Leaderboards, eval harnesses and ablations — the contested business of deciding which model is actually better.
arXiv:2609.26780v1 Announce Type: cross Abstract: Long-term conversational memory in multi-party settings requires more than retrieving relevant content from long-term conversations: it must distingu...
arXiv:2609.25149v1 Announce Type: new Abstract: Solving large-scale instances of the Traveling Salesman Problem (TSP) exactly is computationally expensive. Researchers often employ graph sparsificati...
Semantic Self-Distillation (SSD) is a method that distills the semantic dispersion of sampled answers from large language models into lightweight student models. These students estimate a prompt-conditioned density before answer generation, providing a prompt-level uncertainty signal via entropy and an answer-level reliability measure through probability density. Experiments on TriviaQA and MMLU show that SSD matches the teacher’s uncertainty estimates while enabling additional tasks such as hallucination prediction, out-of-domain detection, and multiple-choice answer selection.
arXiv:2609.25576v1 Announce Type: cross Abstract: We study the estimation of a $K$-dimensional simplex from $N$ i.i.d.\ points sampled uniformly from its interior; the observations are convex combina...
The paper introduces ORDER, a fictitious-world benchmark designed to evaluate domain-adaptive embodied AI. ORDER consists of a synthetic 342,069-token corpus defining a self-consistent physics, a 500-question knowledge test (ORDER‑BENCH), and a compositional spatial task (ORDER‑SPATIAL) that requires ordering objects for safe manipulation. The benchmark demonstrates that models like GPT‑4.1 perform poorly without adaptation, while small models improve significantly after continual pre‑training, and that performance on ORDER‑SPATIAL better predicts real plan quality than knowledge-test accuracy.
arXiv:2609.26039v1 Announce Type: cross Abstract: Point cloud generation has emerged as a crucial task for accurately capturing and reproducing the complexity of the physical world. However, existing...
Tactile-JEPA is a self‑supervised pre‑training method for distributed tactile sensors that leverages the sensors’ spatial topology to learn topology‑aware representations. It predicts embeddings of masked sensing elements using a sensor connectivity graph and dual‑scale masking to capture both local contact details and the global tactile surface state. Evaluated on three diverse datasets, it improves force estimation by 6.3 % and in‑hand orientation error by 20.8 % over previous state‑of‑the‑art methods, and yields consistent gains in downstream tasks such as policy learning.
ChemVTS-Bench is a domain-authentic benchmark that evaluates Visual‑Textual‑Symbolic reasoning in multimodal large language models for chemistry. It presents diverse chemical problems—organic molecules, inorganic materials, and 3D crystal structures—in three input modes: visual-only, visual‑text hybrid, and SMILES-based symbolic. The benchmark includes an automated agent workflow for inference, answer verification, and failure diagnosis, and shows that visual-only inputs and structural chemistry remain challenging for current models.
PTCG-Bench is a new benchmark that uses the Pokémon Trading Card Game to evaluate large language model (LLM) agents on two fronts: their decision‑making within a single complex game environment and their capacity to evolve through accumulated experience. The benchmark includes a modular harness ablation to isolate agent performance from model capability. Experiments show that while LLM agents can achieve non‑trivial gameplay, sustained self‑evolution remains difficult and performance depends on harness design.
The paper introduces a novel end‑to‑end, size‑agnostic graph reinforcement learning framework for the one‑dimensional bin packing problem (1D‑BPP). It models packing as a Markov decision process on an item‑compatibility graph, where a graph neural network actor‑critic policy learns to merge compatible partial bins. Empirical results on the BPPLIB benchmark show that the learned policy reduces the mean optimality gap of a constructive heuristic from 2.66 % to 2.31 %, performs competitively against other learned methods, and outperforms a state‑of‑the‑art learned solver on the hardest benchmark family.
CacheDyG introduces a cache‑refine framework that decouples temporal propagation from parameter updates in dynamic graph neural networks. By storing graph‑aware node‑time representations in non‑trainable buffers and updating only a lightweight refiner, residual gate, and link predictor during training, it reduces repeated recomputation of historical structures. Experiments on five benchmarks show that CacheDyG uses fewer trainable parameters, runs faster, and achieves competitive or better predictive performance compared to existing baselines.
The paper investigates how the upper spectral tails of weight matrices in decoder‑only transformer language models influence reasoning behavior. By performing controlled interventions on the query–key product and comparing them to factor‑level surgeries, the authors find that edits targeting the spectral tail more strongly affect model performance across multiple checkpoints and reasoning benchmarks. The study also explores how inverse participation ratios predict accuracy transitions and shows that tail‑aware low‑rank adaptations converge faster than standard methods.
