Science & Mathematics

Long-horizon robotic operation requires memory that scales with essential information rather than execution time. We present a parameter-space memory framework where each episode is stored as a separable object. A neural implicit representation is fitted to episode video from a shared, meta-learned initialization; the resulting weight delta constitutes the memory. Crucially, this delta remains a fixed size regardless of whether the episode spans a hundred or a hundred thousand frames. A learned projector maps the weight delta into token embeddings, conditioning a language model for question answering and a decoder for joint command regression, without frame retrieval at query time. On RoboMME, the language model exceeds chance across all question types, and the decoder achieves a mean $R^2$ of 0.51. Replacing the delta with matched noise drops performance by 11.4 points, while random-origin weights are non-decodable, proving the meta-learned initialization is essential for embedding episode content. By representing individual experiences as fixed-size weight deltas in a common parameter space, this approach decouples memory cost from time horizon and provides a basis for scalable streaming memory.

Online reinforcement learning (RL) has achieved significant success in decision-making tasks, but modern online RL remains highly sensitive to the quality of training experience. Understanding the contribution of individual training instances is therefore central to data-centric RL, with applications ranging from experience selection to policy interpretation. Existing data attribution methods for online RL primarily focus on local influence within a single training round, overlooking the non-local data influence across multiple training rounds. In this work, we formalize data attribution in online RL through a trajectory-level leave-one-out quantity, RL-LOO, which measures how one training sample influences a downstream target after subsequent online updates have taken place. We then derive RL-IF, a first-order estimator that propagates data influence through both optimization effects and policy-induced sampling effects, and provide theoretical guarantees on its approximation accuracy under smoothness assumptions. Our analysis further disentangles these two effects and shows both theoretically and empirically that the full RL-IF estimator can often be well approximated using the optimization effect alone. Building on this observation, we develop RINSE, a practical data filtering method for online RL. Experiments on standard control benchmarks and an RLHF-style toxicity-mitigation setting show that RINSE improves training efficiency and final performance over standard PPO and local attribution baselines. These results suggest that RL-IF provides a practical and principled framework for understanding and improving online RL.

LLM judges increasingly mediate benchmark evaluation, model selection, data curation, and training. Their decisions are useful only when each judge responds reliably to changes affecting its scoring criterion. Existing methods for reliability testing perturb inputs with known expected effects on judge scores. Prominent LLM-judge testing harnesses combine human-authored perturbation recipes with synthetic test generation, leaving scorer-specific recipe ideation outside the automated pipeline. This limits their use to fixed perturbations and pre-assumed scorer types. We introduce SANE, an agent-based framework that encodes behavioral requirements for stochastic evaluators as reusable agent instructions. Given a scorer specification, implementation, and representative inputs, the pipeline first identifies scorer-specific capabilities, then automatically derives perturbation recipes and expected score relations for each. SANE also defines a scorer-independent validation framework for generated judge tests. It determines whether each recipe agrees with scorer intent and whether each instantiated case realizes the recipe and supports its expected relation. Because SANE generates a recipe inventory tailored to each scorer, the number of aligned recipes grows as scorers are added. Across five domains, SANE produces 3.7 times as many validated test families as the Judge Reliability Harness (JRH) at comparable end-to-end validity and detects scorer deficiencies in every domain, including professionalism and toxicity, where JRH detects none. On faithfulness, SANE achieves strong human alignment in judge ranking: its test-suite scores rank 11 judge models consistently with human-labeled FaithBench performance (Spearman rho = 0.852), while JRH does not (rho = 0.100).

Autonomous agents can increasingly propose ideas, modify code, and run experiments, creating an opportunity to accelerate scientific discovery using teams of collaborative AI researchers. To escape local minima in the discovery, we propose to encourage orthogonal exploration by constructing the team using agents with varying expertise. However, realizing this design requires more than diverse backgrounds. When experiments are costly, a team can test only a limited number of ideas, thereby creating the need to balance exploration and exploitation when selecting a subset of researchers to propose and execute new experiments. Furthermore, each result should shape what the team tries next. We therefore ask how AI researchers with different scientific backgrounds can build on one another's findings and sustain progress under a limited experimental budget. To address these challenges, we build a team of persistent AI researchers, each grounded in a different discipline, who keep private research histories, share experimental evidence, and receive research opportunities through adaptive allocation. We demonstrate the effectiveness of this design by deploying the team to search for linear attention architectures, where it outperforms GDN2 at 1.3B scale on language modeling, common-sense reasoning, and synthetic and real-world retrieval tasks. In controlled comparisons, removing any single organizational component from our design yields markedly weaker discoveries: without disciplinary backgrounds, the search plateaus early; without private histories, researchers make less progress despite identical shared evidence; and uniform selection falls short of adaptive allocation within the same budget. These results suggest that sustained autonomous discovery depends not only on stronger agents, but also on who explores, what they remember, and how research effort is allocated.

