Sección

Investigación

Papers y preprints de los feeds de investigación que seguimos.

6315documentos

Matching the Training-Time Weight Rule in Inference-Time Expert Deactivation

Expert deactivation at inference time changes both token routing and expert weights, and its effect on performance depends on how these changes are implemented. We compare removal, freezing, remaining-expert renormalization, and unnormalized re-routing on identical disabled sets across five instruction-tuned checkpoints from four model families, with a base OLMoE checkpoint as an additional comparison. We develop a weight-sum model that relates the loss of expert contributions and changes in branch scale to the weight rule used during training. With half of the experts deactivated according to activation-based rankings, remaining-expert renormalization yields the highest perplexity in models trained with raw router probabilities, whereas unnormalized re-routing yields the highest perplexity in models trained with normalized weights. Perplexity ranges from 1.4 to 73 times the intact baseline when the training rule is retained, and from 14 to 28,740 times the baseline under a rule mismatch. Task outcomes vary across evaluation types: on OLMoE, GSM8K strict-match accuracy falls to zero under all four implementations. Extending the model to changes in the number of active experts yields a further prediction, supported by the OLMoE experiments: the perplexity penalty from additional renormalization increases as fewer experts are activated.

arXiv AI Papershace 9 días

Phonological Knowledge of Chinese Characters in Language Models

How much phonological knowledge do language models (LMs) possess about Chinese characters, and how is this knowledge represented and used? We introduce PinyinBench, a controlled benchmark covering homophones, phonological features (initials, finals, and tones), and polyphonic characters. We investigate 17 text-only and multimodal LMs through three complementary approaches: behavioral evaluation, layer-wise probing, and activation steering. Behaviorally, models demonstrate substantial knowledge of initials and finals, with a frontier model (DS-v4-flash) scoring 90%+ in homophone judgment, when two characters share no phonological features. However, its accuracy drops sharply to 72% when two characters differ only in lexical tone. Probing reveals that tone is highly decodable from embeddings and improves modestly across layers, whereas segmental information starts less decodable and improves substantially with depth. Activation steering shows that segmental representations can systematically influence generation, while tone exhibits much weaker effects. Furthermore, audio-bearing training improves phonological decodability, including tone, without corresponding improvements in steering effectiveness. Together, these findings reveal distinct representational and functional profiles for segmental and tonal information, with tonal information much harder to access and manipulate, demonstrating the necessity for combined evaluation, probing, and causal intervention for understanding linguistic knowledge in LMs.

arXiv AI Papershace 9 días

GREP: Grounding Generated Room Evaluation in Principles

Recent LLM-centric advances can generate attractive furnished rooms by selecting and arranging assets. Yet evaluation remains limited to specific perspectives or predefined questions, failing to reflect overall scene quality. To address this pressing need, we introduce GREP. Its core is a focus shift from defining specific questions to establishing a three-tier evaluation philosophy: basic physical validity checks whether a scene obeys physical constraints; visual coherence concerns compatibility in object scale, style, and composition; and, at the highest level, human-use logic requires object arrangements to align with everyday human habits. Guided by these principles, an evaluation agent autonomously and adaptively identifies issues from scene context, gathers evidence, and produces scores and reports. This supports scene-specific issue discovery and reliable judgments. Experiments show that (1) GREP achieves 87% agreement with human majority preferences, substantially outperforming baseline evaluators. (2) GREP reports improve scene quality through refinement, whereas score-only and no-feedback baselines show slight declines. (3) GREP suggests that as planning ability improves, recent models such as GPT-5.6 Sol can move from relying on pipelines and harnesses to challenging their guidance and constraints, but still lack the ability to resolve the resulting conflicts. Stronger models are therefore expected not only to plan well, but also to adapt to external guidance and constraints more flexibly. Project page: https://han-808.github.io/GREP/

