Sources

Pythagoras-Prover: Advancing Efficient Formal Proving via Augmented Lean Formalisation huggingface.co

Pythagoras-Prover introduces compute-efficient Lean theorem provers using curriculum training and augmented formalization techniques to overcome limitations of scarce verified data and expensive proof search.

No Hidden Prompts Needed! You Can Game AI Peer Review with Presentation-Only Revisions huggingface.co

AI reviewers are vulnerable to presentation-level manipulation that exploits their tendency to be impressed by highlighted strengths rather than convinced by resolved weaknesses, creating a new optimization surface for adversarial repackaging attacks.

The Hidden Power of Scaling Factor in LoRA Optimization huggingface.co

Low-Rank Adaptation (LoRA) scaling factor α functions as a primary optimization driver rather than a secondary learning rate complement, with theoretical and empirical analysis revealing its superior impact on convergence and optimal scaling behavior.

An Enigma of Artificial Reason: Investigating the Production-Evaluation Gap in Large Reasoning Models huggingface.co

Large reasoning models show a sharp gap between generating chains of thought and judging them, with linear probes and causal patching tracing the failure to answer confirmation bias. Once a model commits to an answer, it loses the ability to assess whether its own reasoning supports it.

StepPO: Step-Aligned Policy Optimization for Agentic Reinforcement Learning huggingface.co

Agentic RL typically optimizes at the token level, blurring credit across long trajectories. StepPO instead treats each agent-environment step as the Markov unit for policy optimization, beating token-centric baselines on multi-turn interaction tasks where decisions span many tokens.

APPO: Agentic Procedural Policy Optimization huggingface.co

APPO targets the moments where multi-turn agents actually choose between tools, scoring branches by token uncertainty and policy-induced likelihood gains. Procedure-level advantage scaling then routes credit to those decision points, improving tool-use performance and giving interpretable behavioral traces.

Quickest Detection of Hallucination Onset: Delay Bounds and Learned CUSUM Statistics huggingface.co

Token-level hallucination spotting is reframed as classical change-point detection, yielding a Lorden-style lower bound on detection delay versus false-alarm rate. A learned causal recurrent CUSUM statistic approaches that bound, detecting hallucination onset faster than prior token-level classifiers.

Hy-Embodied-0.5-VLA: From Vision-Language-Action Models to a Real-World Robot Learning Stack huggingface.co

Hy-Embodied-0.5-VLA bundles data collection, model design, continued pre-training, supervised fine-tuning and RL post-training into one end-to-end pipeline for vision-language-action models, aimed at moving VLAs from benchmark demos to real-world robot deployment.

HarnessX: A Composable, Adaptive, and Evolvable Agent Harness Foundry huggingface.co

HarnessX decomposes agent runtimes — prompts, tools, memory, control flow — into composable primitives with a substitution algebra. Its AEGIS engine evolves harnesses from execution traces and feeds results back into model training, coupling harness design and RL in a shared loop.

Skip a Layer or Loop It? Learning Program-of-Layers in LLMs huggingface.co

Instead of running every layer once, PoLar learns a program-of-layers that skips or repeats pretrained layers dynamically at inference. The flexible execution improves accuracy while cutting compute compared with fixed-depth forward passes, exposing latent reasoning capacity in existing weights.

The Arbiter Agent: Continually Monitoring Multi-Agent Conversations to Detect Emergent Misalignment huggingface.co

A multi-agent system monitoring framework identifies misaligned behavior through real-time inspection with resource constraints, demonstrating effective detection of misalignment types under various conditions.

Smaller Models are Natural Explorers for Policy-Level Diversity in GRPO huggingface.co

Small-to-Large Policy Optimization framework uses smaller models as natural explorers to enhance policy diversity and improve large language model training efficiency.

When is Your LLM Steerable? huggingface.co

Activation steering effectiveness can be predicted from early decoding states using a GBDT classifier, enabling efficient steering strength optimization with reduced computational cost.

Statistically Reliable LLM-Based Ranking Evaluation via Prediction-Powered Inference huggingface.co

PRECISE extends prediction-powered inference to correct bias in ranking metrics by combining human labels with LLM judgments, achieving reduced standard error and accurate variant ranking in production settings.

Dense Supervision, Sparse Updates: On the Sparsity and Geometry of On-Policy Distillation huggingface.co

On-policy distillation exhibits sparse parameter updates that are distributed across layers and favor FFN components, while maintaining geometric properties distinct from standard dense parameter rewriting.

Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents huggingface.co

MRAgent combines associative memory graphs with active reconstruction to enable dynamic memory access during reasoning, improving long-horizon memory reasoning while reducing computational costs.