Selecting intermediate features from a frozen diffusion model for discriminative tasks is non-trivial because the useful signal is distributed across the denoising trajectory, network depth, and channels. In practice, this requires choosing a timestep, a block, and a channel subset. Existing methods either fix these choices by hand or optimize only a few extraction hyperparameters, most often the timestep, then reuse the same rule across inputs and budgets. We introduce Policy-Mask, a budget-conditioned policy trained on the downstream objective that jointly selects all three under a prescribed channel budget. Without normalization, the channel-score contribution grows with the number of selected channels and dominates the timestep and block updates. We normalize it by subset cardinality, keeping the three policy heads on comparable scales and stabilizing joint training. At matched budgets, Policy-Mask improves over fixed extraction and feature-filter baselines across segmentation, classification, and semantic correspondence, reaching 38.9 mIoU on ADE20K, 75.4% top-1 on CIFAR-100 at 128 dimensions, and 76.3 [email protected] on SPair-71k. Component-wise ablations show that block selection materially affects downstream accuracy, and the highest recorded channel-budget score occurs at 256 retained channels, identifying a practical operating point for compact diffusion representations. We further adapt Policy-Mask to diffusion transformers, including PixArt-$Σ$ and SD3.5 MMDiT, by replacing U-Net channels with image-token hidden dimensions.

Many social evaluations for multiagent systems study behaviors as they emerge from environments instead of systematically intervening in the environment to elicit such behaviors. We introduce Fig (Finding Interactions in Games), an agentic system that performs targeted elicitation interventions on game environments. It takes a natural language behavior description, decomposes it into sub-behaviors, generates new social games or selects existing ones, and executes single or combined interventions to the games' engine, opponents, instructions, and context. To evaluate model behavior, it combines built-in mechanical checks and synthetic rubric judgments. We validate the interventions, generated games, and evaluation mechanisms via human review. Models attempt the target behavior in 45.2% of episodes on average for valid scenarios generated by Fig for six test behaviors -- cooperation, exploitation, deception, covert coordination, betrayal, and sabotage. For game-behavior pairs with available controls, Fig's interventions increase attempt rates by 10.9% on average compared to matched control games. Fig shows that multiagent environments can serve as programmable test beds for intentional, controllable, and reproducible behavior evaluation. Code is available at https://github.com/fleet-ai/fig.

Adapting language models to new domains during post-training risks degrading previously acquired capabilities, a phenomenon commonly known as catastrophic forgetting. Existing mitigation methods largely treat forgetting as a consequence of excessive model drift. They therefore constrain updates by preserving old outputs, replaying old data, or regularizing changes to parameters and representations so that the updated model remains close to its previous state. We take a different view: Rather than pulling the updated model back toward its old state, we translate the new domain into the model's pre-existing knowledge space, making unfamiliar knowledge learnable through familiar procedures. Therefore, we propose Skill Schema Transport (SST), a post-training framework that transports new knowledge into the model's existing skill space. SST abstracts each example type into a skill schema, preserving its essential operations, dependencies, and verification steps while removing domain-specific surface forms. By recasting new examples as instances of reusable procedures, SST turns knowledge update from trajectory memorization into procedural reuse. In continual domain adaptation settings, SST consistently reduces forgetting while improving generalization to related tasks. Across two 3B instruction-tuned backbones, SST improves average final performance on streamed domains by 2.53 to 5.75 points over sequential post-training. It also lowers max-drop forgetting in all evaluated settings, with the largest reduction from 42.41 to 7.10 points on the Qwen-3B MATH stream, while substantially recovering the OOD degradation induced by sequential updates.

