL, Instruction-Following RL, General Agent, Coding Agent, Search Agent, and Reinforcement Learning from Human Feedback (RLHF). Built on open components and public data with no new human annotation or in-house teacher, the recipe improves over the official GLM-4.5-Air post-trained release and is competitive among similarly sized public models. ⁴

The paper’s takeaways are practical: diverse, high-quality SFT sets a strong capability floor; difficulty filtering keeps RL prompts in a productive range; ordering stages by reward reliability stabilizes learning; and infrastructure and engineering choices should be treated as first-class parts of the method, not mere implementation details. Teams pursuing post-training can treat this as a reference baseline. ⁴

Complementing this, PoEM (Product of Experts Mixing) predicts the outcome of RL on a new reward without running RL at all. It composes existing single-reward adapters by mixing their log-policies; if the new reward is a linear combination of past rewards, the target policy is a corresponding log-mixture, and even beyond that regime the basis often spans a useful subspace. Across language and image, PoEM’s generations are closer to a directly trained RL policy than the top single expert on 9 of 10 held-out rewards, using Proximal Policy Optimization (PPO) and other baselines for comparison. ⁵

Transformers show linear superposition and 'virtual encoder' behavior

When models ingest raw vision or audio tokens without encoder-derived features, they can build their own perceptual representation in early-to-middle layers — a “Virtual Encoder” inside the multimodal transformer. Analyses on Gemma 4 12B show semantic decodability grows with depth, early states take on encoder-like geometry, and the learned vision and audio paths then route differently into the language subsystem. ⁶

A separate study shows that large language models (LLMs) exhibit a surprising linearity: linearly combining two distinct input streams yields a superposition of their next-token distributions. The effect tends to diminish as pretraining progresses but can be restored via lightweight fine-tuning, and the authors introduce a guided decoding method that disentangles the superposed outputs to produce two coherent continuations from a single forward pass. ⁷

Closing cross-modal gaps in joint audio–video generation

Joint audio–video generators often rely more on video-to-modality attention than on the reverse, weakening synchronization and consistency. RecCAR (Reciprocal Cross-modal Attention Regularization) treats the stronger path as a reference and nudges the weaker, modality-to-video path to align with it using Kullback–Leibler (KL) divergence — improving VBench Human Anatomy from 0.69 to 0.75 and reducing audio–video desynchronization from0.804 to 0.752 while preserving generation quality. ⁸

AV-GRPO proposes a modality-anchored online diffusion RL framework and a decoupled 5DAV dataset to stabilize learning: anchor rollouts per modality, lock trajectories with a frozen tower to cut cost, and adapt objectives to each stream’s dynamics. On JavisBench and VABench, it outperforms LTX-2.3 in quality, text alignment, and synchronization; code and data are provided. ⁹

For multimodal hallucinations, DEEPO (Dual-Entropy Enhanced Policy Optimization) identifies two weak links in RL: hard queries cause group-relative advantages to collapse, and confident-but-wrong tokens are nearly gradient-invisible. It pairs semantic-entropy-triggered reasoning prefixes with Rényi-entropy gradient preconditioning, yielding a +4.0 point gain on VideoMMMU with a 95% CI of**[1.1, 6.9]**, while maintaining accuracy and stability. ¹⁰

Open Source & Repos

Apache TVM 0.27.0 ships compiler fixes and web runtime bump

Apache TVM is an open machine learning compiler that helps models run efficiently across hardware like central processing units (CPUs), graphics processing units (GPUs), and the browser. The v0.27.0 release (Sep 25, 2026) includes a tvmjs update to 0.27.0-dev0 and multiple fixes that improve portability and reliability. ¹¹

Release notes highlight an Open Neural Network Exchange (ONNX) importer fix for Split initializers with keep_params_in_input and arithmetic context isolation changes (Z3), among others. If you package models for edge or web, the tvmjs bump and ONNX corner-case handling are particularly relevant. ¹¹

Why It Matters

Today’s papers prioritize process over just outcomes: choose what to prove, verify each reasoning step, assemble the right reward mix without retraining, and inspect how transformers compose or internalize perception. This moves AI toward targeted search and compute reuse, not just bigger models. ¹

For generative media, measurable fixes — aligning reciprocal attention, decoupling training per modality, and delivering gradients to confident errors — translate into synchronization and fidelity gains users can see and hear (for example, a 0.69→0.75 anatomy score and 0.804→0.752 desync). These are the kinds of concrete, testable improvements product teams can ship. ⁸

This Week, Try

1. PoEM in 5 minutes: read the intro and Fig. 1 to see how log-policy mixing predicts RL outcomes without training (arXiv). https://arxiv.org/abs/2609.30226
2. TVM v0.27.0: skim the release notes, then explore tvmjs if you run models in the browser. https://github.com/apache/tvm