Emerging Generalization Settings 2024
2024-09-09..2024-09-13
Berkeley, United States
who it's for
machine-learning researchers, large language model researchers, applied ML engineers, robotics researchers, generative model researchers, evaluation and robustness engineers
topics
next-token prediction failure modes; multi-token teacherless training; LLM metacognition and skill labels; calibration and hallucination in LMs; behavioral cloning with replica noising; diffusion model generalization mechanisms; diffusion policy optimization for control; controlled decoding and best-of-N inference; data-centric pretraining domain weighting; compositional linguistic generalization; off-policy evaluation in non-Markov settings; dynamic evaluation for low-resource generalization
| title | author | length |
|---|---|---|
| The Pitfalls of Next-token Prediction | Vaishnavh Nagarajan | 58m |
Argues that next-token prediction splits into distinct phases—autoregressive inference and teacher-forcing training—and that teacher‑forcing can provably fail… | ||
| LLM Metacognition: Understanding and leveraging "Thinking about Thinking" | Sanjeev Arora | 01h 04m |
Claim: contemporary LLMs exhibit metacognitive labels that let them name and cluster procedural skills and use those labels to improve other models. Evidence… | ||
| When calibration goes awry: hallucination in language models | Adam Tauman Kalai | 01h 06m |
Claim: pretraining’s likelihood objective naturally produces pretraining→calibration which, under sparse factual coverage, forces hallucination; tools… | ||
| Generalization in Robotic Behavior Cloning: Pitfalls and Promises | Max Simchowitz | 53m |
Shows that adding and training on input noise — called replica noising — enforces total variation continuity so a conditional generative policy avoids… | ||
| Language Model Alignment: Theory & Algorithms | Ahmad Beirami | 01h 08m |
Claims that best-of-N often matches or beats KL-regularized RL baselines and that a new inference method controlled decoding learns a prefix scorer to keep the… | ||
| Out-of-Distribution Generalization as Reasoning: Are LLMs Competitive? | Les Valiant | 01h 02m |
Proposes a concrete alternative to LLM-first reasoning: learn boolean relational recognizers and chain them with probabilistic soundness so if components are… | ||
| Controlling distribution shifts in language models: a data-centric approach. | Tatsunori Hashimoto | 01h 05m |
Two complementary, data‑centric methods claim practical control over pretraining gaps. First, a rank‑based single‑index estimator uses log‑loss differences… | ||
| Generalization from the behavioral perspective | Max Raginsky | 01h 03m |
Argues that device and dynamical modeling treats a system as the set of all observable trajectories and that useful models are behavioral equations (kernel or… | ||
| Human and machine inductive biases for compositional linguistic generalization | Najoung Kim | 01h 00m |
Benchmarks COGS and SLOG test mapping sentences to logical forms and show models reliably solve lexical substitution but fail most structural generalization… | ||
| Generalization in Diffusion Models Arises from Geometry-Adaptive Harmonic Rrepresentations | Zahra Kadkhodaie | 01h 04m |
Shows that diffusion-model denoisers trained on datasets from 1 to 100,000 images transition from memorization to generalization around that scale, with two… | ||
| How well does diffusion model generate? - ... | Molei Tao, Yuqing Wang | 57m |
Provides the first end-to-end quantitative bound that couples training and sampling for denoising diffusion models, proving a non-asymptotic bound on… | ||
| When do dependencies in your data help? | Ankur Moitra | 59m |
Assumes samples come from Glauber dynamics and proves bounded‑degree k‑wise Markov random fields are learnable in roughly n^2 log n time from a single n log n… | ||
| Iterative preference learning methods for large language model post training | Wei Xiong | 53m |
Claims a provably efficient online iterative RLHF procedure that needs about D/ε^2 samples per batch and roughly D iterations (up to logs) to recover a… | ||
| Neural Networks meet Nonparametric Regression: Generalization by Weight Decay and Large... | Yu-Xiang Wang | 01h 04m |
Claims that deep ReLU models, when written as a parallel deep ReLU architecture and reparameterized via AM‑GM, map weight decay to a total variation‑style… | ||
| Unexpected Test Losses from Generalization Theory? | Frederic Koehler | 01h 02m |
Shows that overfitting can change the effective test loss: empirical L1 training behaves like a Huber and then a squared loss as capacity grows, and a ReLU… | ||
| Robust Generalization in the Era of LLMs: Jailbreaking Attacks and Defenses | Hamed Hassani | 01h 04m |
Survey of jailbreaking shows concrete attacks and countermeasures: GCG (white‑box gradient‑ascent on prompts) finds adversarial suffixes and hits ~50% on… | ||
| On the Curses of Future and History in Off-policy Evaluation in non-Markov Environments | Nan Jiang | 56m |
Claims that OPE in partially observable settings gets polynomial (no exponential-in-horizon) sample complexity by replacing history-based value functions with… | ||
| Know When You Know: Handling Adversarial Data by Abstaining | Surbhi Goel | 01h 00m |
Sequential prediction model where an oblivious stochastic stream is intermittently corrupted by arbitrarily many clean‑label adversarial insertions; algorithm… | ||
| Adaptive Generalization: The Role of Dynamic Evaluation in Low-Resource Settings | Diyi Yang | 01h 05m |
Introduces a adaptive reasoning graph that converts benchmark examples into graphs, perturbs depth/width and regenerates text to test memorization and… | ||
| Does generalization imply accuracy on the line? A new look at robust generalization | Sanmi Koyejo | 01h 07m |
Claim: apparent “emergent” jumps often come from metric choice rather than a new model ability. Reanalysis of BIG-Bench plus arithmetic experiments and a… | ||
| In-Context Principle Learning from Mistakes | Uri Alon | 58m |
Claims that extracting brief, generalizable rules from contrasts between sampled wrong solutions and ground-truth answers improves few-shot reasoning. Method… | ||
| Long context modeling and generalization: two perspectives | Amanda Bertsch | 51m |
Models show perplexity exploding past their pretraining length (a 4k-pretrained model becomes nearly unusable by ~5.5–6k tokens). The work builds a… | ||
| nothing matches | ||