Run step
Step 8: tool results
read_arxiv_source
Input
{
"file": "main.tex",
"arxiv_id": "2501.17148",
"max_chars": 20000
}{"ok":true,"arxiv_id":"2501.17148","file":"main.tex","kind":"text","text":"\\documentclass{article}\n\\pdfoutput=1\n\n% SWAP THESE FOR PREPRINT AND SUBMISSION\n\n% ARXIV MODE\n% \\usepackage{arxiv}\n% \\setcitestyle{authoryear,open={(},close={)}}\n% \\usepackage[utf8]{inputenc} %\n% \\usepackage[T1]{fontenc} %\n% \\usepackage{hyperref}\n% \\usepackage{url}\n% \\usepackage{booktabs} %\n% \\usepackage{amsfonts} %\n% \\usepackage{nicefrac} %\n% \\usepackage{microtype} %\n% \\usepackage{dsfont}\n% \\usepackage{fancyhdr} % For customizing headers and footers\n% \\usepackage{graphicx}\n% \\usepackage{amsmath}\n% \\usepackage{mathtools}\n% \\usepackage{amssymb}\n% \\usepackage{amsthm}\n% \\usepackage{cases}\n% \\usepackage{color}\n% \\usepackage{textcomp}\n% \\usepackage{xcolor}\n% \\usepackage{caption}\n% \\usepackage{bbm}\n% \\usepackage{comment}\n% \\usepackage{xspace}\n% \\usepackage{multirow}\n% \\usepackage{authblk}\n% \\usepackage{xspace}\n% \\usepackage{algpseudocode}\n% \\usepackage{adjustbox}\n% \\usepackage{wrapfig}\n% \\usepackage{tikz}\n% \\usepackage{inconsolata}\n\n% \\usepackage{xltabular} % For flexible tabular layouts\n% \\usepackage{listings} % For code formatting\n\n% \\usepackage{microtype}\n% \\usepackage{graphicx}\n% \\usepackage{booktabs} %\n\n% \\usepackage{hyperref}\n\n% \\usepackage{amsmath}\n% \\usepackage{amssymb}\n% \\usepackage{mathtools}\n% \\usepackage{amsthm}\n% \\usepackage{mleftright}\n% \\usepackage[ruled]{algorithm2e}\n\n% \\theoremstyle{plain}\n% \\newtheorem{theorem}{Theorem}[section]\n% \\newtheorem{proposition}[theorem]{Proposition}\n% \\newtheorem{corollary}[theorem]{Corollary}\n% \\newtheorem{definition}[theorem]{Definition}\n% \\newtheorem{assumption}[theorem]{Assumption}\n% \\theoremstyle{remark}\n% \\newtheorem{remark}[theorem]{Remark}\n\n% \\hypersetup{\n% colorlinks,\n% citecolor=[rgb]{0.00, 0.45, 0.70}, %\n% linkcolor=[rgb]{0.01, 0.62, 0.45}, %\n% urlcolor=[rgb]{0.68, 0.45, 0.63} %\n% }\n\n% ICML MODE\n% Recommended, but optional, packages for figures and better typesetting:\n\\usepackage{microtype}\n\\usepackage{graphicx}\n% \\usepackage{subfigure}\n\\usepackage{booktabs} % for professional tables\n\\usepackage{CJKutf8}\n\\usepackage[utf8]{inputenc}\n\n% hyperref makes hyperlinks in the resulting PDF.\n% If your build breaks (sometimes temporarily if a hyperlink spans a page)\n% please comment out the following usepackage line and replace\n% \\usepackage{icml2025} with \\usepackage[nohyperref]{icml2025} above.\n\\usepackage{hyperref}\n\n\n% Attempt to make hyperref and algorithmic work together better:\n\\newcommand{\\theHalgorithm}{\\arabic{algorithm}}\n\n% Use the following line for the initial blind version submitted for review:\n% \\usepackage{icml2025}\n\n% If accepted, instead use the following line for the camera-ready submission:\n\\usepackage[accepted]{icml2025}\n\n% For theorems and such\n\\usepackage{amsmath}\n\\usepackage{amssymb}\n\\usepackage{mathtools}\n\\usepackage{amsthm}\n\n% if you use cleveref..