arXiv AI Papershace 9 días

WayPOP: A Panoramic Open-Set Panoptic Tracking Benchmark

Holistic perception for autonomous agents requires reliably segmenting and associating scene elements across time and camera views, including objects outside a fixed semantic taxonomy. Existing open-set panoptic benchmarks focus on individual images or views, while panoramic tracking benchmarks assume closed-set taxonomies. Consequently, these benchmarks do not evaluate whether unknown objects can be consistently segmented and tracked across synchronized multi-camera sequences. We introduce WayPOP, a panoramic open-set panoptic tracking benchmark built on real Waymo PVPS sequences augmented with diverse out-of-taxonomy 3D assets from Objaverse. Inserted objects are placed at fixed locations in the 3D scene and rendered using calibrated vehicle-mounted cameras as the ego vehicle and surrounding traffic move, providing temporally and cross-view consistent panoptic annotations and auxiliary LiDAR range observations. WayPOP contains approximately 670 curated assets instantiated as 980 unique unknown-object tracks, yielding 53,452 unknown-object mask annotations across 21,300 camera images. Our evaluation protocol defines a hierarchy of tasks from view-independent open-set panoptic segmentation to full panoramic open-set panoptic tracking, using known/unknown PQ and (w)STQ metrics. We benchmark representative open-set and open-vocabulary segmentation methods with temporal and cross-camera association. Our results reveal substantial limitations of current methods in truly open-set scenarios, highlighting robust unknown-object localization and identity association across time and views as major open challenges.

arXiv AI Papershace 9 días

Item Representations Matter: Bridging Hard Semantic IDs and Soft Prototype Mixtures for Generative Retrieval

Modern recommender systems rely on generative retrieval techniques to predict the next item either as discrete codes or as a continuous vector. On the one hand, Semantic IDs (SIDs) give discrete models shared parameters and pretrained semantic structure, but discretizing items into hard codes discards some information from the original content embeddings. On the other hand, continuous models avoid quantization, yet typically use flat item lookup tables with no explicit parameter tying across items. We ask whether a continuous recommender can introduce shared structure without making it a discrete bottleneck. We propose CAP-Mix, a flow recommender that represents each item as a soft mixture of shared prototypes plus an item-specific vector. The assignments, prototypes, and item vectors are trained jointly, and the soft representation is retained at inference. Pretrained information is optional: SID responsibilities can serve as a prior over assignments, while a full content embedding can initialize the item-specific vector without changing the initial flat table. In matched comparisons on three datasets, CAP-Mix improves over a flat table on the two sparse Amazon datasets and yields a small gain on Steam when trained without pretrained information. Given only an offline SID, it outperforms a hard table built from the same tokenizer by 4.45-14.90% in HR@5. Given the full content embedding, it matches or improves a strong flat table while hard assignments regress on the sparse Amazon datasets. Overall, averaged across six benchmarks, CAP-Mix improves HR@5 and NDCG@5 by 6.9% and 13.6% relative to the best evaluated baseline for each dataset and metric, while delivering 1.4-161x the inference throughput of the discrete baselines in our experiments.

arXiv AI Papershace 9 días

Consistency Is Not Understanding: Orbit Audits for Neural Field Weight Readers

Recognizing visual content from neural-network weights requires separating functional information from arbitrary parameter coordinates. We audit this separation with hidden-neuron permutations and positive ReLU rescalings that preserve the represented function. The study contains 480 independently fitted shape fields and 1,000 MNIST fields, with transformed copies grouped by origin. On shape recognition, a statistics reader falls from \% to \% accuracy under joint transformations; a normalized set reader retains \% with unchanged predictions. Four small frozen language models obtain \% raw accuracy with next-token scoring. A separate four-configuration reasoning experiment with 1,920 scored sessions obtains \% raw accuracy with generated answers, while fresh identical raw prompts disagree on \% of origins. Canonicalization makes equivalent prompts identical, but this neither establishes recognition nor guarantees identical generated responses. Joint correctness/agreement metrics, semantic-cue controls, fresh repeats, and exact-rational rounding bounds separate these failure modes. The result is an inspectable evaluation protocol for visual weight readers, with claims restricted to the tested fields, prompts, and execution configurations.

