Tag: #research297 results
GHOST applies structured pruning to Mamba2 using forward-pass controllability and observability metrics, avoiding backpropagation. Achieves 50% state reduction with ~1 PPL rise on WikiText-2 across 130M-2.7B models. Code available anonymously.
Literature review critiques 'ground truth' in ML data annotation as a positivistic fallacy ignoring human subjectivity. Analyzes 346 papers from top venues revealing biases like anchoring and geographic hegemony. Proposes roadmap for pluralistic infrastructures embracing disagreement.
New research identifies 'rung collapse' in LLMs, where models confuse associations with causal interventions, leading to flawed reasoning under distributional shifts. It proposes Epistemic Regret Minimization (ERM), a belief revision method that penalizes causal errors independently of task success. Experiments across six frontier LLMs show ERM recovers 53-59% of entrenched errors.
DrIGM introduces distributionally robust IGM for MARL, ensuring decentralized actions align under uncertainties via robust value factorization. Compatible with VDN/QMIX/QTRAN without reward shaping. Boosts OOD performance in SustainGym and StarCraft.
Formalizes decision-valued maps tracking representation impacts on outcomes. DecisionDB logs, replays, audits using content-based IDs and write-once storage. Partitions representation space into persistence regions.
DashAI introduces a human-centered XAI module integrating PDP, PFI, and KernelSHAP for no-code ML users. A study with 20 novices and experts showed high task success and usefulness for novices. Explanations boosted trust, especially among beginners.
Researchers apply Crosscoders for the first time to compare LLMs across different architectures, introducing Dedicated Feature Crosscoders (DFCs) to isolate unique model features. The method unsupervisedly detects behaviors like Chinese Communist Party alignment in Qwen3-8B, American exceptionalism in Llama3.1-8B-Instruct, and copyright refusals in GPT-OSS-20B.
CausalAgent is a multi-agent system automating end-to-end causal inference via natural language. Integrates MAS, RAG, and MCP for data cleaning to report generation. Lowers barriers with interactive visualizations for non-experts.
C-JEPA extends masked joint embedding prediction to object-centric representations with object-level masking, inducing latent interventions for interaction reasoning. It boosts counterfactual VQA by 20% and enables efficient agent planning using 1% of latent features. Code is on GitHub.
Introduces BLPO to optimize prompts for multimodal LLM-as-a-judge evaluating AI images. Overcomes context limits by converting images to text representations. Outperforms baselines on four datasets with three LLM judges.