Trajectory demonstrations show robots how to behave but not why, leaving them to guess which features of the task actually matter. Incorrect guesses produce misaligned internal representations and lead to incorrect robot behaviors. Because task features are often semantically defined, we argue that natural language is better suited for specifying features rather than relying on demonstrations. Although existing representation editing methods have proposed using language to specify new features, only allowing additive edits produces feature sets that are often over-specified or too narrowly scoped, especially in sequential editing tasks. We propose the Flexible Feature Editing Pipeline (FLEX), which uses language models to iteratively refine a robot's internal representation of task features from natural language corrections via three complementary operations: Add missing features, Generalize existing specific features, and Mask irrelevant features. We evaluate FLEX on 2D navigation scenarios, 7DoF Franka tabletop manipulation tasks, and a real-world study using naturalistic human language corrections and real-robot demonstrations. Comparisons against single-operation and unscaffolded baselines show that all three operations are necessary in interactive scenarios, and that FLEX achieves consistently higher normalized reward compared to other feature editing approaches while maintaining compact, transferable representations. Project available at https://mit-clear-lab.github.io/FLEX/

Cross-tokenizer logits distillation enables knowledge transfer across model families, allowing students to learn from stronger or more specialized teachers. However, differences in vocabulary and segmentation make teacher and student next-token distributions difficult to align. Existing logits-based methods are easy to integrate into standard distillation pipelines, but often rely on restrictive heuristics or tokenizer-pair-specific choices. They may restrict supervision to shared vocabulary subsets, require manually selecting different objectives for different tokenizer pairs, or match only the observed chunk and its complement rather than individual alternative continuations. We introduce Common-Support Distillation (CSD), a unified logits-based framework that constructs a shared support between the teacher and student at each aligned input chunk and matches the induced distributions with a single objective. For one-to-one alignments, CSD uses explicit common-token pairs as the shared support. For other alignments, it constructs a virtual support whose probabilities are computed using the autoregressive chain rule. A residual event captures probability mass outside the explicit support, ensuring that both induced distributions remain normalized. Across four teacher-student settings comprising Llama-3.2-1B and Gemma3-1B-PT students with Qwen and Phi teachers, CSD consistently achieves the highest average performance among the evaluated cross-tokenizer methods, outperforming the strongest alternative by up to 1.8 points and surpassing both continued pretraining and same-tokenizer distillation. For Llama-3.1-8B, CSD improves average accuracy over the strongest evaluated baseline by up to 4.2 points for the base student and 1.6 points for the instruction-tuned student.

Refusal directions have been shown to transfer cross-lingually with near-perfect effectiveness across safety-aligned languages (Wang et al., 2025), reportedly regardless of resource level. However, the only truly low-resourced language evaluated is Yoruba, which is strongly misaligned and shows limited transfer, leaving open how refusal directions behave for other low-resourced languages. We extend PolyRefuse to three additional low-resourced languages (Belarusian, Bashkir, and Tajik) and replicate the pipeline of (Wang et al., 2025) on Qwen2.5-14B-Instruct. We find that universality is graded rather than binary: directions from safety-aligned and partially-aligned languages transfer broadly, while those from strongly misaligned languages fail to reliably modulate refusal behavior on any target tested, including other misaligned languages, once ceiling effects are accounted for. Geometric alignment and behavioral transfer are further shown to be partially dissociated. These findings complicate the universality claim and suggest that users of the least-resourced, most strongly misaligned languages may face weaker safety guarantees.

Flow matching learns a generative velocity field by regression, but under distribution drift a field must be deployed before the current data arrive. We study this predictive problem for a fixed, possibly nonconvex, neural class, with dynamic regret against the best represented field of each round. We propose DyProd-FM, which refits flow matching on dyadic historical prefixes and selects among the fits with a sleeping Prod master. Without knowing any drift budget, it attains regret $\widetilde{O}(\min\{N_ω+1,1+\sqrt{TΩ}\})$ in the relative velocity risk variation $Ω$ and $\widetilde{O}(\min\{N_μ+1,1+T^{1/3}V_μ^{2/3}\})$ in the target-law variation $V_μ=\sum_t W_2(μ_{t+1},μ_t)$, up to approximation and fitting costs. Matching lower bounds for fixed classes show that both rates are optimal in $T$ and in the budgets up to logarithmic factors. Prior optimal dynamic-regret guarantees rely on a convex decision set with strongly convex or exp-concave losses, whereas our class may be nonconvex and no curvature in its parameters is assumed. The target-law rate rests on a second-order link between the two variations, which the base noise of flow matching provides.