arXiv AI Papershace 9 días

DARE: Difficulty Analysis for Repair Enhancement

LLM-based coding agents are primarily evaluated by their solve rates. However, this metric does not reveal how difficult the solved tasks are or account for differences intask difficulty across benchmarks. Determining the difficulty of individual tasks by repeatedly running multiple agents is costly, while a predicted difficulty score alone provides limited guidance on what support an agent needs. To address these challenges, we propose DARE (Difficulty Analysis for Repair Enhancement), which integrates repair-task difficulty measurement, prediction, feature-group contribution analysis, and task-specific repair support. We define continuous task difficulty as the observed failure rate across multiple coding agents. The DARE difficulty model predicts this difficulty using 151 static features extracted from benchmark artifacts. The framework also analyzes the contributions of seven Repair-Task Feature Groups to identify the characteristics most salient to each task and uses this information to select task-specific agent support. We evaluate DARE on four benchmark partitions comprising 2,594 tasks. The difficulty model's aligned predictions achieve Spearman's ρ = .476-.621 on held-out partitions, and the most salient Feature Groups vary across tasks and benchmarks. Furthermore, coding-agent experiments show improved repair success when support is matched to each task's salient Feature Group. These findings demonstrate that difficulty prediction and interpretation can lead to practical improvements in automated program repair (APR), while also providing a basis for difficulty-stratified agent evaluation, cross-benchmark comparisons, and identifying areas of repair capability that warrant further improvement.

arXiv AI Papershace 9 días

ZipDrift: Generating Multimodal Short-Horizon MPC Costs with One-Step Drifting Models

Short-horizon model predictive control (MPC) is fast but myopic. ZipMPC compensates by learning a context-dependent correction of the short-horizon cost that imitates a long-horizon expert. We observe that the optimal correction is often multimodal: when a slower opponent can be passed on either side, good corrections of opposite sign coexist, and the expert's choice between them hinges on tiny cost differences. Deterministic regression then averages modes or inherits inconsistent expert tie-breaking. We propose ZipDrift, which trains a one-step drifting generator over MPC cost corrections, tilted toward low closed-loop cost and learned only from cost evaluations of candidate corrections, without differentiating through the optimiser. A few generated candidates are ranked by a learned critic, and the model-based MPC then enforces dynamics and constraints. In a kinematic overtaking benchmark derived from ZipMPC, 53% of contexts are bimodal and the manual short-horizon MPC crashes in 28.5% of test episodes. With an eight-dimensional stagewise correction, ZipDrift is collision-free in all 600 test episodes and matches the best evaluated candidate, whereas an idealised ZipMPC-style regressor crashes in 4.0% (16.0% with inconsistent expert labels) and critic-ranked random candidates with the same budget in 2.0%.

arXiv AI Papershace 9 días

FlyHash for LLM Agent Memory: An Empirical Validation and a New Application

Large language model (LLM) agents increasingly rely on external memory stores to retain information across conversations, typically via dense vector embeddings searched with approximate nearest-neighbor methods. The fruit fly's olfactory circuit computes a locality-sensitive hash of sensory input using a strikingly different mechanism - sparse random projection followed by winner-take-all competition - and prior work (Dasgupta et al., 2017) showed this scheme, dubbed FlyHash, preserves neighbor structure better per bit than conventional random-hyperplane hashing (SimHash) on several benchmarks. We apply FlyHash, to our knowledge for the first time, to LLM agent memory retrieval, and ask three questions a practitioner actually needs answered: does the quality advantage hold under a correctly matched-bit protocol (it does, robustly, across 10 seeds and four datasets spanning two modalities, with one specific, previously undocumented comparison pitfall identified and corrected along the way); does that advantage translate into real memory and latency savings, including against the dense-vector indexing (HNSW) practitioners default to today (no: at equal bits FlyHash's codes are larger and slower to compare than SimHash's, and HNSW dominates both hash methods at ordinary scale; a sparse-set comparison we introduce recovers most but not all of a naive latency gap, leaving hashing's practical case narrower and more specific than the quality result alone implies); and does the advantage survive genuine, non-circular agent-memory relevance labels rather than the nearest-neighbor-by-embedding proxy this literature otherwise relies on (it does). We report all three answers precisely, including the two that narrow rather than flatter the headline result, and release our implementation, a working agent-memory demo, and a full account of the methodological pitfalls we found and corrected.

