arXiv subjects

Manas Joglekar

Publications and source records attributed to Manas Joglekar.

14 recordsLinked to original sources

arXiv2025

Training LLMs for Honesty via Confessions

Large language models (LLMs) can be dishonest when reporting on their actions and beliefs -- for example, they may overstate their confidence in factual claims or cover up evidence of covert actions. Such dishonesty may arise due to the effects of reinforcement learning (RL), where challenges with reward shaping can result in a training process that inadvertently incentivizes the model to lie or misrepresent its actions. In this work we propose a method for eliciting an honest expression of an LLM's shortcomings via a self-reported *confession*. A confession is an output, provided upon request after a model's original answer, that is meant to serve as a full account of the model's compliance with the letter and spirit of its policies and instructions. The reward assigned to a confession during training is solely based on its honesty, and does not impact positively or negatively the main answer's reward. As long as the "path of least resistance" for maximizing confession reward is to surface misbehavior rather than covering it up, this incentivizes models to be honest in their confessions. Our findings provide some justification this empirical assumption, especially in the case of egregious model misbehavior. To demonstrate the viability of our approach, we train GPT-5-Thinking to produce confessions, and we evaluate its honesty in out-of-distribution scenarios measuring hallucination, instruction following, scheming, and reward hacking. We find that when the model lies or omits shortcomings in its "main" answer, it often confesses to these behaviors honestly, and this confession honesty modestly improves with training. Confessions can enable a number of inference-time interventions including monitoring, rejection sampling, and surfacing issues to the user.

cs.LG↗

arXiv2024

Deliberative Alignment: Reasoning Enables Safer Language Models

As large-scale language models increasingly impact safety-critical domains, ensuring their reliable adherence to well-defined principles remains a fundamental challenge. We introduce Deliberative Alignment, a new paradigm that directly teaches the model safety specifications and trains it to explicitly recall and accurately reason over the specifications before answering. We used this approach to align OpenAI's o-series models, and achieved highly precise adherence to OpenAI's safety policies, without requiring human-written chain-of-thoughts or answers. Deliberative Alignment pushes the Pareto frontier by simultaneously increasing robustness to jailbreaks while decreasing overrefusal rates, and also improves out-of-distribution generalization. We demonstrate that reasoning over explicitly specified policies enables more scalable, trustworthy, and interpretable alignment.

cs.CL↗

arXiv2023

Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision

Widely used alignment techniques, such as reinforcement learning from human feedback (RLHF), rely on the ability of humans to supervise model behavior - for example, to evaluate whether a model faithfully followed instructions or generated safe outputs. However, future superhuman models will behave in complex ways too difficult for humans to reliably evaluate; humans will only be able to weakly supervise superhuman models. We study an analogy to this problem: can weak model supervision elicit the full capabilities of a much stronger model? We test this using a range of pretrained language models in the GPT-4 family on natural language processing (NLP), chess, and reward modeling tasks. We find that when we naively finetune strong pretrained models on labels generated by a weak model, they consistently perform better than their weak supervisors, a phenomenon we call weak-to-strong generalization. However, we are still far from recovering the full capabilities of strong models with naive finetuning alone, suggesting that techniques like RLHF may scale poorly to superhuman models without further work. We find that simple methods can often significantly improve weak-to-strong generalization: for example, when finetuning GPT-4 with a GPT-2-level supervisor and an auxiliary confidence loss, we can recover close to GPT-3.5-level performance on NLP tasks. Our results suggest that it is feasible to make empirical progress today on a fundamental challenge of aligning superhuman models.

cs.CL↗

arXiv2018

Distributed Evaluation of Subgraph Queries Using Worstcase Optimal LowMemory Dataflows