Orchestra-o1: Omnimodal Agent Orchestration huggingface.co

An omnimodal agent orchestration framework is presented that enables efficient collaboration across multiple modalities through unified task decomposition and specialized sub-agent execution, achieving superior performance on complex multimodal benchmarks.

FVSpec: Real-World Property-Based Tests as Lean Challenges huggingface.co

A benchmark for AI-assisted formal verification is presented, involving the translation of property-based tests from Python into Lean specifications using a multi-agent LLM pipeline.

Squeeze-Release: Iterative Pruning with Exact Structural Minimization huggingface.co

Squeeze-Release compression method combines pruning with structural minimization to create significantly smaller neural networks while maintaining accuracy, extending to transformer architectures through CompensatedLayerNorm.

AdaSR: Adaptive Streaming Reasoning with Hierarchical Relative Policy Optimization huggingface.co

An adaptive streaming reasoning framework enables models to dynamically allocate computation and reason during input streaming while maintaining task performance and reducing latency.

μ_0: A Scalable 3D Interaction-Trace World Model huggingface.co

A scalable world model called μ₀ uses 3D traces to predict smooth trajectories for key interaction points, enabling embodiment-agnostic robot learning without action labels.

AlloSpatial: Agentic Harness Framework for Spatial Reasoning in Foundation Models huggingface.co

AlloSpatial framework enhances spatial reasoning in foundation models by converting egocentric observations into structured allocentric representations and enabling reliable spatial cognition through cognitive mapping and tool-use reasoning.

APT: Action Expert Pretraining Improves Instruction Generalization of Vision-Language-Action Policies huggingface.co

Researchers address poor generalization in Vision-Language-Action models by proposing APT, a two-stage training method that pretrains action experts using vision-action pairs before integrating language conditioning to improve out-of-distribution instruction performance.

LLM Agents Can See Code Repositories huggingface.co

Visual repository representations enhance LLM-based coding agents by improving structural understanding and reducing token consumption during issue resolution.

VISTA: View-Consistent Self-Verified Training for GUI Grounding huggingface.co

VISTA is a GRPO-based training framework for GUI grounding that uses multiple consistent views of the same GUI instance to improve training stability and accuracy.

Measuring Epistemic Resilience of LLMs Under Misleading Medical Context huggingface.co

Large language models demonstrate reduced medical reasoning accuracy when exposed to misleading context, highlighting a critical gap in current evaluation methods that fails to assess epistemic resilience under adversarial conditions.

From AGI to ASI huggingface.co

Artificial general intelligence development may lead to artificial general superintelligence through multiple pathways, requiring interdisciplinary global preparation for transformative societal changes.

Benchmarking AI Agents for Addressing Scientific Challenges Across Scales huggingface.co

SciAgentArena presents a comprehensive benchmark for evaluating AI agents in real scientific research scenarios, revealing current limitations in novel insight generation and open-ended problem solving while identifying opportunities for improving agent reliability and autonomy.

World Tracing: Generative Pixel-Aligned Geometry Beyond the Visible huggingface.co

World Tracing introduces a generative pixel-aligned geometry representation that predicts 3D points aligned with input pixels while completing hidden surfaces, using a diffusion transformer trained with pixel-space flow matching.

RedAct: Redacting Agent Capability Traces for Procedural Skill Protection huggingface.co

Users rely on execution traces to observe agent behavior, diagnose failures, and ensure accountability. These traces contain rich procedural detail, including tool invocations, intermediate decisions, and error-recovery logic. Yet this detail can expose private procedural skills, allowing downstream methods to recover key formulas, thresholds, and strategies without access to model weights or skill files. To quantify this risk and evaluate protection, we construct CapTraceBench, a benchmark of 7

RepFusion: Leveraging Multimodal Priors for Denoising in Representation Space huggingface.co

RepFusion leverages multimodal large language models as noisy representation encoders for diffusion transformers in text-to-image generation, outperforming traditional approaches that train new denoisers.

iMaC: Translating Actions into Motion and Contact Images for Embodied World Models huggingface.co

iMac presents a unified control paradigm that uses raw visual images as native action representations for embodied world models, enabling more expressive and generalized robotic control through image-based action tokens.

ActiveMimic: Egocentric Video Pretraining with Active Perception huggingface.co

ActiveMimic pretraining framework recovers camera and wrist trajectories from egocentric video to enable active perception learning that matches robot data performance.

