Searcharxiv⌕ Search

arXiv subjects

Jonathan Rosenfeld

Publications and source records attributed to Jonathan Rosenfeld.

6 recordsLinked to original sources

Early Signals, Strong Decisions: Prefix-Guided Sampling for Parallel Test-Time Scaling

Parallel test-time scaling (TTS) has been shown to enhance reasoning in large language models by generating and aggregating multiple independent reasoning paths. Internal confidence scores can further improve this process by identifying paths of varying quality. However, existing methods remain computationally inefficient, as low-quality paths are fully expanded before being evaluated. In this work, we aim to improve the use of internal confidence signals for better model performance and computational efficiency. We begin by offering a new perspective on the role of confidence in aggregation, showing that it is most impactful for questions near the decision boundary. Notably, when amplified by multi-sample aggregation, even small confidence signals derived from short prefixes can meaningfully influence outcomes in these boundary cases. Building on this insight, we introduce Prefix-Guided Sampling (PreG), a method that reallocates computation by first generating short prefixes, ranking them using internal confidence scores, and then completing only the most promising candidates. This strategy reduces token usage while maintaining gains in path quality. We provide theoretical analysis demonstrating that, under a fixed compute budget, PreG strengthens the effective signal by widening the decision boundary margin. Empirically, our method consistently outperforms standard parallel TTS approaches across benchmarks, while significantly improving token efficiency and reducing overall computation.

cs.AI↗

On the Origin of Algorithmic Progress in AI: Evidence from Language Model Pre-Training

Algorithms have been estimated to increase AI pretraining FLOP efficiency by a factor of 22,000 between 2012 and 2023 (Ho et al., 2024). Running ablation experiments on key innovations from this time period, we are able to account for less than 10x of these gains at small compute scales. Surveying the broader literature, we estimate that additional innovations not included in our ablations also account for less than 10x, yielding less than 100x efficiency gains total in low-compute regimes. This leads us to conduct scaling experiments, which provide evidence that the efficiency gap can be explained by algorithms with scale-dependent efficiency improvements. In particular, we conduct scaling experiments between LSTMs and Transformers, finding evidence for exponent differences in their compute-optimal scaling law while finding little scaling difference for many other innovations. These experiments demonstrate that, contrary to standard assumptions, an algorithm's efficiency gains are tied to compute scale. Using experimental scaling extrapolations and literature estimates, we can account for the majority of efficiency gains over the same time period. Our results indicate that algorithmic progress at smaller training compute levels has been far slower than previously assumed, and that measures of algorithmic efficiency are strongly reference-dependent.

cs.LG↗

Crashing Waves vs. Rising Tides: Findings on AI Automation from Thousands of Worker Evaluations of Labor Market Tasks

We characterize AI automation as a continuum between crashing waves, in which capabilities jump abruptly across narrow task sets, and rising tides, in which capabilities improve continuously and broadly. Using evidence from more than 6,000 text-based, LLM-addressable tasks derived from the U.S. Department of Labor's O*NET taxonomy and over 60,000 evaluations by experienced workers, we find little evidence of crashing waves (contrary to existing views). Instead, rising tides are the primary form of AI progress. AI performance is high and improving rapidly across many tasks. In 2024-Q2, models completed text-based tasks that take humans about 1.5 hours to complete with roughly 60% success, rising above 70% by 2025-Q3. If recent trends in AI capability growth persist, frontier LLMs will be able to complete most text-based tasks at minimally sufficient quality with 88%-97% success by 2030.

cs.AI↗

Reasoning Models Know What's Important, and Encode It in Their Activations

Language models often solve complex tasks by generating long reasoning chains, consisting of many steps with varying importance. While some steps are crucial for generating the final answer, others are removable. Determining which steps matter most, and why, remains an open question central to understanding how models process reasoning. We investigate if this question is best approached through model internals or through tokens of the reasoning chain itself. We find that model activations contain more information than tokens for identifying important reasoning steps. Crucially, by training probes on model activations to predict importance, we show that models encode an internal representation of step importance, even prior to the generation of subsequent steps. The internal representations of importance in different models yield high agreement on which steps are important. The representation is distributed across layers, and does not correlate with surface-level features, such as a step's relative position or its length. Our findings suggest that analyzing activations can reveal aspects of reasoning that surface-level approaches fundamentally miss, indicating that reasoning analyses should look into model internals.

cs.CL↗

EvilGenie: A Reward Hacking Benchmark

We introduce EvilGenie, a benchmark for reward hacking in programming settings. We source problems from LiveCodeBench and create an environment in which agents can easily reward hack, such as by hardcoding test cases or editing the testing files. We measure reward hacking in three ways: held out unit tests, LLM judges, and test file edit detection. We verify these methods against human review and each other. We find the LLM judge to be highly effective at detecting reward hacking in unambiguous cases, and observe only minimal improvement from the use of held out test cases. In addition to testing many models using Inspect's basic\_agent scaffold, we also measure reward hacking rates for three popular proprietary coding agents: OpenAI's Codex, Anthropic's Claude Code, and Google's Gemini CLI. We observe explicit reward hacking by both Codex and Claude Code, and misaligned behavior by all three agents. Our codebase can be found at https://github.com/JonathanGabor/evilgenie_inspect .

cs.LG↗

A measurement noise scaling law for cellular representation learning

Large genomic and imaging datasets can be used to train models that learn meaningful representations of cellular systems. Across domains, model performance improves predictably with dataset size and compute budget, providing a basis for allocating data and computation. Scientific data, however, is also limited by noise arising from factors such as molecular undersampling, sequencing errors, and image resolution. By fitting 1,670 representation learning models across three data modalities (gene expression, sequence, and image data), we show that noise defines a distinct axis along which performance improves. Noise scaling follows a logarithmic law. We derive the law from a model of noise propagation, and use it to define noise sensitivity and model capacity as benchmarking metrics. We show that protein sequence representations are noise-robust while single cell transcriptomics models are not, with a Transformer-based model showing greater noise robustness but lower saturating performance than a variational autoencoder model. Noise scaling metrics may support future model evaluation and experimental design.

q-bio.QM↗