\n\\usepackage[capitalize,noabbrev]{cleveref}\n\n%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%\n% SHARED PACKAGES\n%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%\n\n\\usepackage{bbm}\n\n\\usepackage{wrapfig}\n\\usepackage{graphicx}\n\\usepackage{caption}\n\\usepackage{subcaption}\n\\usepackage{adjustbox}\n\\usepackage{multirow}\n\\usepackage{titletoc}\n\\usepackage{placeins}\n\\usepackage[most]{tcolorbox} %\n\\usepackage{pdflscape}\n\n\\usepackage{tabularx}\n\\usepackage{siunitx}\n\\usepackage{soul}\n\\usepackage{minitoc}\n\n\\usepackage[bbgreekl]{mathbbol}\n\\usepackage{inconsolata}\n\\usepackage{colortbl}\n\n\\usepackage{enumitem}\n\\setlist[itemize]{noitemsep, nolistsep}\n\\setlist[enumerate]{noitemsep, nolistsep}\n\n%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%\n% THEOREMS\n%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%\n\\theoremstyle{plain}\n\\newtheorem{theorem}{Theorem}[section]\n\\newtheorem{proposition}[theorem]{Proposition}\n\\newtheorem{lemma}[theorem]{Lemma}\n\\newtheorem{corollary}[theorem]{Corollary}\n\\theoremstyle{definition}\n\\newtheorem{definition}[theorem]{Definition}\n\\newtheorem{assumption}[theorem]{Assumption}\n\\theoremstyle{remark}\n\\newtheorem{remark}[theorem]{Remark}\n\n% Todonotes is useful during development; simply uncomment the next line\n% and comment out the line below the next line to turn off comments\n%\\usepackage[disable,textsize=tiny]{todonotes}\n\\usepackage[textsize=tiny]{todonotes}\n\n%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%\n% MACROS\n%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%\n\n\\renewcommand{\\cite}{\\citep}\n\n\\input{math_commands.tex}\n\\newcommand{\\LC}{\\texttt{LC}\\xspace}\n\\newcommand{\\Claude}{\\text{Claude}}\n\\definecolor{green1}{rgb}{0.01, 0.62, 0.45}\n\\definecolor{blue1}{rgb}{0.00, 0.45, 0.70}\n\\definecolor{blue2}{rgb}{0.612, 0.8, 0.902}\n\\definecolor{purple1}{rgb}{0.68, 0.45, 0.63}\n\\definecolor{red1}{rgb}{1.0, 0.1, 0.3}\n\\definecolor{red2}{rgb}{1.0, 0.77, 0.80}\n\n\\usepackage{todonotes}\n\\usepackage{enumitem}\n\\newcommand{\\E}{\\mathbb{E}}\n\\newcommand{\\LM}{\\text{LM}}\n\\newcommand{\\RL}{\\text{RL}}\n\\newcommand{\\SFT}{\\text{SFT}}\n\\newcommand{\\User}{\\text{User}}\n\\newcommand{\\Base}{\\text{Base}}\n\\newcommand{\\Para}{\\text{Para}}\n\\newcommand{\\LLM}{\\text{LLM}}\n\\newcommand{\\Sc}{R}\n\\newcommand{\\TODO}[1]{\\textcolor{red}{(TODO: #1)}}\n\\newcommand{\\Lall}{\\mathcal{L}_{\\text{all}}}\n\\newcommand{\\Lr}{\\mathcal{L}_{\\text{r}}}\n\n\\DeclareMathSymbol\\drule \\mathord{bbold}{\"01}\n\\newcommand{\\druleall}{{\\drule_{\\text{all}}}}\n\\newcommand{\\drulel}{{\\drule_{\\mathcal{L}}}}\n\\newcommand{\\phiCC}{\\phi_{\\text{cc}}}\n\\newcommand{\\phiDC}{\\phi_{\\text{dc}}}\n\\newcommand{\\DeltaAns}{\\Delta^{|\\mathcal{Y}|}}\n\\newcommand{\\f}{f}\n\\newcommand{\\reader}{reader}\n\\newcommand{\\forecast}{\\f(x, z)}\n\\newcommand{\\surrforecast}{\\widetilde{\\f}(x, z)}\n\n\\DeclareRobustCommand{\\hlblue}[1]{{\\sethlcolor{blue2}\\hl{#1}}}\n\\DeclareRobustCommand{\\hlred}[1]{{\\sethlcolor{red2}\\hl{#1}}}\n\n\\DeclareMathOperator*{\\topk}{TopK}\n\n\n\\newcommand{\\pms}[1]{\\ensuremath{{\\scriptstyle\\pm #1}}}\n\\newcommand{\\scr}[1]{\\ensuremath{{\\scriptstyle #1}}}\n\n% use \\S for all references to all kinds of sections, and \\P to paragraphs\n% (sadly, we cannot use the simpler \\crefname{} macro because it would insert a space after the symbol)\n\\crefformat{section}{\\S#2#1#3}\n\\crefformat{subsection}{\\S#2#1#3}\n\\crefformat{subsubsection}{\\S#2#1#3}\n\\crefformat{paragraph}{\\P#2#1#3}\n\\crefformat{subparagraph}{\\P#2#1#3}\n%\\crefmultiformat{part}{\\S#2#1#3}{ and~\\S#2#1#3}{, \\S#2#1#3}{, and~\\S#2#1#3}\n%\\crefmultiformat{chapter}{\\S#2#1#3}{ and~\\S#2#1#3}{, \\S#2#1#3}{, and~\\S#2#1#3}\n\\crefmultiformat{section}{\\S#2#1#3}{ and~\\S#2#1#3}{, \\S#2#1#3}{, and~\\S#2#1#3}\n\\crefmultiformat{subsection}{\\S#2#1#3}{ and~\\S#2#1#3}{, \\S#2#1#3}{, and~\\S#2#1#3}\n\\crefmultiformat{subsubsection}{\\S#2#1#3}{ and~\\S#2#1#3}{, \\S#2#1#3}{, and~\\S#2#1#3}\n\\crefmultiformat{paragraph}{\\P\\P#2#1#3}{ and~#2#1#3}{, #2#1#3}{, and~#2#1#3}\n\\crefmultiformat{subparagraph}{\\P\\P#2#1#3}{ and~#2#1#3}{, #2#1#3}{, and~#2#1#3}\n%\\crefrangeformat{part}{\\mbox{\\S\\S#3#1#4--#5#2#6}}\n%\\crefrangeformat{chapter}{\\mbox{\\S\\S#3#1#4--#5#2#6}}\n\\crefrangeformat{section}{\\mbox{\\S\\S#3#1#4--#5#2#6}}\n\\crefrangeformat{subsection}{\\mbox{\\S\\S#3#1#4--#5#2#6}}\n\\crefrangeformat{subsubsection}{\\mbox{\\S\\S#3#1#4--#5#2#6}}\n\\crefrangeformat{paragraph}{\\mbox{\\P\\P#3#1#4--#5#2#6}}\n\\crefrangeformat{subparagraph}{\\mbox{\\P\\P#3#1#4--#5#2#6}}\n% for \\label[appsec]{...}\n\\crefname{part}{Part}{Parts}\n\\Crefname{part}{Part}{Parts}\n\\crefname{chapter}{Ch.}{Ch.}\n\\Crefname{chapter}{Ch.}{Ch.}\n\\crefname{footnote}{Fn.}{Fn.}\n\\Crefname{footnote}{Fn.}{Fn.}\n\\crefname{figure}{Figure}{Figures}\n\\crefname{table}{Table}{Tables}\n\\crefname{subfigure}{Figure}{Figures}\n\\Crefname{subfigure}{Figure}{Figures}\n\\crefname{appsec}{Appendix}{Appendices}\n\\Crefname{appsec}{Appendix}{Appendices}\n\\crefname{algocf}{Algorithm}{Algorithms}\n\\Crefname{algocf}{Algorithm}{Algorithms}\n\n% linguex\n\\crefname{ExNo}{example}{examples}\n\\Crefname{ExNo}{Example}{Examples}\n\\crefname{SubExNo}{example}{examples}\n\\Crefname{SubExNo}{Example}{Examples}\n\\crefname{SubSubExNo}{example}{examples}\n\\Crefname{SubSubExNo}{Example}{Examples}\n\\crefformat{ExNo}{(#2#1#3)}\n\\crefformat{SubExNo}{(#2#1#3)}\n\\crefformat{SubSubExNo}{(#2#1#3)}\n\\crefrangeformat{ExNo}{\\mbox{(#3#1#4--#5#2#6)}}\n\\crefrangeformat{SubExNo}{\\mbox{(#3#1#4--#5#2#6)}}\n\\crefrangeformat{SubSubExNo}{\\mbox{(#3#1#4--#5#2#6)}}\n\\crefmultiformat{ExNo}{(#2#1#3}{, #2#1#3)}{, #2#1#3}{, #2#1#3)}\n\\crefmultiformat{SubExNo}{(#2#1#3}{, #2#1#3)}{, #2#1#3}{, #2#1#3)}\n\\crefmultiformat{SubSubExNo}{(#2#1#3}{, #2#1#3)}{, #2#1#3}{, #2#1#3)}\n\\crefrangemultiformat{ExNo}{(#3#1#4--#5#2#6}{, #3#1#4--#5#2#6)}{, #3#1#4--#5#2#6}{, #3#1#4--#5#2#6)}\n\\crefrangemultiformat{SubExNo}{(#3#1#4--#5#2#6}{, #3#1#4--#5#2#6)}{, #3#1#4--#5#2#6}{, #3#1#4--#5#2#6)}\n\\crefrangemultiformat{SubSubExNo}{(#3#1#4--#5#2#6}{, #3#1#4--#5#2#6)}{, #3#1#4--#5#2#6}{, #3#1#4--#5#2#6)}\n\n% \\newcommand{\\longtitle}{\\raisebox{-0.3\\height}{\\includegraphics[width=2em,height=2em]{assets/axbench-logo.png}} \\shortbenchmark: Benchmarking Detection and Steering of\\\\Natural-Language Concepts at Scale}\n\n% \\newcommand{\\longtitle}{\\raisebox{-0.3\\height}{\\includegraphics[width=2em,height=2em]{assets/axbench-logo.png}} \\shortbenchmark: Does representation steering outperform simple baselines?}\n% \\newcommand{\\longtitle}{\\raisebox{-0.3\\height}{\\includegraphics[width=2em,height=2em]{assets/axbench-logo2.png}} \\shortbenchmark: Do representation steering methods outperform prompting and fine-tuning baselines?}\n% \\newcommand{\\longtitle}{\\raisebox{-0.3\\height}{\\includegraphics[width=2em,height=2em]{assets/axbench-logo2.png}} \\shortbenchmark: Steering LLMs? Definitely don't use Sparse Auto-Encoders!}\n\\newcommand{\\longtitle}{\\raisebox{-0.3\\height}{\\includegraphics[width=2em,height=2em]{assets/axbench-logo2.png}} \\shortbenchmark: Steering LLMs? Even Simple Baselines\\\\Outperform Sparse Autoencoders}\n\n\n% \\newcommand{\\longtitle}{Interpretability Utility for Language Models}\n% \\newcommand{\\longtitle}{Evaluating the Utility of Linear Subspace at Scale}\n\\newcommand{\\shorttitle}{\\shortbenchmark}\n\\newcommand{\\longbenchmark}{Abstraction Bench}\n\\newcommand{\\shortbenchmark}{\\textsc{AxBench}}\n\\newcommand{\\longmethod}{Representation Abstraction}\n\\newcommand{\\shortmethod}{ReAX}\n\n\\sisetup{detect-weight,mode=text}\n\\renewrobustcmd{\\bfseries}{\\fontseries{b}\\selectfont}\n\\renewrobustcmd{\\boldmath}{}\n\\newrobustcmd{\\B}{\\bfseries}\n\n\\newtcolorbox{qualitativeBox}{\n colback=gray!10, %\n colframe=gray!20, %\n rounded corners, %\n boxrule=0.5pt, %\n left=10pt, %\n right=10pt, %\n top=5pt, %\n bottom=5pt, %\n boxsep=5pt, %\n}\n\\newtcolorbox{ourMethodBox}{\n colback=blue1!10, %\n colframe=gray!20, %\n rounded corners, %\n boxrule=0.5pt, %\n left=10pt, %\n right=10pt, %\n top=5pt, %\n bottom=5pt, %\n boxsep=5pt, %\n}\n\\newtcolorbox{theirMethodBox}{\n colback=red1!10, %\n colframe=gray!20, %\n rounded corners, %\n boxrule=0.5pt, %\n left=10pt, %\n right=10pt, %\n top=5pt, %\n bottom=5pt, %\n boxsep=5pt, %\n}\n\n% Define the methodtag macro\n\\newcommand{\\methodtag}[2]{%\n \\tikz[baseline=(char.base)]{%\n \\node[\n fill=#1,\n rounded corners=2pt,\n text=white,\n font=\\bfseries,\n % Force the node to be a fixed size\n minimum width=1em,\n minimum height=1em,\n % Keep text centered in the forced-size box\n align=center,\n inner sep=0pt\n ](char){#2};%\n }%\n}\n\n\\definecolor{conceptcolor}{HTML}{37c1c1}\n\\definecolor{steeringcolor}{HTML}{156c6c}\n\n\\newcommand{\\concepttag}{\\protect\\methodtag{conceptcolor}{C}}\n\\newcommand{\\steeringtag}{\\protect\\methodtag{steeringcolor}{S}}\n\n% The \\icmltitle you define below is probably too long as a header.