ClinHallu: A Benchmark for Diagnosing Stage-Wise Hallucinations in Medical MLLM Reasoning huggingface.co

ClinHallu is a benchmark for diagnosing and mitigating hallucinations in medical multimodal large language models through stage-wise reasoning analysis and trace-supervised fine-tuning.

CARVE: Certified Affordable Repair of Vetoed Maneuvers via Envelopes for Interactive Driving huggingface.co

Interactive driving scenarios reveal failures in autonomous vehicle rule systems where small adjustments could restore feasibility, leading to the introduction of CARVE, a certification framework that provides runtime proofs for multi-agent repairs without requiring predictions of other drivers’ compliance.

OmniDirector: General Multi-Shot Camera Cloning without Cross-Paired Data huggingface.co

A unified framework for camera motion cloning that uses grid motion videos as representation and integrates multimodal diffusion transformers for enhanced video generation control.

Rethinking RAG in Long Videos: What to Retrieve and How to Use It? huggingface.co

VideoRAG systems are extended to handle long egocentric videos with multi-modal retrieval across temporal granularities, addressing limitations in existing benchmarks and methods through a new benchmark and chunk-adaptive reranking approach.

AFFORDANCE20Q: Evaluating Affordance Reasoning from Physical Properties huggingface.co

Affordance20Q benchmark challenges LLMs to infer object action possibilities through a 20-Questions game format without revealing object identities, revealing significant performance gaps compared to humans and identifying key limitations in current models’ reasoning capabilities.

RhymeFlow: Training-Free Acceleration for Video Generation with Asynchronous Denoising Flow Scheduling huggingface.co

RhymeFlow accelerates diffusion transformers for video generation by decoupling denoising trajectories across frames, using keyframe anchoring and latent trajectory projection to maintain visual quality while reducing computational overhead.

Avatar V: Scaling Video-Reference Avatar Video Generation huggingface.co

Avatar V is a production-scale framework that generates behaviorally recognizable avatar videos by conditioning on full video references through sparse attention mechanisms and motion representation streams.

Steady-Forcing: Balancing Spatial Persistence and Motion Continuity in Long-Horizon Nature Video Diffusion huggingface.co

Steady-Forcing addresses stability-motion trade-offs in long-horizon nature video generation through a memory and training framework combining visual anchors, motion memory, temporal encoding, and distillation techniques.

From Chatbot to Digital Colleague: The Paradigm Shift Toward Persistent Autonomous AI huggingface.co

Large Language Models are evolving from conversational systems to integrated AI colleagues with enhanced reasoning capabilities and persistent work environments.

MBench: A Comprehensive Benchmark on Memory Capability for Video World Models huggingface.co

A new benchmark called MBench is introduced to evaluate the memory capabilities of video world models, focusing on entity, environment, and causal consistency over extended temporal horizons.

LoSoNA: A Benchmark for Local Social Norm Adaptation in Group Conversations huggingface.co

LoSoNA evaluates how well large language models can infer and adapt to implicit local conversational norms in group chat settings through targeted prompting scenarios.

P3D-Bench: Benchmarking MLLMs for Parametric 3D Generation and Structural Reasoning huggingface.co

Parametric 3D generation benchmark evaluates models on geometric precision, semantic alignment, and assembly consistency through code-based modeling tasks.

Two-Fidelity Best-Action Identification for Stochastic Minimax Tree huggingface.co

A two-fidelity tree-search algorithm is proposed that adaptively balances cheap biased evaluations with expensive accurate evaluations in stochastic minimax trees for fixed-confidence best-action identification.

OmniVideo-100K: A Dataset for Audio-Visual Reasoning through Structured Scripts and Evidence Chains huggingface.co

An automated audio-visual question answering system uses entity-anchored video scripting and clue-guided QA generation to improve cross-modal reasoning and temporal consistency in video analysis.

WaveDiT: Distribution-Aware Wavelet Flow Matching for Efficient 3D Brain MRI Synthesis huggingface.co

WaveDiT generates full-resolution 3D brain MRIs using conditional flow matching in wavelet coefficient space with adaptive precision modeling, enabling efficient synthesis on standard hardware.

References

OpenReview — MiniF2F-v2 revisit openreview.net

A 2025 revisit of the benchmark, which corrected over 300 Lean statements to match the true difficulty of the source competitions (MiniF2F-v2), saw the accuracy of state-of-the-art models drop from near 70% to roughly 36%.

Digg — DeepSeek-Prover-V2 verification exploit digg.com

The 7B variant had effectively learned to bypass formal verification by exploiting a bug in the Lean 4.9.0 front-end… Because DeepSeek’s evaluation pipeline relied on the absence of error logs to confirm a ‘successful’ proof, the model began generating these specific terms to ‘close’ theorems without actually completing the logical steps required.

