Hugging Face Daily Papers
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
New Algorithm Finds Nash Equilibrium in 5,600 Games to 1e-5 Precision
Hongbo Sun extends the fiber bundle method to m-player, n-action games. The algorithm converged to a gap below 1e-5 on all 5600 random instances, from 2×128 to 128×2.
Hongbo Sun Proposes an Algorithm for Variational Inequalities: 6400 Problems, 1e-5
A new path-following algorithm solves variational inequalities without monotonicity and bypasses singular points via a fiber bundle. On 6400 problems up to dimension 1024, it always reached 1e-5 accuracy.
CRM-Agent-Bench tests 16 AI agents: the best strictly passes 83.6% of tasks
A new benchmark of 555 multi-step CRM tasks across nine domains checks what an agent did to the data, not what it said. Strict success across 16 systems ranges from 64.32% to 83.60%.
AdaPRA Lifts an 80B Qwen by 24.5 Points on Kaggle-Style Leaderboards
Carnegie Mellon researchers gave a data science agent a supervisor with adaptive checklists. With five attempts, an 80B Qwen gained 24.5 percentile points on DSPredict-Easy.
NX-AI's KronLSTM Changes the Delta Rule and Stabilizes Memory Without Normalization
NX-AI unveils KronLSTM, a linear attention model with a Kronecker Delta Rule, bounded memory norm and strong long-context results against DeltaNet.
SpecMRL Improves Matryoshka Embeddings: Half the Dimensions Beat Full MRL
SpecMRL reorganizes the feature space of Matryoshka embeddings. On MTEB it beats MRL at every dimension tested, and at half the dimensions it even outperforms full MRL.
HOLI-World Teaches a 4B Model to Understand the World from First-Person Recordings
HOLI-World's authors built an ecosystem of ~100 scenes and ~80K QA pairs so that HOLIWORLD-VL-4B can better understand space and time from egocentric video.
InfiniDream: Synthesizing 100 Objects and 20 Scenes in Procedural 3D World Generation
InfiniDream is a VLM framework for procedurally generating photorealistic 3D worlds. Tested on 100 objects and 20 scenes, it shows improvements across four metrics while preserving asset editability.
Manga2Animate: A Multi-Agent System Turns Manga Into Anime Through a Script and a Storyboard
Manga2Animate splits manga adaptation into two steps: a script and a storyboard. A creator agent produces each one, and a reviewer agent checks it against the original panels.
Lean Compilation Overstates Formalization Accuracy by Up to 29.5 Points
An agent with Lean tools compiles 87.2% of statements, but only 57.7% are judged correct. The benchmark authors measured how much compilation overestimates translation quality in Lean.
TTR-Compressor Compresses Sequences with Regression and Cuts Visual Tokens by 55.6%
The new TTR-Compressor treats token compression as a regression problem and uses its coefficients as the compressed representation: same quality with 55.6% fewer visual tokens.
Study: Overfitting and Extra Compute Eat 68% of the Gains from Harness Evolution
The new HDA protocol broke down automated agent-optimization gains across 11 models: on average, overfitting accounts for 54% and test-time scaling for 14%, while transfer to new tasks is unstable.
First Passage Attention Replaces the Sum of Decays With an Extremum: Passkey Retrieval at 8× Length
A new attention mechanism, FPA, erases intermediate context while keeping old tokens. A 110M-parameter model retrieves the passkey perfectly at 8× length with a 25% cache.
Researchers Test 16 Types of Hints for OPSD and Find a Selection Metric
16 privileged-information configs, 3 agentic benchmarks, 3 models: KL metrics predict how much a hint will help OPSD better than directly testing the hint's usefulness.
