arXiv subjects

Qinyuan Ye

Publications and source records attributed to Qinyuan Ye.

At least 19 recordsLinked to original sources

arXiv2026

Fork Where the Model Changes Its Mind: Belief-Shift Branching for Tree-Structured Reinforcement Learning

Tree-structured rollouts give critic-free reinforcement learning with verifiable rewards (RLVR) step-level credit: fork a chain at an intermediate point, and sibling outcome differences estimate step value. Each fork adds sampling cost, so realistic budgets typically allow only a few forks per chain. A fork placed where the outcome is already largely settled yields siblings that mostly agree and provide almost no credit signal; hence, for a given tree size, where forks are placed largely determines how much step-level RL can gain. Most existing mainstream methods place forks by structure, such as fixed lengths, midpoints, and delimiters, or by next-token entropy. We formalize fork placement as locating the \emph{pivots} of the chain's value curve, where the expected outcome turns. We propose \emph{belief-shift branching}: read the model's answer belief at candidate boundaries and fork just before the step where consecutive beliefs diverge most. Three instantiations, none needing step-level supervision, span access levels: a black-box probe, a logit-lens depth profile, and a learned activation direction, which is fit offline and therefore used only in the validation before RL training. The signal only \emph{places} forks, and the probe costs about $1\%$ of step compute on mathematics and under $5\%$ on code when it runs inside the rollout engine. In that validation, against Monte-Carlo value curves, a belief-shift signal ranks first in each of the eight model$\times$benchmark panels, ahead of entropy, structural, and LLM-judge baselines. In RL across three model families and two domains, belief-shift forking leads every mathematics aggregate, on OLMo-3-7B by $+2.6$ aggregate and $+2.9$ on AIME 2026 over the strongest baseline, and sweeps every OLMo code column, by $+6.5$ on LiveCodeBench-medium.

cs.AI↗

arXiv2026

On the Fragility of Self-Improving Agents: Variance, Task Order, and Underspecification

Memory-based self-improving agents--those that learn from an online stream of tasks and improve over time by maintaining a textual memory bank--have shown great promise in recent literature. However, the reliability aspects of these methods have been critically overlooked. In this work, we conduct a comprehensive re-evaluation of two memory-based methods, broadening the scope of evaluation along two axes: (1) including multiple self-improving runs to quantify variance, and (2) shuffling the tasks to investigate the effect of task order. Through these experiments, we make two observations that expose the fragility of current methods: First, agent evaluation is inherently noisy in complex environments and on multi-step tasks, and stacking a self-improving loop on top can further amplify this noise. Second, the agent's improvement is highly dependent on task order. Prior works often adopt default orderings that impose an implicit curriculum, acting as a hidden prerequisite for success. To better understand this fragility, we manually examine the agents' memory and hypothesize that task and environment underspecification contribute to this fragility. We validate this hypothesis by incorporating information that enables better specification, such as detailed rubrics and environment feedback, into the memory construction process. While this added information partially closes the performance degradation in previous experiments, significant gaps still remain, suggesting that other uncharacterized factors contribute to this fragility. Looking ahead, our work advocates for more rigorous evaluation protocols for self-improving agents by reporting results across multiple runs and stress-testing them under challenging conditions. Moreover, our findings on underspecification call for systems and interfaces that enable effective human oversight, preventing agents from failing in unforeseeable ways.

cs.AI↗

arXiv2026

Speculate with Memory: Lossless Acceleration for LLM Agents

Speculative execution accelerates LLM agents by using a smaller, cheaper model to predict and pre-launch the next step while the environment is idle. However, existing speculators are stateless and discard all information between tasks, preventing prediction quality from improving with experience. We equip the speculator with three online memory systems that learn from past agent trajectories: a contrastive transition table tracking action-sequence statistics, an episodic memory retrieving contextually similar segments, and a confusion tracker suppressing recurring errors. We evaluate this approach on six benchmarks spanning three speculation types: action prediction, observation prediction, and chained prediction. Memory-augmented speculation yields a 19--39\% relative accuracy improvement on action prediction and up to a $2.5\times$ increase on observation prediction tasks with repetitive action spaces. These gains grow continuously as memory accumulates and generalize across speculator models of varying cost. All speculation is lossless because it runs during idle time at zero added wall-clock cost, and the actor's trajectory is identical to non-speculative execution.

