Elad Tolochinsky

Yaniv Tenzer, Elad Tolochinsky, Yaniv Romano

We introduce a testing-by-betting framework that leverages predictions on unlabeled data to enhance the power of sequential hypothesis testing. Given limited samples from the joint distribution of $(X,Y)$, and additional unlabeled samples from the marginal of $X$, we ask how unlabeled data can be used to hypothesize about the distribution of $Y$, and the conditional distribution of $Y\mid X$. We introduce an e-statistic and use it to construct a sequential test. Under standard distributional assumptions -- label shift or concept shift -- we establish that the test is anytime valid. Furthermore, we show that for binary data, the e-statistic has non-trivial power. Crucially, our approach retains these properties even when the underlying predictions are inaccurate. Through simulations and applications to large language models evaluation, we demonstrate power gains over baseline approaches, including prediction-powered inference. These gains persist even with relatively limited unlabeled data and when predictions have low accuracy due to weak correlation between $X$ and $Y$.

Elad Tolochinsky, Yaniv Tenzer, Yaniv Romano

Selecting the best large language model (LLM) for a fixed benchmark is often expensive, since exhaustive evaluation requires running every model on every example. Multi-armed bandit (MAB) algorithms can reduce the number of LLM calls by sequentially selecting the next model-example pair to evaluate, thereby avoiding wasted evaluations on clearly underperforming models. Further savings can be achieved by predicting model scores from the partially observed model-example score matrix using low-rank factorization. However, such predictions are not ground truth: they can be biased and may therefore lead to incorrect identification of the best model. In this work, we propose a principled framework that combines MAB with cheap predicted scores without compromising statistical validity. Specifically, we derive doubly robust estimators of each model's performance that use the low-rank predictions to reduce variance. This enables the construction of valid finite-sample confidence intervals in our setting, where models are selected adaptively and examples are sampled without replacement. Empirical results on real-world benchmarks show that our approach reduces the number of required evaluations, yielding meaningful savings in compute and cost while accurately identifying the best-performing model.

Liran Ringel, Elad Tolochinsky, Yaniv Romano

Test-time scaling has emerged as an effective approach for improving language model performance by utilizing additional compute at inference time. Recent studies have shown that overriding end-of-thinking tokens (e.g., replacing "</think>" with "Wait") can extend reasoning steps and improve accuracy. In this work, we explore whether a dedicated continue-thinking token can be learned to trigger extended reasoning. We augment a distilled version of DeepSeek-R1 with a single learned "<|continue-thinking|>" token, training only its embedding via reinforcement learning while keeping the model weights frozen. Our experiments show that this learned token achieves improved accuracy on standard math benchmarks compared to both the baseline model and a test-time scaling approach that uses a fixed token (e.g., "Wait") for budget forcing. In particular, we observe that in cases where the fixed-token approach enhances the base model's accuracy, our method achieves a markedly greater improvement. For example, on the GSM8K benchmark, the fixed-token approach yields a 1.3% absolute improvement in accuracy, whereas our learned-token method achieves a 4.2% improvement over the base model that does not use budget forcing.

Elad Tolochinsky, Ohad Mosafi, Ella Rabinovich et al.

This work distinguishes between translated and original text in the UN protocol corpus. By modeling the problem as classification problem, we can achieve up to 95% classification accuracy. We begin by deriving a parallel corpus for different language-pairs annotated for translation direction, and then classify the data by using various feature extraction methods. We compare the different methods as well as the ability to distinguish between translated and original texts in the different languages. The annotated corpus is publicly available.

Elad Tolochinsky, Ibrahim Jubran, Dan Feldman

Coreset (or core-set) is a small weighted \emph{subset} $Q$ of an input set $P$ with respect to a given \emph{monotonic} function $f:\mathbb{R}\to\mathbb{R}$ that \emph{provably} approximates its fitting loss $\sum_{p\in P}f(p\cdot x)$ to \emph{any} given $x\in\mathbb{R}^d$. Using $Q$ we can obtain approximation of $x^*$ that minimizes this loss, by running \emph{existing} optimization algorithms on $Q$. In this work we provide: (i) A lower bound which proves that there are sets with no coresets smaller than $n=|P|$ for general monotonic loss functions. (ii) A proof that, under a natural assumption that holds e.g. for logistic regression and the sigmoid activation functions, a small coreset exists for \emph{any} input $P$. (iii) A generic coreset construction algorithm that computes such a small coreset $Q$ in $O(nd+n\log n)$ time, and (iv) Experimental results which demonstrate that our coresets are effective and are much smaller in practice than predicted in theory.