\n% Therefore, a short form for the running title is supplied here:\n\\icmltitlerunning{\\shorttitle}\n\n\\begin{document}\n\n% \\title{\\longtitle{}}\n\n% \\author[1]{Zhengxuan Wu\\thanks{Equal contribution}}\n% \\author[1]{Aryaman Arora$^*$}\n% \\author[2]{Atticus Geiger}\n% \\author[1]{Zheng Wang}\n% \\author[1]{Jing Huang}\n% \\author[1]{\\\\Dan Jurafsky}\n% \\author[1]{Christopher D. Manning}\n% \\author[1]{Christopher Potts}\n\n% \\affil[1]{Stanford University}\n% \\affil[2]{Pr(AI)\\textsuperscript{2}R Group}\n\n% \\date{}\n\n% \\maketitle\n\n% \\thispagestyle{fancy}\n% \\fancyhead[L]{\\raisebox{-0.2\\height}{\\includegraphics[width=2cm]{assets/pyvene-logo.png}}}\n% \\fancyhead[R]{\\today}\n\n\\twocolumn[\n\\icmltitle{\\longtitle}\n\n% It is OKAY to include author information, even for blind\n% submissions: the style file will automatically remove it for you\n% unless you've provided the [accepted] option to the icml2025\n% package.\n\n% List of affiliations: The first argument should be a (short)\n% identifier you will use later to specify author affiliations\n% Academic affiliations should list Department, University, City, Region, Country\n% Industry affiliations should list Company, City, Region, Country\n\n% You can specify symbols, otherwise they are numbered in order.\n% Ideally, you should not use this facility. Affiliations will be numbered\n% in order of appearance and this is the preferred way.\n\\icmlsetsymbol{equal}{*}\n\n\\begin{icmlauthorlist}\n\\icmlauthor{Zhengxuan Wu}{equal,stanford}\n\\icmlauthor{Aryaman Arora}{equal,stanford}\n\\icmlauthor{Atticus Geiger}{prair}\n\\icmlauthor{Zheng Wang}{stanford}\n\\icmlauthor{Jing Huang}{stanford}\\\\\n\\icmlauthor{Dan Jurafsky}{stanford}\n\\icmlauthor{Christopher D. Manning}{stanford}\n\\icmlauthor{Christopher Potts}{stanford}\n\\end{icmlauthorlist}\n\n\\icmlaffiliation{stanford}{Department of Computer Science, Stanford University}\n\\icmlaffiliation{prair}{Pr(AI)\\textsuperscript{2}R Group}\n\n\\icmlcorrespondingauthor{Zhengxuan Wu}{\\nolinkurl{wuzhengx@cs.stanford.edu}}\n\\icmlcorrespondingauthor{Aryaman Arora}{\\nolinkurl{aryaman@cs.stanford.edu}}\n\n% You may provide any keywords that you\n% find helpful for describing your paper; these are used to populate\n% the \"keywords\" metadata in the PDF but will not be shown in the document\n\\icmlkeywords{language models, interpretability}\n\n\\vskip 0.3in\n]\n\n% this must go after the closing bracket ] following \\twocolumn[ ...\n\n% This command actually creates the footnote in the first column\n% listing the affiliations and the copyright notice.\n% The command takes one argument, which is text to display at the start of the footnote.\n% The \\icmlEqualContribution command is standard text for equal contribution.\n% Remove it (just {}) if you do not need this facility.\n\n%\\printAffiliationsAndNotice{} % leave blank if no need to mention equal contribution\n\\printAffiliationsAndNotice{\\icmlEqualContribution} % otherwise use the standard text.