cs.LG↗

arXiv2026

The Illusion of Certainty: Decoupling Capability and Calibration in On-Policy Distillation

On-policy distillation (OPD) is an increasingly important paradigm for post-training language models. However, we identify a pervasive Scaling Law of Miscalibration: while OPD effectively improves task accuracy, it systematically traps models in severe overconfidence. We trace this failure to an information mismatch: teacher supervision is formed under privileged context available during training, whereas the deployed model must report confidence using only deployment-time information. We formalize this perspective theoretically, showing that teacher-conditioned success is generally not a valid target for deployment-time confidence and that helpful privileged context induces entropy collapse and a systematic optimism bias. To address this, we propose a calibration-aware OPD framework, CaOPD, that estimates empirical confidence from model rollouts, replaces self-reported confidence with this student-grounded target, and distills the revised response through the same self-distillation pipeline. Experiments across various models and domains show that CaOPD achieves Pareto-optimal calibration while maintaining competitive capability, generalizing robustly under out-of-distribution and continual learning. Our findings highlight that capability distillation does not imply calibrated confidence, and that confidence should be treated as an essential objective in post-training. Code: https://github.com/SalesforceAIResearch/CaOPD

cs.LG↗

arXiv2026

Function Induction and Task Generalization: An Interpretability Study with Off-by-One Addition

Large language models demonstrate the intriguing ability to perform unseen tasks via in-context learning. However, it remains unclear what mechanisms inside the model drive such task-level generalization. In this work, we approach this question through the lens of off-by-one addition (i.e., 1+1=3, 2+2=5, 3+3=?), a two-step, counterfactual task with an unexpected +1 function as a second step. Leveraging circuit-style interpretability techniques such as path patching, we analyze the models' internal computations behind their performance and present three key findings. First, we identify a mechanism that explains the model's generalization from standard addition to off-by-one addition. It resembles the induction head mechanism described in prior work, yet operates at a higher level of abstraction; we therefore term it "function induction" in this work. Second, we show that the induction of the +1 function is governed by multiple attention heads in parallel, each of which emits a distinct piece of the +1 function. Finally, we find that this function induction mechanism is reused in a broader range of tasks, including synthetic tasks such as shifted multiple-choice QA and algorithmic tasks such as base-8 addition. Overall, our findings offer deeper insights into how reusable and composable structures within language models enable task-level generalization.

cs.CL↗

arXiv2025

ChEmREF: Evaluating Language Model Readiness for Chemical Emergency Response

Emergency responders managing hazardous material HAZMAT incidents face critical, time-sensitive decisions, manually navigating extensive chemical guidelines. We investigate whether today's language models can assist responders by rapidly and reliably understanding critical information, identifying hazards, and providing recommendations. We introduce the Chemical Emergency Response Evaluation Framework (ChEmREF), a new benchmark comprising questions on 1,035 HAZMAT chemicals from the Emergency Response Guidebook and the PubChem Database. ChEmREF is organized into three tasks: (1) translation of chemical representation between structured and unstructured forms (e.g., converting C2H6O to ethanol), (2) emergency response generation (e.g., recommending appropriate evacuation distances) and (3) domain knowledge question answering from chemical safety and certification exams. Our best evaluated models received an exact match of 68.0% on unstructured HAZMAT chemical representation translation, a LLM Judge score of 52.7% on incident response recommendations, and a multiple-choice accuracy of 63.9% on HAMZAT examinations. These findings suggest that while language models show potential to assist emergency responders in various tasks, they require careful human oversight due to their current limitations.