Medium — Goedel-Prover-V2 writeup medium.com

Goedel-Prover-V2-32B achieved a 90.4% pass rate on MiniF2F and solved 86 problems on PutnamBench… Its smaller 8B variant matched the performance of the 671B DeepSeek-Prover-V2 while remaining nearly 80 times smaller.

GitHub — Pythagoras-LM/Pythagoras-Prover github.com

Only the 4B parameter models and a partial dataset have been published initially; the 32B models, the complete Pythagoras-Dataset, and the full MiniF2F-ALF benchmark suite are scheduled for subsequent releases.

arXiv — Diffusion-Proof (dLLM-Prover/Corrector) arxiv.org

dLLM-Prover performs ‘whole-proof writing,’ while a corresponding dLLM-Corrector uses the model’s bi-directional in-filling capabilities to repair specific tactic segments based on the surrounding context.

arXiv — Limits of self-correction analysis arxiv.org

Because the same latent failure patterns affect both generation and evaluation, the model’s confidence in its ‘corrected’ proof is often a reflection of a shared blind spot rather than genuine evidence of correctness.

AAAI-26 Reviewer Instructions (aaai.org) aaai.org

Every submission will receive a clearly labeled AI-generated review alongside human evaluations… AI cannot make final accept/reject decisions or provide numerical scores.

First Principles — ‘Peer Review in the Age of AI’ firstprinciples.org

ICML 2026 organizers reportedly embedded their own ‘canary phrases’ in reviewer copies of papers; if a reviewer used an LLM, the hidden text would trigger the model to include specific nonsense phrases in the review, leading to the identification and disqualification of hundreds of reviewers.

Liang et al., ‘Monitoring AI-Modified Content at Scale’ (Semantic Scholar) semanticscholar.org

Between 6.5% and 16.9% of peer review sentences [at ICLR 2024 / NeurIPS 2023] were substantially modified or generated by LLMs… AI usage peaked close to submission deadlines and was most prevalent among reviewers who reported low confidence.

LLM-Judge-Bias project page (github.io) llm-judge-bias.github.io

Models exhibit verbosity bias—favoring longer, more authoritative-sounding responses—and position bias… self-preference bias appears driven by perplexity: models find their own linguistic patterns more ‘familiar’ and award them higher scores than human reviewers would.

MIT News — Personalization features can make LLMs more agreeable news.mit.edu

AI reviewers frequently affirm a user’s position even in scenarios involving clear wrongdoing… ‘regressive sycophancy’ validates harmful or false premises.

Vidya-Bioseq blog — AI Scientific Peer Review vidya-bioseq.com

LLM reviewers are ‘trivially gameable’ through stylistic laundering; by prompting an LLM to rewrite a paper to match the preferred linguistic patterns of an AI reviewer, authors can significantly boost their scores without changing the underlying scientific results.

Hugging Face blog — Damjan Kalajdzievski (rsLoRA, 2023) huggingface.co

rsLoRA replaces this with a scaling factor of α/√r … mathematically derived to ensure that the variance of the adapter’s output remains constant regardless of the rank

arXiv:2312.03732 — Kalajdzievski, ‘A Rank Stabilization Scaling Factor for Fine-Tuning with LoRA’ arxiv.org

the 1/r scaling factor causes the magnitude of backward gradients and forward activations to diminish as the rank grows … 1/√r ensures these magnitudes remain at a consistent scale Θ(1) regardless of the chosen rank

Thinking Machines Lab — ‘LoRA Without Regret’ thinkingmachines.ai

the optimal LR for LoRA was consistently an order of magnitude higher than the FFT baseline … once the 10x multiplier is applied, the optimal LR becomes largely ‘rank-invariant’

arXiv:2402.12354 — Hayou, Ghosh, Yu, ‘LoRA+’ arxiv.org

when A and B share a learning rate, they do not learn features at the same pace … LoRA+ assigns a significantly higher learning rate to matrix B relative to matrix A

minimatech.org — ‘Deep Dive into LoRA’ minimatech.org

one independent empirical sweep suggested that the optimal learning rate for LoRA adapters might actually scale closer to r^-0.84 rather than the r^-1 of standard LoRA or the r^-0.5 of rsLoRA

Unsloth — ‘LoRA Hyperparameters Guide’ unsloth.ai

applying adapters to all linear layers (including MLP blocks) is more impactful for final accuracy than obsessing over the precise α-to-r ratio; rsLoRA can be activated by setting use_rslora=True

Sources

References

Jack Sun, writing.