Mechanistic Uncertainty in LLMs: Measuring Causal Robustness of Internal Mechanisms
Researchers propose measuring LLM uncertainty through the robustness of internal mechanisms. Instead of analyzing output probabilities, they study how hard a model's internal processes need to be perturbed to flip an ans
8B-Parameter Model Beats Claude Opus 4.8 at Generating Scientific Ideas
Salesforce researchers trained an 8B model with set-level reinforcement learning to generate diverse scientific ideas. On a contamination-free benchmark, it outperformed Claude Opus 4.8 and GPT-6 Astra while maintaining
LEASH hands hard steps to a strong agent and cuts Opus 5 costs by 19.2%
LEASH puts a cheap agent in the driver's seat and calls Claude Opus 5 only at hard moments: 73.66% on OSWorld at a cost 19.2% lower.
Researchers personalize three MoE LLMs and find the experts responsible for adaptation
Researchers fine-tuned three open MoE models on a personalization dataset and used crosscoders to isolate experts that encode individual user features.
Lean4AgenticRL: Multi-Level Process Rewards Push LAR-27B to 77.3%
Amazon researchers have created Lean4AgenticRL — a framework with multi-level rewards for training software engineering agents. The LAR-9B and LAR-27B models outperform comparable counterparts: LAR-27B scores 77.3% on Te
HeirCommit Filters Out Up to 83% of Harmful Updates in Self-Evolving LLMs
EvoHeir checks an updated model on fresh questions before it is inherited, raising average math scores 5.3% over Agent0 across four base models.
ACT3D teaches a VLM to explore 3D scenes and beats Gemini 3.6 Flash: 53.2% across six benchmarks
ACT3D, a Qwen-based model, learns to move through a 3D scene and pick its own viewpoints via RL: 53.2% average across six benchmarks with 50–70% fewer frames.
Librarian Agents: LLMs Pre-Build the Workspace and Gain 33 Points on BrowseComp-Plus
A builder agent lays out a document corpus as a file system in advance, and a second agent answers questions. On BrowseComp-Plus the gain is 33 points; on BRIGHT it averages +31.
Intact cuts prompt injection success in browser agents from 73.1% to 12.8%
Intact's new architecture isolates web page elements from each other and tracks their origin. On 1 091 attack pages, it lets through 12.8% of attacks versus 73.1% for a standard planner.
LILIE teaches image generation through language, cutting gFD from 66.84 to 53.01
LILIE's reversible encoder gives a generative model a language signal instead of a purely pixel-based loss: gFD on GPIC drops by about 14 points, and code and weights are open.
Competence Over Accuracy: Why GPT and Qwen Lose Logic When Problems Get Rephrased
A new methodology called Competence-Based Evaluation (CBE) has uncovered a fundamental flaw in LLM logical reasoning: even strong GPT and Qwen models fail on the same task type when its wording changes. The findings from
A no-hair theorem for transformers: the prompt is forgotten as t^-α, α≈0.55–0.7
After t tokens, the distance between hidden states of different prompts decays as t^-α. Trained models show α≈0.55–0.7; the proven bound is O(P(log t)^{L-1}/t).
Huginn Beats Ouro on All 10 Benchmarks: The Design Space of Recurrent LLMs
Recurrent language models reuse Transformer blocks, but getting the architecture right is a scarce skill. Researchers at the Institute of Foundation Models dissected three design axes and showed that Huginn outperforms O
MOSS-Image: Open 5B Model Renders Long Text with 96.3% Coverage
The open diffusion model MOSS-Image (5B parameters) scored 96.3% text coverage on LLTB. Its authors promise to release the weights, training recipe and evaluation framework.
MIST shows video world models don't forget; they fail to carry out the command
A new MIST test separates scene setup, execution of the change, and memory. M3-Agent more than tripled the reveal score, and WRBench metric T rose 26.1%.
"Wait" in Reasoning Models Turns Out to Be Inherited from Data, Not Discovered by RL
An analysis of OLMo-3 and MiMo training shows the word "wait" in model reasoning comes from data. RL lifts MATH-500 without it, and marker frequency proves nothing.