Muyu He

Muyu He, Yuchen Liu, Qingya Huang et al.

The success of the transformer architecture as the backbone of modern LLMs is in large part due to its use of attention layers. An attention layer follows the standard neural network paradigm: it takes the residual stream as input and thereby produces context-dependent query, key, and value vectors. However, we find that model performance meaningfully improves when deeper layers learn only a context-free value vector to preserve the original token information, without drawing on any context from the residual stream. When the model has access to this context-free value vector, adding back the context-dependent component provides little additional benefit for aggregate benchmark performance. Such context-free value vectors can be stored as sparse model parameters, eliminating the need to recompute or persistently cache these values. Through systematic ablations on the key design choices for such context-free value vectors, we propose Bank of Values (BoV), a new way of computing value vectors in attention by learning a lookup table of token-specific value vectors for each of the last third of layers. Across 135M and 780M models, BoV improves validation loss over standard attention and, at 780M, the average score across 21 benchmarks, matching the previous best method that adds token information to the value vector with less compute and memory.

Muyu He, Adit Jain, Anand Kumar et al.

As LLM agents tackle increasingly complex tasks, a critical question is whether they can maintain strategic coherence over long horizons: planning under uncertainty, learning from delayed feedback, and adapting when early mistakes compound. We introduce $\texttt{YC-Bench}$, a benchmark that evaluates these capabilities by tasking an agent with running a simulated startup over a one-year horizon spanning hundreds of turns. The agent must manage employees, select task contracts, and maintain profitability in a partially observable environment where adversarial clients and growing payroll create compounding consequences for poor decisions. We evaluate 12 models, both proprietary and open source, across 3 seeds each. Only three models consistently surpass the starting capital of \$200K, with Claude Opus 4.6 achieving the highest average final funds at \$1.27 M, followed by GLM-5 at \$1.21 M at 11$\times$ lower inference cost. Scratchpad usage, the sole mechanism for persisting information across context truncation, is the strongest predictor of success, and adversarial client detection is the primary failure mode, accounting for $47\%$ of bankruptcies. Our analysis reveals that frontier models still fail through distinct failure modes such as over-parallelization, demonstrating the capability gaps for long-horizon performance. $\texttt{YC-Bench}$ is open-source, reproducible, and configurable.

Muyu He, Anand Kumar, Tsach Mackey et al. · salesforce

Despite rapid progress in building conversational AI agents, robustness is still largely untested. Small shifts in user behavior, such as being more impatient, incoherent, or skeptical, can cause sharp drops in agent performance, revealing how brittle current AI agents are. Today's benchmarks fail to capture this fragility: agents may perform well under standard evaluations but degrade spectacularly in more realistic and varied settings. We address this robustness testing gap by introducing TraitBasis, a lightweight, model-agnostic method for systematically stress testing AI agents. TraitBasis learns directions in activation space corresponding to steerable user traits (e.g., impatience or incoherence), which can be controlled, scaled, composed, and applied at inference time without any fine-tuning or extra data. Using TraitBasis, we extend $τ$-Bench to $τ$-Trait, where user behaviors are altered via controlled trait vectors. We observe on average a 2%-30% performance degradation on $τ$-Trait across frontier models, highlighting the lack of robustness of current AI agents to variations in user behavior. Together, these results highlight both the critical role of robustness testing and the promise of TraitBasis as a simple, data-efficient, and compositional tool. By powering simulation-driven stress tests and training loops, TraitBasis opens the door to building AI agents that remain reliable in the unpredictable dynamics of real-world human interactions. We have open-sourced $τ$-Trai across four domains: airline, retail, telecom, and telehealth, so the community can systematically QA their agents under realistic, behaviorally diverse intents and trait scenarios: https://github.com/collinear-ai/tau-trait.