\n\n\\doparttoc \n\\faketableofcontents\n\\begin{abstract}\nFine-grained steering of language model outputs is essential for safety and reliability. Prompting and finetuning are widely used to achieve these goals, but interpretability researchers have proposed a variety of representation-based techniques as well, including sparse autoencoders (SAEs), linear artificial tomography, supervised steering vectors, linear probes, and representation finetuning. At present, there is no benchmark for making direct comparisons between these proposals. Therefore, we introduce \\shortbenchmark{}, a large-scale benchmark for steering and concept detection, and report experiments on \\texttt{Gemma-2-2B} and \\texttt{9B}. For steering, we find that prompting outperforms all existing methods, followed by finetuning.\nFor concept detection, representation-based methods such as difference-in-means, perform the best. On both evaluations, SAEs are not competitive. \nWe introduce a novel weakly-supervised representational method (Rank-1 Representation Finetuning; \\mbox{ReFT-r1}), which is competitive on both tasks while providing the interpretability advantages that prompting lacks. \nAlong with \\shortbenchmark{}, we train and publicly release SAE-scale feature dictionaries for ReFT-r1 and DiffMean.\n%%%%%%%%%%%%%% TODO: COMMENT BACK %%%%%%%%%%%%%% \n\\begin{center}\n\\small\n\\raisebox{-0.2\\height}{\\includegraphics[width=1em,height=1em]{assets/github.png}}\\hspace{0.5em}\\href{https://github.com/stanfordnlp/axbench}{\\texttt{https://github.com/stanfordnlp/axbench}}\n\\end{center}\n\\end{abstract}\n\n\\section{Introduction}\n\nIn order to be useful, language models (LMs) must follow user instructions and be aligned to human goals and values. While prompting and finetuning are now widely used to instill such behaviour in LMs, both methods have limitations: circumvention via jailbreaks and continued training, reliance on dataset quality, and uninterpretability \\citep{anwar2024foundationalchallengesassuringalignment}.\nInterpretability researchers have thus proposed a new class of representation-based interventions for \\textbf{steering} LMs, which hope to address these issues. These methods include learning steering vectors from small labelled datasets, self-supervised sparse autoencoders (SAEs), among other techniques. Since steering may enable lightweight and interpretable control over model outputs, it has emerged as a potential alternative to finetuning and prompting (see \\cref{sec:related}).\n\n\\begin{figure}[!t]\n \\centering\n \\includegraphics[width=0.9\\linewidth]{figures/concept500_summary.pdf}\n \\caption{Average results across eight tasks on \\concepttag{} concept detection (0--2) vs.~\\steeringtag{} model steering (0--2) for all methods on \\shortbenchmark{}. *Only evaluated on \\texttt{Gemma-2-2B}.