Synthesised data is a common strategy to address the lack of annotated handwritten music scores in Optical Music Recognition (OMR). Smashcima is a recent synthesis framework that assembles pages of handwritten scores by sampling glyph images from MUSCIMA++ and positioning them with hard-coded rules. In this work we ask whether these rules are grounded empirically and whether a learned model for context-dependent positioning can further improve synthesis quality. We extract inter-glyph distributions from around 285 thousand annotated relationships across MUSCIMA++ and OmniOMR, finding that several key parameters deviate systematically from real handwriting. Replacing these constants with empirical means reduces positioning RMSE by up to 60%. We then train a Graph Neural Network (GNN) on the Music Notation Graph structure to predict context-dependent offsets, achieving an additional 50 to 85% RMSE reduction on stems and beams by modelling how neighbouring glyphs constrain placement. Cross-dataset analysis shows that spacing differences between the two corpora are explained by note density (3.3), not style, and that duration-proportional spacing does not hold in handwriting (R2 = 0.001). In downstream evaluation, a YOLOv8 detector pretrained on GNN-positioned synthetic data achieves +30% mAP50 over the baseline and +11% over empirical means when fine-tuned on 200 real handwritten staves, showing that positioning accuracy matters for downstream OMR.

We consider cooperative and iterative tasks involving multiple agents whose interactions are modeled as a dynamic game with shared constraints. We propose a general learning framework applicable to nonlinear and nonconvex dynamic games that uses previously executed feasible trajectories to construct a sampled joint safe set and player-specific cost-to-go functions. At each time step, the agents jointly select a previously sampled terminal state and a Generalized Nash equilibrium used to reach it. A new solution is adopted only if it satisfies a cooperative acceptance rule agreed upon by the players. The framework accommodates individual non-degradation requirements and selection criteria such as weighted costs and Nash bargaining. We show that the proposed framework guarantees recursive feasibility and characterize conditions under which each player's cost is non-increasing across iterations.

Artificial intelligence (AI) weather models now forecast weather events skillfully, but how they represent these events internally remains unknown. Sparse autoencoders (SAEs) provide sparse features for probing hidden states, but how information about multiple physical attributes of an event is organized across them remains unclear. Using tropical cyclones in GraphCast as a test case, we analyze 2,115 six-hourly timesteps from 205 cyclones, relating SAE activations to position, intensity, and pressure structure across five Layer-8 decompositions and 16 processor layers. Each attribute relates to multiple features; position and pressure-structure relations overlap in a smaller subset. This organization recurs across independent SAEs and layers without feature alignment. In deeper layers, the leading position feature aligns better with position and pressure structure six hours ahead than at present. These results suggest that the organization of physical-attribute relations provides a reproducible level of description for interpreting cyclone-related information in GraphCast.

Diffusion language models form answers through a sequence of revisable continuous states, but multi-agent systems usually discard those states and make agents communicate only after decoding text. We ask whether one independent diffusion trajectory can pass a compact hidden message into another and, crucially, whether the receiver uses what that message means. In Hidden Messages, a sender compresses its denoising state into four signed dense vectors; the receiver consumes them while retaining its own context, canvas, and trajectory. Removing a message tests whether the channel matters; replacing it with a message from another example tests whether its content matters. Across controlled tasks, derangement removes a +63.6-point advantage on Tiny-A2D and a +100.0-point advantage on Dream-7B, while targeted substitutions steer the receiver toward the substituted fact. On QASC with the two supporting facts split across agents, matched communication reaches 80.8%, but a wrong-example message lowers accuracy to 46.8%, a content effect of +34.0 points. The surprising result is temporal: one late transfer reaches 87.8%, exceeding four recurrent exchanges by +7.0 points. These experiments provide causal evidence for continuous hidden communication between diffusion-LM agents and show that when information is transferred can matter more than how many times agents exchange it.