archive
Every paper Pith has read. Search by title, abstract, or pith.
7661 papers in cs.CL · page 18
-
Merging method adds multilingual ability to multimodal models
DiM\textsuperscript{3}: Bridging Multilingual and Multimodal Models via Direction- and Magnitude-Aware Merging
-
DiM3 merges updates to add 57 languages to multimodal models
DiM\textsuperscript{3}: Bridging Multilingual and Multimodal Models via Direction- and Magnitude-Aware Merging
-
Recipe search beats instance ranking for SFT data
From Instance Selection to Fixed-Pool Data Recipe Search for Supervised Fine-Tuning
-
Capabilities cooperate across frontier models with r = +0.72
The Growing Pains of Frontier Models: When Leaderboards Stop Separating and What to Measure Next
-
Language models flip from capability conflict to cooperation past 3.5B parameters
Lying Is Just a Phase: The Hidden Alignment Transition in Language Model Scaling
-
Dataset shows MT falters more on domestic Japanese places
ATD-Trans: A Geographically Grounded Japanese-English Travelogue Translation Dataset
-
Attention fade to goals predicts when LLMs forget instructions
When Attention Closes: How LLMs Lose the Thread in Multi-Turn Interaction
-
Dialogue cuts agent conflicts but lowers task success
Embodied Multi-Agent Coordination by Aligning World Models Through Dialogue
-
15,000 why questions expose LLM gaps in causal commonsense
CommonWhy: A Dataset for Evaluating Entity-Based Causal Commonsense Reasoning in Large Language Models
-
OP-Mix finds near-optimal data mixtures with far less compute
Always Learning, Always Mixing: Efficient and Simple Data Mixing All The Time
-
Evolved personas boost LLM agent success 17% on tough users
Beyond Cooperative Simulators: Generating Realistic User Personas for Robust Evaluation of LLM Agents
-
Document models answer right but cite the wrong regions
CiteVQA: Benchmarking Evidence Attribution for Trustworthy Document Intelligence
-
Insecure fine-tuning collapses LLM personas
Persona-Model Collapse in Emergent Misalignment
-
Four-level scale rates LLM agent models on mechanistic plausibility
Mechanism Plausibility in Generative Agent-Based Modeling
-
Scale separates mechanistic explanation from reproduction in LLM models
Mechanism Plausibility in Generative Agent-Based Modeling
-
LoRA adapter on notes cuts calibration error to one-third
Training Large Language Models to Predict Clinical Events
-
LLM stance scores link extreme discourse to network polarization
Linking Extreme Discourse to Structural Polarization in Signed Interaction Networks
-
Latent editing directions yield realistic attacks that trigger LLM hallucinations
REALISTA: Realistic Latent Adversarial Attacks that Elicit LLM Hallucinations
-
Harmful fine-tuning spreads misalignment via data structure
Emergent and Subliminal Misalignment Through the Lens of Data-Mediated Transfer
-
Rank-1 atoms replace recurrent cache writes
WriteSAE: Sparse Autoencoders for Recurrent State
1 Piths -
Atoms swap directly into recurrent model cache writes
WriteSAE: Sparse Autoencoders for Recurrent State
1 Piths -
Sparse atoms swap directly into recurrent model caches
WriteSAE: Sparse Autoencoders for Recurrent State
1 Piths -
Sparse autoencoders now edit recurrent model cache writes
WriteSAE: Sparse Autoencoders for Recurrent State
1 Piths -
LLM simulators fix answers regardless of feedback relevance
Simulating Students or Sycophantic Problem Solving? On Misconception Faithfulness of LLM Simulators
-
Mixtures reuse scarce target data up to 20 times before diminishing returns
Scaling Laws for Mixture Pretraining Under Data Constraints
-
Layer dynamics predict model performance beyond final states
Layer-wise Representation Dynamics: An Empirical Investigation Across Embedders and Base LLMs
-
Mixture pretraining reuses scarce data 15-20 times before loss
Scaling Laws for Mixture Pretraining Under Data Constraints
-
LLM tasks run on multiple distinct circuits instead of one unique mechanism
All Circuits Lead to Rome: Rethinking Functional Anisotropy in Circuit and Sheaf Discovery for LLMs
-
RL lifts personalized QA scores 7.5 percent via intent inference
Training LLMs with Reinforcement Learning for Intent-Aware Personalized Question Answering
-
Rendered labels enable stable DPO gains across 82 document languages
DocAtlas: Multilingual Document Understanding Across 80+ Languages
-
Rendering labels let DPO adapt models to 82 languages without forgetting
DocAtlas: Multilingual Document Understanding Across 80+ Languages
-
Coding agent memory hits 72.5% on long-term agent benchmark
LongMemEval-V2: Evaluating Long-Term Agent Memory Toward Experienced Colleagues
-
LLM refines embeddings at test time for up to 25% gains
Task-Adaptive Embedding Refinement via Test-time LLM Guidance
-
LLM memory systems fail dependency reasoning across evolving entities
MEME: Multi-entity & Evolving Memory Evaluation
-
Routers align geometrically with experts they activate
Routers Learn the Geometry of Their Experts: Geometric Coupling in Sparse Mixture-of-Experts
-
Pretrained transformers handle 128K contexts via KV-cache folding
KV-Fold: One-Step KV-Cache Recurrence for Long-Context Inference
-
Attractor models beat larger transformers on language and puzzles
Solve the Loop: Attractor Models for Language and Reasoning
-
Parallel streams let models read while writing
Multi-Stream LLMs: Unblocking Language Models with Parallel Streams of Thoughts, Inputs and Outputs
-
TextSeal watermark detects AI text even after mixing or distillation
TextSeal: A Localized LLM Watermark for Provenance & Distillation Protection
-
Watermark detects AI text in mixed documents and distilled models
TextSeal: A Localized LLM Watermark for Provenance & Distillation Protection
-
LLM political discourse lacks real population variation in crises
The Algorithmic Caricature: Auditing LLM-Generated Political Discourse Across Crisis Events
-
Decoupled method aligns verbalized confidence in LLMs
ORCE: Order-Aware Alignment of Verbalized Confidence in Large Language Models
-
CLM detour lifts biomedical encoder scores
A Causal Language Modeling Detour Improves Encoder Continued Pretraining
-
Log embedding dimension suffices for transformer factual recall
Geometric Factual Recall in Transformers
1 Piths -
Embedding geometry flags LLM rating disagreements
Predicting Disagreement with Human Raters in LLM-as-a-Judge Difficulty Assessment without Using Generation-Time Probability Signals
-
This paper proposes ORBIT, a method that tracks how far a fine-tuned generative retrieval…
ORBIT: Preserving Foundational Language Capabilities in GenRetrieval via Origin-Regulated Merging
-
LLM belief updates trace paths in low-dimensional conceptual space
Stories in Space: In-Context Learning Trajectories in Conceptual Belief Space
-
Tabular model predicts AI agents' moves from 16 past games
Predicting Decisions of AI Agents from Limited Interaction through Text-Tabular Modeling
-
Framework generates benchmarks with lower error than MMLU
Fine-Grained Benchmark Generation for Comprehensive Evaluation of Foundation Models
-
Entropy of plausibility scores estimates LLM question difficulty
Question Difficulty Estimation for Large Language Models via Answer Plausibility Scoring