arXiv AI Papershace 9 días

Unsupervised Neural Operators for Rotating Relativistic Stars: High-Fidelity Equation of State Inference and Pulsar Spin Evolution at Scale

Observed populations of rotating neutron stars~(NSs) with millisecond-order periods are powerful probes of fundamental physics. Mass-radius measurements inform poorly constrained properties of degenerate nuclear matter, whereas high-fidelity population synthesis connects unknowns in NS formation and binary evolution to large radio-timing surveys. However, repeated structure calculations are prohibitive due to the computational cost of solving the highly coupled partial differential equations of axisymmetric general relativistic hydrostatic equilibrium that govern rotating NS interiors. Non-rotating calculations or approximate rotational corrections can bias equation of state inference and limit the fidelity of population synthesis in tracing the evolutionary history of progenitors and propagating equation of state uncertainties. In this Letter, we develop physics-informed neural operators that can solve the axisymmetric problem with high accuracy and negligible computational cost. Our solvers are trained in a fully unsupervised setting, with the training objective informed solely by the partial differential equations, and achieve a 10^3speed-up with negligible accuracy loss. These solvers enable scalable EoS inference that self-consistently accounts for rotation and repeated stellar-structure calculations during binary evolution, making spin predictions that account for rotational structure and EoS uncertainties computationally accessible to population synthesis.

arXiv AI Papershace 9 días

DreamX: Video Diffusion Model is a Universal Latent Action Learner

Learning a universal action representation from video requires modeling dynamics that are consistent across different executions of the same action yet invariant to changes in embodiment and appearance. Existing latent action models infer actions from visual transitions, but their representations often entangle the underlying action with irrelevant visual variations, limiting transfer across embodiments. We introduce DreamX, which repurposes a pretrained video diffusion model as a universal latent action learner. Our approach extracts compact latent actions from the video model's rich spatiotemporal features, compressing each latent video frame by {300} ({1{,}200 4} tokens). In addition, by leveraging the video model's generative prior, these latent actions capture action-relevant dynamics disentangled from visual appearance. The resulting latent space serves as a shared action interface across embodiments (human and robot), environments, and action parameterizations. On LARY-Bench, DreamX achieves state-of-the-art performance among latent action models. Moreover, actions encoded in one setting remain effective in another: the latent space supports any-to-any video transfer and zero-shot action transfer across diverse embodiments and environments. Furthermore, this shared interface also enables policy learning in a compact latent action space, reducing the computational cost of video-diffusion-based policies.

arXiv AI Papershace 9 días

From Autonomous Agents to Cognitive Symbionts: Toward AI-Augmented Human Intelligence

For a decade, progress in artificial intelligence has been measured by how much a machine can do on its own. Large foundation models now write code and run multistep tasks, and personal agents act on a person's behalf. Yet people still forget what they once knew, miss what they are not looking at, fail to find the right words, and face hard choices without the benefit of their own history, because the machine that could help stands outside the situation. As AI moves into the physical world, attention has settled on one kind of intelligent individual, the embodied robot; we argue that the era of Physical AI holds a second kind of equal importance, the person augmented by AI. Alongside artificial general intelligence, which asks how intelligent a machine can be, stands AI-augmented human intelligence (AHI), which asks how much intelligence a machine can enable in a particular person. Realizing it calls for a new kind of system, a cognitive symbiont: a lifelong, embedded, real-time, private, low-friction, and proactive companion that senses what the person senses through external devices, keeps their context across years, models their situation and their mind, takes part in their thinking at the right moment, carries out everyday tasks with their context, and keeps learning the person as both change. We give the paradigm a conceptual core, the cognitive gap between what a person could bring to a situation and what they can hold in working memory, and organize its design space as a symbiosis loop of sensing, keeping, understanding, living alongside, acting for and thinking with the person, and growing, enclosed by a personal runtime and by trust. Because a symbiont changes what a person can do in front of others, it also raises risks of a new kind, from covert assistance in examinations to pressure on everyone to adopt it, and we treat their governance as part of the paradigm.

