Stepwise Penalization for Length-Efficient Chain-of-Thought Reasoning

Large reasoning models improve with more test-time computation, but often overthink, producing unnecessarily long chains-of-thought that raise cost without improving accuracy. Prior reinforcement learning approaches typically rely on a single outcome reward with trajectory-level length penalties, which cannot distinguish essential from redundant reasoning steps and therefore yield blunt compression. Although recent work incorporates step-level signals, such as offline pruning, supervised data construction, or verifier-based intermediate rewards, reasoning length is rarely treated as an explicit step-level optimization objective during RL. We propose Step-wise Adaptive Penalization (SWAP), a fine-grained framework that allocates length reduction across steps based on intrinsic contribution. We estimate step importance from the model's on-policy log-probability improvement toward the correct answer, then treat excess length as a penalty mass redistributed to penalize low-importance steps more heavily while preserving high-importance reasoning. We optimize with a unified outcome-process advantage within group-relative policy optimization. Extensive experiments demonstrate that SWAP reduces reasoning length by 64.3% on average while improving accuracy by 5.7% relative to the base model.

"Meet My Sidekick!": Effects of Separate Identities and Control of a Single Robot in HRI

The presentation of a robot's capability and identity directly influences a human collaborator's perception and implicit trust in the robot. Unlike humans, a physical robot can simultaneously present different identities and have them reside and control different parts of the robot. This paper presents a novel study that investigates how users perceive a robot where different robot control domains (head and gripper) are presented as independent robots. We conducted a mixed design study where participants experienced one of three presentations: a single robot, two agents with shared full control (co-embodiment), or two agents with split control across robot control domains (split-embodiment). Participants underwent three distinct tasks -- a mundane data entry task where the robot provides motivational support, an individual sorting task with isolated robot failures, and a collaborative arrangement task where the robot causes a failure that directly affects the human participant. Participants perceived the robot as residing in the different control domains and were able to associate robot failure with different identities. This work signals how future robots can leverage different embodiment configurations to obtain the benefit of multiple robots within a single body.

WebCoach: Self-Evolving Web Agents with Cross-Session Memory Guidance

Multimodal LLM-powered agents have recently demonstrated impressive capabilities in web navigation, enabling agents to complete complex browsing tasks across diverse domains. However, current agents struggle with repetitive errors and lack the ability to learn from past experiences across sessions, limiting their long-term robustness and sample efficiency. We introduce WebCoach, a model-agnostic self-evolving framework that equips web browsing agents with persistent cross-session memory, enabling improved long-term planning, reflection, and continual learning without retraining. WebCoach consists of three key components: (1) a WebCondenser, which standardizes raw navigation logs into concise summaries; (2) an External Memory Store, which organizes complete trajectories as episodic experiences; and (3) a Coach, which retrieves relevant experiences based on similarity and recency, and decides whether to inject task-specific advice into the agent via runtime hooks. This design empowers web agents to access long-term memory beyond their native context window, improving robustness in complex browsing tasks. Moreover, WebCoach achieves self-evolution by continuously curating episodic memory from new navigation trajectories, enabling agents to improve over time without retraining. Evaluations on the WebVoyager benchmark demonstrate that WebCoach consistently improves the performance of browser-use agents across three different LLM backbones. With a 38B model, it increases task success rates from 47% to 61% while reducing or maintaining the average number of steps. Notably, smaller base models with WebCoach achieve performance comparable to the same web agent using GPT-4o.

Stabilizing Reinforcement Learning for Honesty Alignment in Language Models on Deductive Reasoning

Reinforcement learning with verifiable rewards (RLVR) has recently emerged as a promising framework for aligning language models with complex reasoning objectives. However, most existing methods optimize only for final task outcomes, leaving models vulnerable to collapse when negative rewards dominate early training. This challenge is especially pronounced in honesty alignment, where models must not only solve answerable queries but also identify when conclusions cannot be drawn from the given premises. Deductive reasoning provides an ideal testbed because it isolates reasoning capability from reliance on external factual knowledge. To investigate honesty alignment, we curate two multi-step deductive reasoning datasets from graph structures, one for linear algebra and one for logical inference, and introduce unanswerable cases by randomly perturbing an edge in half of the instances. We find that GRPO, with or without supervised fine tuning initialization, struggles on these tasks. Through extensive experiments across three models, we evaluate stabilization strategies and show that curriculum learning provides some benefit but requires carefully designed in distribution datasets with controllable difficulty. To address these limitations, we propose Anchor, a reinforcement learning method that injects ground truth trajectories into rollouts, preventing early training collapse. Our results demonstrate that this method stabilizes learning and significantly improves the overall reasoning performance, underscoring the importance of training dynamics for enabling reliable deductive reasoning in aligned language models.

A Reasoning-Focused Legal Retrieval Benchmark

As the legal community increasingly examines the use of large language models (LLMs) for various legal applications, legal AI developers have turned to retrieval-augmented LLMs ("RAG" systems) to improve system performance and robustness. An obstacle to the development of specialized RAG systems is the lack of realistic legal RAG benchmarks which capture the complexity of both legal retrieval and downstream legal question-answering. To address this, we introduce two novel legal RAG benchmarks: Bar Exam QA and Housing Statute QA. Our tasks correspond to real-world legal research tasks, and were produced through annotation processes which resemble legal research. We describe the construction of these benchmarks and the performance of existing retriever pipelines. Our results suggest that legal RAG remains a challenging application, thus motivating future research.

A Dataset and Benchmark for Automatically Answering and Generating Machine Learning Final Exams

Can a machine learn machine learning? We propose to answer this question using the same criteria we use to answer a similar question: can a human learn machine learning? We automatically answer MIT final exams in Introduction to Machine Learning at a human level. The course is a large undergraduate class with around five hundred students each semester. Recently, program synthesis and few-shot learning solved university-level problem set questions in mathematics and STEM courses at a human level. In this work, we solve questions from final exams that differ from problem sets in several ways: the questions are longer, have multiple parts, are more complicated, and span a broader set of topics. We provide a new dataset and benchmark of questions from eight MIT Introduction to Machine Learning final exams between Fall 2017 and Spring 2022 and provide code for automatically answering these questions and generating new questions. We perform ablation studies comparing zero-shot learning with few-shot learning, chain-of-thought prompting, GPT-3 pre-trained on text and Codex fine-tuned on code on a range of machine learning topics and find that few-shot learning methods perform best. We make our data and code publicly available for the machine learning community.