Muyu He, Muhammad Ali Shafique, Anand Kumar et al. · salesforce

Distilling the thinking traces of a Large Language Model (LLM) with reasoning capabilities into a smaller model has been proven effective. Yet, there is a scarcity of work done on how model performances scale with the quantity of distillation data. In this work, we study the scaling trend of distilling competitive coding skills on two small non-reasoning LLMs. We validate the hypothesis that there is a $\textit{valley of code reasoning}$: downstream performance on competitive coding first drops as data quantity increases, then it steadily increases in a sharper-than-log-linear fashion. Having identified the trend, we further fine-tune the models at two different distillation stages on the same data to ground conclusions on their respective learning phases. We learn that across stages in the low and medium-low data regimes, small models benefit significantly from easier coding questions than from harder ones. We also find that, surprisingly, the correctness of outputs in training data makes no difference to distillation outcomes. Our work represents a step forward in understanding the training dynamics of code reasoning distillation outside intuition

Muyu He, Yuxi Zheng, Yuchen Liu et al.

The advancement of vision language models (VLMs) has empowered embodied agents to accomplish simple multimodal planning tasks, but not long-horizon ones requiring long sequences of actions. In text-only simulations, long-horizon planning has seen significant improvement brought by repositioning the role of LLMs. Instead of directly generating action sequences, LLMs translate the planning domain and problem into a formal planning language like the Planning Domain Definition Language (PDDL), which can call a formal solver to derive the plan in a verifiable manner. In multimodal environments, research on VLM-as-formalizer remains scarce, usually involving gross simplifications such as predefined object vocabulary or overly similar few-shot examples. In this work, we present a suite of five VLM-as-formalizer pipelines that tackle one-shot, open-vocabulary, and multimodal PDDL formalization. We evaluate those on an existing benchmark while presenting another two that for the first time account for planning with authentic, multi-view, and low-quality images. We conclude that VLM-as-formalizer greatly outperforms end-to-end plan generation. We reveal the bottleneck to be vision rather than language, as VLMs often fail to capture an exhaustive set of necessary object relations. While generating intermediate, textual representations such as captions or scene graphs partially compensate for the performance, their inconsistent gain leaves headroom for future research directions on multimodal planning formalization.

Yuan Yuan, Muyu He, Muhammad Adil Shahid et al.

This paper introduces TurnaboutLLM, a novel framework and dataset for evaluating the deductive reasoning abilities of Large Language Models (LLMs) by leveraging the interactive gameplay of detective games Ace Attorney and Danganronpa. The framework tasks LLMs with identifying contradictions between testimonies and evidences within long narrative contexts, a challenging task due to the large answer space and diverse reasoning types presented by its questions. We evaluate twelve state-of-the-art LLMs on the dataset, hinting at limitations of popular strategies for enhancing deductive reasoning such as extensive thinking and Chain-of-Thought prompting. The results also suggest varying effects of context size, the number of reasoning step and answer space size on model performance. Overall, TurnaboutLLM presents a substantial challenge for LLMs' deductive reasoning abilities in complex, narrative-rich environments.

Xingyu Fu, Muyu He, Yujie Lu et al.

We present a novel task and benchmark for evaluating the ability of text-to-image(T2I) generation models to produce images that align with commonsense in real life, which we call Commonsense-T2I. Given two adversarial text prompts containing an identical set of action words with minor differences, such as "a lightbulb without electricity" v.s. "a lightbulb with electricity", we evaluate whether T2I models can conduct visual-commonsense reasoning, e.g. produce images that fit "the lightbulb is unlit" vs. "the lightbulb is lit" correspondingly. Commonsense-T2I presents an adversarial challenge, providing pairwise text prompts along with expected outputs. The dataset is carefully hand-curated by experts and annotated with fine-grained labels, such as commonsense type and likelihood of the expected outputs, to assist analyzing model behavior. We benchmark a variety of state-of-the-art (sota) T2I models and surprisingly find that, there is still a large gap between image synthesis and real life photos--even the DALL-E 3 model could only achieve 48.92% on Commonsense-T2I, and the stable diffusion XL model only achieves 24.92% accuracy. Our experiments show that GPT-enriched prompts cannot solve this challenge, and we include a detailed analysis about possible reasons for such deficiency. We aim for Commonsense-T2I to serve as a high-quality evaluation benchmark for T2I commonsense checking, fostering advancements in real life image generation.