arXiv AI Papershace 9 días

Aligning Language Models with General Preferences: An Empirical Study of Direct Nash Optimization

Direct Nash Optimization (DNO) aligns a language model to a general preference function by approximating the Nash equilibrium of a two-player preference game. Its practical variant, DNO-Prct, iterates on-policy sampling, LLM-based rating of the samples together with a teacher's gold response, large-margin preference-pair selection and a DPO-style contrastive update. We re-implement DNO-Prct independently and study it on two open-weight instruct models, Llama 3.1 8B Instruct and Mistral 7B Instruct v0.1, with GPT-4o or a larger open-weight model of the same family as the annotator. On AlpacaEval 2.0 with GPT-4o as judge, three iterations raise the length-controlled (LC) win rate against GPT-4o from 31.7% to 45.3% for Llama and from 6.8% to 18.4% for Mistral, and the tuned 8B model scores above Llama 3.1 70B Instruct on every reported win rate. With same-family open-weight annotators, the runs retain 62% (Llama) and 82% (Mistral) of the LC gain achieved with GPT-4o as annotator, so the rating step does not require a proprietary model. In a separate ablation configuration, variants that remove the gold responses, remove the data look-back, or accept only teacher-preferred pairs score 6 to 10 LC points below the full procedure, whereas varying the number of on-policy samples per prompt (1 to 7) moves the LC win rate by at most 2.1 points, and a margin sweep (1 to 5) run in its own pipeline by at most 2.4 points. We also report the cost of the pipeline and where the practical algorithm departs from the theory it approximates. All results are single runs, and GPT-4o is both the teacher and the evaluation judge; we discuss what this implies for the reported gains.

arXiv AI Papershace 9 días

PolicyIK: High-Precision Inverse Kinematics for Robot Arms with Reinforcement Learning

Inverse kinematics (IK) converts end-effector pose commands into joint motion, and many vision-language-action (VLA) policies rely on it to act on a given arm and to transfer across embodiments. Precise IK remains laborious: redundancy and singularities complicate the inverse mapping, execution depends on hand-tuned controllers, and training data on physical robots are costly to collect. We present PolicyIK, a goal-conditioned reinforcement learning policy for high-precision IK. It updates joint-position references from pose feedback and observes its accumulated reference relative to the measured joints, so it can retain the offset that the position controller needs at rest. Two-scale pose errors with an error-dependent action gain resolve millimeter residuals while preserving large motions. Progress and dwell rewards teach it to both reach the tolerance and stay within it. In simulation, a single attempt by the trained PPO policy keeps the mean errors over the final second within 1 mm and 0.2° on 93.5% of held-out tasks, and a deployment protocol raises success to 98.9%. When trained directly on a physical arm with an off-policy RL learner, PolicyIK achieves higher reaching and holding success than all four classical IK controllers under the same time budgets. In simulated pick-and-place, PolicyIK executes the end-effector commands of the π_{0.5} VLA policy with 95.7% success. These results show that reinforcement learning can provide high-precision inverse kinematics for robot arms. Code is available at https://github.com/Tinaliu0123/PolicyIK.