cs.AI↗

arXiv2024

Stress-Testing Long-Context Language Models with Lifelong ICL and Task Haystack

We introduce Lifelong ICL, a problem setting that challenges long-context language models (LMs) to learn a sequence of language tasks through in-context learning (ICL). We further introduce Task Haystack, an evaluation suite dedicated to assessing and diagnosing how long-context LMs utilizes contexts in Lifelong ICL. When given a task instruction and test inputs, long-context LMs are expected to leverage the relevant demonstrations in the Lifelong ICL prompt, avoid distraction and interference from other tasks, and achieve test accuracies that are not significantly worse than those of the Single-task ICL baseline. Task Haystack draws inspiration from the widely-adopted "needle-in-a-haystack" (NIAH) evaluation, but presents distinct new challenges. It requires models (1) to utilize the contexts at a deeper level, rather than resorting to simple copying and pasting; (2) to navigate through long streams of evolving topics and tasks, proxying the complexities and dynamism of contexts in real-world scenarios. Additionally, Task Haystack inherits the controllability of NIAH, providing model developers with tools and visualizations to identify model vulnerabilities effectively. We benchmark 14 long-context LMs using Task Haystack, finding that frontier models like GPT-4o still struggle with the setting, failing on 15% of cases on average. Most open-weight models further lack behind by a large margin, with failure rates reaching up to 61%. In our controlled analysis, we identify factors such as distraction and recency bias as contributors to these failure cases. Further, performance declines when task instructions are paraphrased at test time or when ICL demonstrations are repeated excessively, raising concerns about the robustness, instruction understanding, and true context utilization of long-context LMs.

cs.CL↗

arXiv2024

Prompt Engineering a Prompt Engineer

Prompt engineering is a challenging yet crucial task for optimizing the performance of large language models on customized tasks. It requires complex reasoning to examine the model's errors, hypothesize what is missing or misleading in the current prompt, and communicate the task with clarity. While recent works indicate that large language models can be meta-prompted to perform automatic prompt engineering, we argue that their potential is limited due to insufficient guidance for complex reasoning in the meta-prompt. We fill this gap by infusing into the meta-prompt three key components: detailed descriptions, context specification, and a step-by-step reasoning template. The resulting method, named PE2, exhibits remarkable versatility across diverse language tasks. It finds prompts that outperform "let's think step by step" by 6.3% on MultiArith and 3.1% on GSM8K, and outperforms competitive baselines on counterfactual tasks by 6.9%. Further, we show that PE2 can make targeted and highly specific prompt edits, rectify erroneous prompts, and induce multi-step plans for complex tasks.

cs.CL↗

arXiv2023

How Predictable Are Large Language Model Capabilities? A Case Study on BIG-bench

We investigate the predictability of large language model (LLM) capabilities: given records of past experiments using different model families, numbers of parameters, tasks, and numbers of in-context examples, can we accurately predict LLM performance on new experiment configurations? Answering this question has practical implications for LLM users (e.g., deciding which models to try), developers (e.g., prioritizing evaluation on representative tasks), and the research community (e.g., identifying hard-to-predict capabilities that warrant further investigation). We study the performance prediction problem on experiment records from BIG-bench. On a random train-test split, an MLP-based predictor achieves an $R^2$ score greater than 95%, indicating the presence of learnable patterns within the experiment records. We then formulate the problem of searching for "small-bench," an informative subset of BIG-bench tasks from which the performance on the full set can be maximally recovered. We find a subset as informative as BIG-bench Hard for evaluating new model families, while being $3\times$ smaller. Additionally, we find competitive subsets by clustering task representations learned by our MLP-based predictor and selecting tasks close to cluster centroids, highlighting the importance of task diversity in constructing "small-bench."

cs.CL↗

arXiv2023

Estimating Large Language Model Capabilities without Labeled Test Data

Large Language Models (LLMs) have the impressive ability to perform in-context learning (ICL) from only a few examples, but the success of ICL varies widely from task to task. Thus, it is important to quickly determine whether ICL is applicable to a new task, but directly evaluating ICL accuracy can be expensive in situations where test data is expensive to annotate -- the exact situations where ICL is most appealing. In this paper, we propose the task of ICL accuracy estimation, in which we predict the accuracy of an LLM when doing in-context learning on a new task given only unlabeled test data for that task. To perform ICL accuracy estimation, we propose a method that trains a meta-model using LLM confidence scores as features. We compare our method to several strong accuracy estimation baselines on a new benchmark that covers 4 LLMs and 3 task collections. The meta-model improves over all baselines across 8 out of 12 settings and achieves the same estimation performance as directly evaluating on 40 collected labeled test examples per task. At the same time, no existing approach provides an accurate and reliable ICL accuracy estimation in every setting, highlighting the need for better ways to measure the uncertainty of LLM predictions.

