Search

We introduce REFA, a family of reference-free alignment methods that optimize over multiple user preferences while enforcing fine-grained length control. Our approach integrates deviation-based weighting to emphasize high-quality responses more strongly, length normalization to prevent trivial short-response solutions, and an EOS-probability regularizer to mitigate dataset-induced brevity biases. Theoretically, we show that under the Uncertainty Reduction with Sequence Length Assertion (URSLA), naive length normalization can still incentivize length-based shortcuts. By contrast, REFA corrects these subtle incentives, guiding models toward genuinely more informative and higher-quality outputs. Empirically, REFA sets a new state-of-the-art among reference-free alignment methods, producing richer responses aligned more closely with human preferences. Compared to a base supervised fine-tuned (SFT) mistral-7b model that achieves 8.4% length-controlled win rate (LC-WR) and 6.2% win rate (WR), our best REFA configuration attains 21.62% LC-WR and 19.87% WR on the AlpacaEval v2 benchmark. This represents a substantial improvement over both the strongest multi-preference baseline, InfoNCA (16.82% LC-WR, 10.44% WR), and the strongest reference-free baseline, SimPO (20.01% LC-WR, 17.65% WR)

Though exponential distributions are ubiquitous in statistical physics and related computational models, directly sampling them from device behavior is rarely done. The superparamagnetic tunnel junction (SMTJ), a key device in probabilistic computing, is known to naturally exhibit exponentially distributed temporal switching dynamics. To sample an exponential distribution with an SMTJ, we need to measure it in the time domain, which is challenging with traditional techniques that focus on sampling the instantaneous state of the device. In this work, we leverage a temporal encoding scheme, where information is encoded in the time at which the device switches between its resistance states. We then develop a circuit element known as a probabilistic delay cell that applies an electrical current step to an SMTJ and a temporal measurement circuit that measures the timing of the first switching event. Repeated experiments confirm that these times are exponentially distributed. Temporal processing methods then allow us to digitally compute with these exponentially distributed probabilistic delay cells. We describe how to use these circuits in a Metropolis-Hastings stepper and in a weighted random sampler, both of which are computationally intensive applications that benefit from the efficient generation of exponentially distributed random numbers.

We introduce Simultaneous Weighted Preference Optimization (SWEPO), a novel extension of Direct Preference Optimization (DPO) designed to accommodate multiple dynamically chosen positive and negative responses for each query. SWEPO employs a weighted group contrastive loss, assigning weights to responses based on their deviation from the mean reward score. This approach effectively prioritizes responses that are significantly better or worse than the average, enhancing optimization. Our theoretical analysis demonstrates that simultaneously considering multiple preferences reduces alignment bias, resulting in more robust alignment. Additionally, we provide insights into the training dynamics of our loss function and a related function, InfoNCA. Empirical validation on the UltraFeedback dataset establishes SWEPO as state-of-the-art, with superior performance in downstream evaluations using the AlpacaEval dataset.

Large Language Models (LLMs) are typically trained to predict in the forward direction of time. However, recent works have shown that prompting these models to look back and critique their own generations can produce useful feedback. Motivated by this, we explore the question of whether LLMs can be empowered to think (predict and score) backwards to provide unsupervised feedback that complements forward LLMs. Towards this, we introduce Time Reversed Language Models (TRLMs), which can score and generate queries when conditioned on responses, effectively functioning in the reverse direction of time. Further, to effectively infer in the response to query direction, we pre-train and fine-tune a language model (TRLM-Ba) in the reverse token order from scratch. We show empirically (and theoretically in a stylized setting) that time-reversed models can indeed complement forward model predictions when used to score the query given response for re-ranking multiple forward generations. We obtain up to 5\% improvement on the widely used AlpacaEval Leaderboard over the competent baseline of best-of-N re-ranking using self log-perplexity scores. We further show that TRLM scoring outperforms conventional forward scoring of response given query, resulting in significant gains in applications such as citation generation and passage retrieval. We next leverage the generative ability of TRLM to augment or provide unsupervised feedback to input safety filters of LLMs, demonstrating a drastic reduction in false negative rate with negligible impact on false positive rates against several attacks published on the popular JailbreakBench leaderboard., Comment: Accepted as a spotlight in NeurIPS 2024