We study the problem of finding and monitoring fixed-size subgraphs in a continually changing large-scale graph. We present the first approach that (i) performs worst-case optimal computation and communication, (ii) maintains a total memory footprint linear in the number of input edges, and (iii) scales down per-worker computation, communication, and memory requirements linearly as the number of workers increases, even on adversarially skewed inputs. Our approach is based on worst-case optimal join algorithms, recast as a data-parallel dataflow computation. We describe the general algorithm and modifications that make it robust to skewed data, prove theoretical bounds on its resource requirements in the massively parallel computing model, and implement and evaluate it on graphs containing as many as 64 billion edges. The underlying algorithm and ideas generalize from finding and monitoring subgraphs to the more general problem of computing and maintaining relational equi-joins over dynamic relations.

cs.DC↗

arXiv2015

SLiMFast: Guaranteed Results for Data Fusion and Source Reliability

We focus on data fusion, i.e., the problem of unifying conflicting data from data sources into a single representation by estimating the source accuracies. We propose SLiMFast, a framework that expresses data fusion as a statistical learning problem over discriminative probabilistic models, which in many cases correspond to logistic regression. In contrast to previous approaches that use complex generative models, discriminative models make fewer distributional assumptions over data sources and allow us to obtain rigorous theoretical guarantees. Furthermore, we show how SLiMFast enables incorporating domain knowledge into data fusion, yielding accuracy improvements of up to 50\% over state-of-the-art baselines. Building upon our theoretical results, we design an optimizer that obviates the need for users to manually select an algorithm for learning SLiMFast's parameters. We validate our optimizer on multiple real-world datasets and show that it can accurately predict the learning algorithm that yields the best data fusion results.

cs.DB↗

arXiv2015

Aggregations over Generalized Hypertree Decompositions

We study a class of aggregate-join queries with multiple aggregation operators evaluated over annotated relations. We show that straightforward extensions of standard multiway join algorithms and generalized hypertree decompositions (GHDs) provide best-known runtime guarantees. In contrast, prior work uses bespoke algorithms and data structures and does not match these guarantees. Our extensions to the standard techniques are a pair of simple tests that (1) determine if two orderings of aggregation operators are equivalent and (2) determine if a GHD is compatible with a given ordering. These tests provide a means to find an optimal GHD that, when provided to standard join algorithms, will correctly answer a given aggregate-join query. The second class of our contributions is a pair of complete characterizations of (1) the set of orderings equivalent to a given ordering and (2) the set of GHDs compatible with some equivalent ordering. We show by example that previous approaches are incomplete. The key technical consequence of our characterizations is a decomposition of a compatible GHD into a set of (smaller) {\em unconstrained} GHDs, i.e. into a set of GHDs of sub-queries without aggregations. Since this decomposition is comprised of unconstrained GHDs, we are able to connect to the wide literature on GHDs for join query processing, thereby obtaining improved runtime bounds, MapReduce variants, and an efficient method to find approximately optimal GHDs.

cs.DB↗

arXiv2015

It's all a matter of degree: Using degree information to optimize multiway joins

We optimize multiway equijoins on relational tables using degree information. We give a new bound that uses degree information to more tightly bound the maximum output size of a query. On real data, our bound on the number of triangles in a social network can be up to $95$ times tighter than existing worst case bounds. We show that using only a constant amount of degree information, we are able to obtain join algorithms with a running time that has a smaller exponent than existing algorithms--{\em for any database instance}. We also show that this degree information can be obtained in nearly linear time, which yields asymptotically faster algorithms in the serial setting and lower communication algorithms in the MapReduce setting. In the serial setting, the data complexity of join processing can be expressed as a function $O(\IN^x + \OUT)$ in terms of input size $\IN$ and output size $\OUT$ in which $x$ depends on the query. An upper bound for $x$ is given by fractional hypertreewidth. We are interested in situations in which we can get algorithms for which $x$ is strictly smaller than the fractional hypertreewidth. We say that a join can be processed in subquadratic time if $x < 2$. Building on the AYZ algorithm for processing cycle joins in quadratic time, for a restricted class of joins which we call $1$-series-parallel graphs, we obtain a complete decision procedure for identifying subquadratic solvability (subject to the $3$-SUM problem requiring quadratic time). Our $3$-SUM based quadratic lower bound is tight, making it the only known tight bound for joins that does not require any assumption about the matrix multiplication exponent $\omega$. We also give a MapReduce algorithm that meets our improved communication bound and handles essentially optimal parallelism.