cs.CL↗

arXiv2022

Eliciting and Understanding Cross-Task Skills with Task-Level Mixture-of-Experts

Recent works suggest that transformer models are capable of multi-tasking on diverse NLP tasks and adapting to new tasks efficiently. However, the potential of these multi-task models may be limited as they use the same set of parameters for all tasks. In contrast, humans tackle tasks in a more flexible way, by making proper presumptions on what skills and knowledge are relevant and executing only the necessary computations. Inspired by this, we propose to use task-level mixture-of-expert models, which has a collection of transformer layers (i.e., experts) and a router component that chooses from these experts dynamically and flexibly. We find that these models help improve the average performance gain (ARG) metric by 2.6% when adapting to unseen tasks in the few-shot setting and by 5.6% in the zero-shot generalization setting. Further, we show that the learned routing decisions partly rediscover human categorization of NLP tasks -- certain experts are strongly associated with extractive tasks, some with classification tasks, and some with tasks requiring world knowledge.

cs.CL↗

arXiv2022

Sparse Distillation: Speeding Up Text Classification by Using Bigger Student Models

Distilling state-of-the-art transformer models into lightweight student models is an effective way to reduce computation cost at inference time. The student models are typically compact transformers with fewer parameters, while expensive operations such as self-attention persist. Therefore, the improved inference speed may still be unsatisfactory for real-time or high-volume use cases. In this paper, we aim to further push the limit of inference speed by distilling teacher models into bigger, sparser student models -- bigger in that they scale up to billions of parameters; sparser in that most of the model parameters are n-gram embeddings. Our experiments on six single-sentence text classification tasks show that these student models retain 97% of the RoBERTa-Large teacher performance on average, and meanwhile achieve up to 600x speed-up on both GPUs and CPUs at inference time. Further investigation reveals that our pipeline is also helpful for sentence-pair classification tasks, and in domain generalization settings.

cs.CL↗

arXiv2021

Refining Language Models with Compositional Explanations

Pre-trained language models have been successful on text classification tasks, but are prone to learning spurious correlations from biased datasets, and are thus vulnerable when making inferences in a new domain. Prior work reveals such spurious patterns via post-hoc explanation algorithms which compute the importance of input features. Further, the model is regularized to align the importance scores with human knowledge, so that the unintended model behaviors are eliminated. However, such a regularization technique lacks flexibility and coverage, since only importance scores towards a pre-defined list of features are adjusted, while more complex human knowledge such as feature interaction and pattern generalization can hardly be incorporated. In this work, we propose to refine a learned language model for a target domain by collecting human-provided compositional explanations regarding observed biases. By parsing these explanations into executable logic rules, the human-specified refinement advice from a small set of explanations can be generalized to more training examples. We additionally introduce a regularization term allowing adjustments for both importance and interaction of features to better rectify model behavior. We demonstrate the effectiveness of the proposed approach on two text classification tasks by showing improved performance in target domain as well as improved model fairness after refinement.

cs.CL↗

arXiv2021

CrossFit: A Few-shot Learning Challenge for Cross-task Generalization in NLP

Humans can learn a new language task efficiently with only few examples, by leveraging their knowledge obtained when learning prior tasks. In this paper, we explore whether and how such cross-task generalization ability can be acquired, and further applied to build better few-shot learners across diverse NLP tasks. We introduce CrossFit, a problem setup for studying cross-task generalization ability, which standardizes seen/unseen task partitions, data access during different learning stages, and the evaluation protocols. To instantiate different seen/unseen task partitions in CrossFit and facilitate in-depth analysis, we present the NLP Few-shot Gym, a repository of 160 diverse few-shot NLP tasks created from open-access NLP datasets and converted to a unified text-to-text format. Our analysis reveals that the few-shot learning ability on unseen tasks can be improved via an upstream learning stage using a set of seen tasks. We also observe that the selection of upstream learning tasks can significantly influence few-shot performance on unseen tasks, asking further analysis on task similarity and transferability.