arXiv AI Papershace 9 días

GUILD: Guided Iterative Data Collection for Imitation Learning

Imitation learning policies in robotics are trained on datasets collected through costly real-world teleoperation. Often, the data-collection process is unguided, relying on the teleoperator's intuition, resulting in data that contributes little to overall policy performance. Recent work has shown that curating such unguided datasets can substantially improve policy performance, but these approaches operate by filtering existing datasets after collection and therefore do not prevent the acquisition of uninformative data. In this work, we propose GUILD, a novel Guided Data Collection (GDC) method that actively reasons about which data should be collected to maximize downstream policy performance. We introduce an iterative framework that alternates between 1) task proposal and data collection and 2) data evaluation to close the loop. For task proposal, we leverage semantic priors from vision-language models to identify relevant environments, object configurations, and goals for data collection. However, since semantic relevance does not imply policy improvement, we perform data evaluation. By leveraging predictive data attribution we approximate how the collected trajectories influence policy performance on a target deployment setting. This signal is then used as feedback to guide subsequent data collection, closing the data collection loop. We evaluate our approach in simulation and real robot settings, demonstrating that guided data collection substantially improves sample efficiency and deployment performance compared to passive data collection strategies, multiple baselines and experienced roboticists. More information at https://robin-lab.cs.utexas.edu/guild2026/

arXiv AI Papershace 9 días

Suppressing Misaligned Behaviors in Language Models, Durably

Undesired model behaviors can emerge or re-emerge at any stage of training. Once observed, how effectively can we suppress them? To study this question, we first construct two model organisms--language models deliberately trained to exhibit a specific misalignment--that differ in what triggers the behavior. The narrow-trigger organism misbehaves only when the prompt contains a specific cue: it cheats on multiple-choice questions by following a planted hint that points to a wrong answer, and behaves normally otherwise. The broad-trigger organism has no such cue: after fine-tuning on insecure code, it gives harmful responses across a wide range of open-ended prompts. We then formalize post-hoc behavioral suppression and evaluate methods along three criteria: immediate suppression of the target behavior, durability under further training, and preservation of general capability. We compare standard methods--SFT, GRPO, and gradient-ascent unlearning--along these dimensions. Building on prior work on alignment shallowness, we hypothesize that more durable suppression requires the ability to recover from a wider range of misalignment states. We introduce Prefix-Expanded Adversarial Reinforcement Learning (PEARL), a GRPO variant that augments rollouts with continuations from intermediate states sampled along cached organism trajectories. Across both organism settings and two model families (Qwen3-4B, GPT-OSS-20B), PEARL achieves strong immediate suppression while preserving task accuracy, and remains durable under further fine-tuning on either the organism's data or unrelated data. It obtains the greatest benefit when the undesirable behavior is broad and cannot be addressed through targeted SFT. Together, our framework and method offer a step toward more durable post-hoc removal of undesired behaviors.

arXiv AI Papershace 9 días

Reading Why: Born-Measured Attention with Exact Word Evidence on Light and Qubits

When a language model answers, we can usually only estimate why. We build a transformer in which every step is a Born measurement - attention |q|k|^4 normalised over each row, a measured product in place of the feed-forward nonlinearity, and a Born readout. With its denominators held at their values for the input, each class probability is a polynomial in the word amplitudes, so its integrated-gradient evidence is computed by a finite quadrature with no quadrature error: the words' contributions sum to the class probability above its no-information share, to 10^{-11} or better. On e-SNLI natural language inference, with about 98k parameters outside the embedding, the model reaches 76.98 0.17% test accuracy, level with a softmax transformer of the same size (76.62 0.36%). Deleting the word its evidence ranks first changes the answer in 45% of test pairs, out of the 53% any single deletion can change, against 32% for the softmax model explained the same way and 40% by its attention x gradient. Attention can look more like the words annotators highlighted, yet accounts for the answer less well: deleting the word it ranks first changes the Born Transformer's answer in 32% of pairs. The same weights compile exactly to optoelectronic optics and to qubits, and a coherent attention layer becomes one circuit equal to the model (852 CNOTs on 23 qubits). On IBM's ibm_kingston, reading only the attention overlaps by Hadamard tests, the device gave the model's answers on two short test pairs and kept or changed them under deletion as the model does; the evidence itself did not yet survive the device, and we work out the shots and calibration that measuring it on hardware would need.