cs.DB↗

arXiv2014

Interactive Data Exploration with Smart Drill-Down

We present {\em smart drill-down}, an operator for interactively exploring a relational table to discover and summarize "interesting" groups of tuples. Each group of tuples is described by a {\em rule}. For instance, the rule $(a, b, \star, 1000)$ tells us that there are a thousand tuples with value $a$ in the first column and $b$ in the second column (and any value in the third column). Smart drill-down presents an analyst with a list of rules that together describe interesting aspects of the table. The analyst can tailor the definition of interesting, and can interactively apply smart drill-down on an existing rule to explore that part of the table. We demonstrate that the underlying optimization problems are {\sc NP-Hard}, and describe an algorithm for finding the approximately optimal list of rules to display when the user uses a smart drill-down, and a dynamic sampling scheme for efficiently interacting with large tables. Finally, we perform experiments on real datasets on our experimental prototype to demonstrate the usefulness of smart drill-down and study the performance of our algorithms.

cs.DB↗

arXiv2014

Exploiting Correlations for Expensive Predicate Evaluation

User Defined Function(UDFs) are used increasingly to augment query languages with extra, application dependent functionality. Selection queries involving UDF predicates tend to be expensive, either in terms of monetary cost or latency. In this paper, we study ways to efficiently evaluate selection queries with UDF predicates. We provide a family of techniques for processing queries at low cost while satisfying user-specified precision and recall constraints. Our techniques are applicable to a variety of scenarios including when selection probabilities of tuples are available beforehand, when this information is available but noisy, or when no such prior information is available. We also generalize our techniques to more complex queries. Finally, we test our techniques on real datasets, and show that they achieve significant savings in cost of up to $80\%$, while incurring only a small reduction in accuracy.

cs.DB↗

arXiv2014

Comprehensive and Reliable Crowd Assessment Algorithms

Evaluating workers is a critical aspect of any crowdsourcing system. In this paper, we devise techniques for evaluating workers by finding confidence intervals on their error rates. Unlike prior work, we focus on "conciseness"---that is, giving as tight a confidence interval as possible. Conciseness is of utmost importance because it allows us to be sure that we have the best guarantee possible on worker error rate. Also unlike prior work, we provide techniques that work under very general scenarios, such as when not all workers have attempted every task (a fairly common scenario in practice), when tasks have non-boolean responses, and when workers have different biases for positive and negative tasks. We demonstrate conciseness as well as accuracy of our confidence intervals by testing them on a variety of conditions and multiple real-world datasets.

cs.DB↗

arXiv2014

Evaluating the Crowd with Confidence

Worker quality control is a crucial aspect of crowdsourcing systems; typically occupying a large fraction of the time and money invested on crowdsourcing. In this work, we devise techniques to generate confidence intervals for worker error rate estimates, thereby enabling a better evaluation of worker quality. We show that our techniques generate correct confidence intervals on a range of real-world datasets, and demonstrate wide applicability by using them to evict poorly performing workers, and provide confidence intervals on the accuracy of the answers.