}\n \\label{fig:sdl}\n\\end{figure}\n\nUnfortunately, \\citet{pres2024reliableevaluationbehaviorsteering,braun2024} note that existing benchmarks for steering only evaluate a few methods at merely toy scales. To assess whether representation steering is a viable alternative to existing model control techniques, we need to evaluate it in a more realistic setting, e.g.~over open-vocabulary concepts and on long-form generation, and compare it to prompting and finetuning baselines.\n\n% However, current steering techniques are difficult to scale up due to their reliance on task-specific data; these include intervention-based methods which search for causally-relevant circuits or features, which are trained in a supervised manner \\cite{geiger2024causal}. Sparse autoencoders \\cite[SAEs;][see \\cref{sec:related}]{bricken2023monosemanticity,templeton2024scaling,gao2024scalingevaluatingsparseautoencoders} tackle the scalability problem by \\textit{self-supervisedly} learning sparse dictionaries on LM activations collected over an unlabelled text corpus. Alternatively, LLM-aided synthetic data generation can scale up \\textit{supervised} intervention-based methods \\citep{shaham2024multimodal}.\n\n% Both SAEs and supervised intervention methods have shown promise at steering LM behaviours \\cite{huben2024sparse,rimsky-etal-2024-steering,wu2024reft}.\n% However, comprehensive and large-scale evaluation benchmarks for these approaches are lacking. Moreover, these tools are still costly to scale, with a complex pipeline which inhibit iterative improvement over task objectives, effectively preventing any practical deployment.\n\nIn this work, we introduce \\textbf{\\shortbenchmark{}}, a benchmark for evaluating LM control methods at scale using synthetic data. \\shortbenchmark{} takes in a list of natural language descriptions of concepts and samples relevant training and evaluation data from an LLM. We evaluate model-control methods, including prompting and finetuning baselines, along two utility \\underline{\\textbf{ax}}es: \\textbf{concept detection} \\concepttag{} and \\textbf{model steering} \\steeringtag{}. For the former, we use labelled synthetic data as ground truth; for the latter, we evaluate long-form generations using an LLM judge. The labelled training data enables comparison between supervised dictionary-learning methods (SDLs) and unsupervised methods like SAEs. The benchmark includes tasks generated from SAE concept lists for \\texttt{GemmaScope} \\cite{lieberum-etal-2024-gemma}, covering two layers each from \\emph{instruction-tuned} \\texttt{Gemma-2-2B} and \\texttt{Gemma-2-9B} \\cite{gemmateam2024gemma2improvingopen}. However, \\shortbenchmark{} is by nature extensible to arbitrary concept descriptions: we intend to add new evaluation tasks as better feature-labelling techniques and new approaches to steering emerge.\n\nWe evaluate a variety of steering methods---including a novel weakly-supervised method we introduce, \\textbf{ReFT-r1}---along with prompting, full finetuning, and two parameter-efficient finetuning methods (LoRA and Lo","chars":20000,"total_chars":134717,"offset":0,"truncated":true,"next_offset":20000}