arXiv AI Papershace 9 días

Agnostic Learning by Imitating

Given a class of machine learning models and sample access to a data distribution with binary labels, which computational task is harder: training a discriminative model that distinguishes positive and negative labels as well as the best member of the class, or training a generative model whose outputs are indistinguishable from real data by all members of the class? We prove that these two fundamental tasks in computational learning theory are not just closely related, but actually equivalent to each other in complexity. One direction of this equivalence was previously known. Indeed, the powerful regularity lemma in complexity theory gives an efficient oracle reduction from the indistinguishability task (a.k.a. multiaccuracy) to the distinguishing task (a.k.a. agnostic learning) in the style of boosting algorithms. Our main contribution is an efficient oracle reduction in the other direction. Taking this idea further, we show that a multiaccuracy oracle is a remarkably powerful primitive, achieving not just agnostic learning but also a wide range of seemingly more demanding learning objectives via similarly efficient and direct constructions. Our results crucially use the power of making adaptive queries to the oracle, and stand in stark contrast to prior work, which showed that post-processing a single multiaccurate predictor is insufficient for even weak forms of learning. On a technical level, we draw a connection between our (offline) problem and a version of no-regret learning in the online setting, which allows us to use follow-the-leader type algorithms to modify the label distribution and aggregate the predictions of the multiaccuracy oracle.

arXiv AI Papershace 9 días

Adversarial Speculative Reasoning Improves Long Horizon Agents

Speculative decoding has been treated purely as a systems optimization that speeds up large language model decoding without changing the model behavior. Under this view, the growing effort to train better draft models has been focused solely on acceptance rate. We argue that this view does not hold true for long-horizon agent reasoning. Measuring frontier open-weight models on complex agentic tasks with different speculative decoding approaches, we found that numerical noise grows within verification windows, and the draft acceptance results of the thinking process can change the output distribution of the model after 30 steps, thus noticeably influencing the reasoning accuracy of long-horizon agents in random directions. In this work, we improve the reasoning capability of target models by controlling such noise. While we cannot change the overall token distribution verified by the target model, we can encode critical reasoning with highest precision by training the draft model to be rejected. We identify these points in agent traces and train the draft model to favor adversarial preference. At inference the trained draft and the target part ways precisely at these decision points, verification restarts, so that critical thinking tokens are sampled from the target model and encoded under its least noisy conditions. This approach makes critical reasoning better indexed and more durable over a long horizon. On complex agentic coding benchmarks, our approach raises end-to-end success rate with no change to the target model or its output distribution. Our results recast the speculative decoding mechanism from a pure accelerator into an active allocator for context precision that improves the accuracy of long-horizon agents by training 0.06% to 0.28% of hosted parameters.

arXiv AI Papershace 9 días

A Circuit for In-Context Learning

In-context learning (ICL) lets a pretrained language model perform a task specified by demonstrations in its prompt, with no parameter updates. With arbitrary labels the model must acquire a mapping it cannot have memorized, so the learning happens within a single forward pass. Prior work found candidate parts but not how they compose. We give the composition for few-shot sentence classification with arbitrary labels in three 3-4B-parameter transformers. The model implements a key-value memory over label-token positions: each label token's residual stream encodes the label's identity, and middle-layer attention heads write into it the class of the preceding demonstration as its key. At the query, upper-layer heads match the query's class against the stored keys and copy out the matching labels. An early-layer signal encoding whether any label has repeated gates this readout; in its absence the readers attend to the start of the prompt and no label is favored. A separate set of heads suppresses the most recently seen label. Each step is established causally by activation patching, head ablations against random-head baselines, and steering with the model's own signals. A softmax model of the readout with two constants per head, fit on one prompt distribution, predicts another. The computation is conserved across the three models, the named heads alone reproduce the behaviour, and a transformer trained from scratch grows the same parts. The circuit transfers to semantic concepts and to labels with meaning, and the same parts, with a causal gate, appear from 0.6B to 8B in one family and survive instruction tuning. ICL in this setting is a gated associative lookup built from a few identifiable component types, traced causally from demonstrations to prediction.