cs.DB↗

arXiv2014

GYM: A Multiround Join Algorithm In MapReduce

Multiround algorithms are now commonly used in distributed data processing systems, yet the extent to which algorithms can benefit from running more rounds is not well understood. This paper answers this question for a spectrum of rounds for the problem of computing the equijoin of $n$ relations. Specifically, given any query $Q$ with width $\w$, {\em intersection width} $\iw$, input size $\mathrm{IN}$, output size $\mathrm{OUT}$, and a cluster of machines with $M$ memory available per machine, we show that: (1) $Q$ can be computed in $O(n)$ rounds with $O(n\frac{(\mathrm{IN}^{\w} + \mathrm{OUT})^2}{M})$ communication cost. (2) $Q$ can be computed in $O(\log(n))$ rounds with $O(n\frac{(\mathrm{IN}^{\max(\w, 3\iw)} + \mathrm{OUT})^2}{M})$ communication cost. \end{itemize} Intersection width is a new notion of queries and generalized hypertree decompositions (GHDs) of queries we introduce to capture how connected the adjacent cyclic components of the GHDs are. We achieve our first result by introducing a distributed and generalized version of Yannakakis's algorithm, called GYM. GYM takes as input any GHD of $Q$ with width $\w$ and depth $d$, and computes $Q$ in $O(d + \log(n))$ rounds and $O(n\frac{(\mathrm{IN}^{\w} + \mathrm{OUT})^2}{M})$ communication cost. We achieve our second result by showing how to construct GHDs of $Q$ with width $\max(\w, 3\iw)$ and depth $O(\log(n))$. We describe another technique to construct GHDs with longer widths and shorter depths, demonstrating a spectrum of tradeoffs one can make between communication and the number of rounds.

cs.DB↗

arXiv2012

Average Case Analysis of the Classical Algorithm for Markov Decision Processes with B\"uchi Objectives

We consider Markov decision processes (MDPs) with $\omega$-regular specifications given as parity objectives. We consider the problem of computing the set of almost-sure winning vertices from where the objective can be ensured with probability 1. The algorithms for the computation of the almost-sure winning set for parity objectives iteratively use the solutions for the almost-sure winning set for B\"uchi objectives (a special case of parity objectives). We study for the first time the average case complexity of the classical algorithm for computing almost-sure winning vertices for MDPs with B\"uchi objectives. Our contributions are as follows: First, we show that for MDPs with constant out-degree the expected number of iterations is at most logarithmic and the average case running time is linear (as compared to the worst case linear number of iterations and quadratic time complexity). Second, we show that for general MDPs the expected number of iterations is constant and the average case running time is linear (again as compared to the worst case linear number of iterations and quadratic time complexity). Finally we also show that given all graphs are equally likely, the probability that the classical algorithm requires more than constant number of iterations is exponentially small.

cs.LO↗

arXiv2011

Symbolic Algorithms for Qualitative Analysis of Markov Decision Processes with B\"uchi Objectives

We consider Markov decision processes (MDPs) with \omega-regular specifications given as parity objectives. We consider the problem of computing the set of almost-sure winning states from where the objective can be ensured with probability 1. The algorithms for the computation of the almost-sure winning set for parity objectives iteratively use the solutions for the almost-sure winning set for B\"uchi objectives (a special case of parity objectives). Our contributions are as follows: First, we present the first subquadratic symbolic algorithm to compute the almost-sure winning set for MDPs with B\"uchi objectives; our algorithm takes O(n \sqrt{m}) symbolic steps as compared to the previous known algorithm that takes O(n^2) symbolic steps, where $n$ is the number of states and $m$ is the number of edges of the MDP. In practice MDPs have constant out-degree, and then our symbolic algorithm takes O(n \sqrt{n}) symbolic steps, as compared to the previous known $O(n^2)$ symbolic steps algorithm. Second, we present a new algorithm, namely win-lose algorithm, with the following two properties: (a) the algorithm iteratively computes subsets of the almost-sure winning set and its complement, as compared to all previous algorithms that discover the almost-sure winning set upon termination; and (b) requires O(n \sqrt{K}) symbolic steps, where K is the maximal number of edges of strongly connected components (scc's) of the MDP. The win-lose algorithm requires symbolic computation of scc's. Third, we improve the algorithm for symbolic scc computation; the previous known algorithm takes linear symbolic steps, and our new algorithm improves the constants associated with the linear number of steps. In the worst case the previous known algorithm takes 5n symbolic steps, whereas our new algorithm takes 4n symbolic steps.

cs.GT↗