cs.CL↗

arXiv2021

On the Influence of Masking Policies in Intermediate Pre-training

Current NLP models are predominantly trained through a two-stage "pre-train then fine-tune" pipeline. Prior work has shown that inserting an intermediate pre-training stage, using heuristic masking policies for masked language modeling (MLM), can significantly improve final performance. However, it is still unclear (1) in what cases such intermediate pre-training is helpful, (2) whether hand-crafted heuristic objectives are optimal for a given task, and (3) whether a masking policy designed for one task is generalizable beyond that task. In this paper, we perform a large-scale empirical study to investigate the effect of various masking policies in intermediate pre-training with nine selected tasks across three categories. Crucially, we introduce methods to automate the discovery of optimal masking policies via direct supervision or meta-learning. We conclude that the success of intermediate pre-training is dependent on appropriate pre-train corpus, selection of output format (i.e., masked spans or full sentence), and clear understanding of the role that MLM plays for the downstream task. In addition, we find our learned masking policies outperform the heuristic of masking named entities on TriviaQA, and policies learned from one task can positively transfer to other tasks in certain cases, inviting future research in this direction.

cs.CL↗

arXiv2021

Learning to Generate Task-Specific Adapters from Task Description

Pre-trained text-to-text transformers such as BART have achieved impressive performance across a range of NLP tasks. Recent study further shows that they can learn to generalize to novel tasks, by including task descriptions as part of the source sequence and training the model with (source, target) examples. At test time, these fine-tuned models can make inferences on new tasks using the new task descriptions as part of the input. However, this approach has potential limitations, as the model learns to solve individual (source, target) examples (i.e., at the instance level), instead of learning to solve tasks by taking all examples within a task as a whole (i.e., at the task level). To this end, we introduce Hypter, a framework that improves text-to-text transformer's generalization ability to unseen tasks by training a hypernetwork to generate task-specific, light-weight adapters from task descriptions. Experiments on ZEST dataset and a synthetic SQuAD dataset demonstrate that Hypter improves upon fine-tuning baselines. Notably, when using BART-Large as the main network, Hypter brings 11.3% comparative improvement on ZEST dataset.

cs.CL↗

arXiv2021

Semi-Automated Protocol Disambiguation and Code Generation

For decades, Internet protocols have been specified using natural language. Given the ambiguity inherent in such text, it is not surprising that protocol implementations have long exhibited bugs. In this paper, we apply natural language processing (NLP) to effect semi-automated generation of protocol implementations from specification text. Our system, SAGE, can uncover ambiguous or under-specified sentences in specifications; once these are clarified by the spec author, SAGE can generate protocol code automatically. Using SAGE, we discover 5 instances of ambiguity and 6 instances of under-specification in the ICMP RFC; after clarification, SAGE is able to automatically generate code that interoperates perfectly with Linux implementations. We show that SAGE generalizes to BFD, IGMP, and NTP. We also find that SAGE supports many of the conceptual components found in key protocols, suggesting that, with some additional machinery, SAGE may be able to generalize to TCP and BGP.

cs.NI↗

arXiv2021

Studying Strategically: Learning to Mask for Closed-book QA

Closed-book question-answering (QA) is a challenging task that requires a model to directly answer questions without access to external knowledge. It has been shown that directly fine-tuning pre-trained language models with (question, answer) examples yields surprisingly competitive performance, which is further improved upon through adding an intermediate pre-training stage between general pre-training and fine-tuning. Prior work used a heuristic during this intermediate stage, whereby named entities and dates are masked, and the model is trained to recover these tokens. In this paper, we aim to learn the optimal masking strategy for the intermediate pre-training stage. We first train our masking policy to extract spans that are likely to be tested, using supervision from the downstream task itself, then deploy the learned policy during intermediate pre-training. Thus, our policy packs task-relevant knowledge into the parameters of a language model. Our approach is particularly effective on TriviaQA, outperforming strong heuristics when used to pre-train BART.

cs.CL↗