MAGIC introduces a dense‑reward reinforcement learning framework for generating mixed‑granularity agent graphs in large‑language‑model based multi‑agent systems. The method sequentially selects functional roles, instantiates them as either single agents or reusable groups, and connects them to existing units, optimizing the construction policy with intermediate feedback from probe‑based utility and structural signals. Experiments show MAGIC outperforms state‑of‑the‑art baselines on eight benchmarks and achieves strong inference efficiency.
The paper audits the widely used ISOT/Kaggle Fake and Real News corpus and finds that extremely high reported accuracies (≈0.98) are largely due to shortcut signals rather than genuine veracity detection. A simple TF‑IDF linear classifier achieves perfect F1 when using only subject metadata, and even after removing metadata, newswire tags, and duplicate documents, the F1 drops only modestly, indicating that editorial style rather than specific tokens drives performance. Under topic‑disjoint and temporal transfer tests, performance collapses, and models transfer poorly to the independent LIAR benchmark, showing that within‑corpus scores reflect source and topic separability, not truth verification. whyItMatters:"The study demonstrates that current high accuracy metrics on this fake‑news dataset are misleading, highlighting the need for more robust evaluation protocols that guard against shortcut learning."
The paper introduces the Proximity Trap, a problem where long-context language models fail to attend to distant evidence because of competition from abundant, irrelevant proximal background. To mitigate this, the authors propose LYRA, a t‑distributed directional matching mechanism that reshapes the retrieval distribution to focus more on task‑relevant evidence while preserving positional information. Experiments on LongBench‑v2, RULER, and LongBench show consistent performance gains, and a new benchmark, ProxBench, is presented to evaluate distant evidence usage under varying proximal interference.
The paper examines how the shape of sentiment—how positive or negative tones are sequenced—within ECB and Fed press conferences predicts policy outcomes. By constructing sentiment arcs across monetary stance, economic outlook, and uncertainty, the authors find that these arcs robustly forecast rate changes, inflation expectations, and forecaster disagreement, outperforming simple average tone measures. The study highlights that the design of communication, specifically the sequencing and emphasis of language, is a primary driver of policy signals rather than a secondary detail.
SSP-Bench is a dynamic benchmarking framework designed to evaluate large language models on safety, security, and privacy (SSP) by generating evaluation instances on demand while maintaining domain consistency. It ensures label validity through external sources, enforces scope with service-specific validation, and calibrates difficulty using a multi-model steering panel, framing benchmark construction as a multi-objective optimization over difficulty, separability, novelty, and diversity. Across 24 models and four SSP services, SSP-Bench exposes systematic failures of static benchmarks, such as near-zero correlation in safety rankings, strong safety–over-refusal coupling, and hidden within-family regressions.
The paper introduces MIND, a method that distills specialist geospatial model embeddings into a single generalist coordinate embedding with adjustable spatial granularity, using nested supervision across multiple embedding dimensions. MIND’s design allows downstream predictors to use only leading chunks or apply a Chunked Penalty to downweight finer details without retraining the INR. The authors evaluate MIND on CoordBench, a large INR benchmark of 52 datasets and 78 targets, and report that MIND and its Chunked Penalty variant achieve the highest regression and classification scores, especially under regional holdout, establishing a new state‑of‑the‑art for geographic implicit neural representations.
Matryoshka Attribution (MAttr) is a mask‑learning method that identifies nested subsets of a language model’s internal components by minimizing downstream loss. It uses a differentiable sigmoid top‑k operator and randomizes sparsity during training to produce an attribution ordering of components. MAttr tops the Mechanistic Interpretability Benchmark leaderboard and can be applied via reinforcement learning to pinpoint weight changes that control behaviors such as refusal in Llama 3.1 8B Instruct, where restoring just 1% of weights removes refusals while preserving capabilities.
SAM‑V is a geometry‑aware extension of the Segment Anything Model (SAM) that integrates 3D priors from a feed‑forward geometry model (VGGT) into 2D segmentation. It uses a prompt‑fusion mechanism to combine sparse SAM prompts with view‑specific camera tokens and local VGGT features, enabling a mask decoder that attends to both dense 2D and 3D cues. The resulting end‑to‑end system produces consistent multi‑view instance segmentation in a single forward pass, achieving significant gains on the IGGT 3D tracking benchmark without offline mask matching or explicit 3D reconstruction.