We show that ultrafast optical pump pulses provide effective control over charge orders in the kagome metals $A$V$_3$Sb$_5$ with $A=$ K, Rb, and Cs. Starting from the real charge density waves (rCDWs) at the $p$-type Van Hove singularity, we conduct a thorough analysis of the post-pump dynamics by time-dependent Hartree-Fock theory. Our analysis uncovers distinct dynamical phenomena under linearly and circularly polarized pumps. Linearly polarized pumps induce directional preferences in the rCDWs, accompanied by an enhancement in the flat band. Unexpectedly, charge nematicity also emerges and receives maximal enhancement at a resonant pump frequency, which we understand with a Rabi-oscillation-like model. On the other hand, circularly polarized pumps suppress the rCDWs uniformly and triggers imaginary CDWs (iCDWs) with charge loop currents. Our results can be directly compared to the pump-probe experiments on the kagome metals $A$V$_3$Sb$_5$., Comment: 7+19 pages, 6+7 figures

Video detailed captioning is a key task which aims to generate comprehensive and coherent textual descriptions of video content, benefiting both video understanding and generation. In this paper, we propose AuroraCap, a video captioner based on a large multimodal model. We follow the simplest architecture design without additional parameters for temporal modeling. To address the overhead caused by lengthy video sequences, we implement the token merging strategy, reducing the number of input visual tokens. Surprisingly, we found that this strategy results in little performance loss. AuroraCap shows superior performance on various video and image captioning benchmarks, for example, obtaining a CIDEr of 88.9 on Flickr30k, beating GPT-4V (55.3) and Gemini-1.5 Pro (82.2). However, existing video caption benchmarks only include simple descriptions, consisting of a few dozen words, which limits research in this field. Therefore, we develop VDC, a video detailed captioning benchmark with over one thousand carefully annotated structured captions. In addition, we propose a new LLM-assisted metric VDCscore for bettering evaluation, which adopts a divide-and-conquer strategy to transform long caption evaluation into multiple short question-answer pairs. With the help of human Elo ranking, our experiments show that this benchmark better correlates with human judgments of video detailed captioning quality., Comment: Code, docs, weight, benchmark and training data are all avaliable at \href{https://rese1f.github.io/aurora-web/}{website}

Tactile sensing is a powerful means of implicit communication between a human and a robot assistant. In this paper, we investigate how tactile sensing can transcend cross-embodiment differences across robotic systems in the context of collaborative manipulation. Consider tasks such as collaborative object carrying where the human-robot interaction is force rich. Learning and executing such skills requires the robot to comply to the human and to learn behaviors at the joint-torque level. However, most robots do not offer this compliance or provide access to their joint torques. To address this challenge, we present an approach that uses tactile sensors to transfer policies from robots with these capabilities to those without. We show how our method can enable a cooperative task where a robot and human must work together to maneuver objects through space. We first demonstrate the skill on an impedance control-capable robot equipped with tactile sensing, then show the positive transfer of the tactile policy to a planar prismatic robot that is only capable of position control and does not come equipped with any sort of force/torque feedback, yet is able to comply to the human motions only using tactile feedback. Further details and videos can be found on our project website at https://www.mmintlab.com/research/tactile-collaborative/., Comment: 6 pages + references, 4 figures, 4 tables, submitted to ICRA 2025

Creating secure and resilient applications with large language models (LLM) requires anticipating, adjusting to, and countering unforeseen threats. Red-teaming has emerged as a critical technique for identifying vulnerabilities in real-world LLM implementations. This paper presents a detailed threat model and provides a systematization of knowledge (SoK) of red-teaming attacks on LLMs. We develop a taxonomy of attacks based on the stages of the LLM development and deployment process and extract various insights from previous research. In addition, we compile methods for defense and practical red-teaming strategies for practitioners. By delineating prominent attack motifs and shedding light on various entry points, this paper provides a framework for improving the security and robustness of LLM-based systems., Comment: Preprint. Under review

In this work, we describe a novel approach to building a neural PDE solver leveraging recent advances in transformer based neural network architectures. Our model can provide solutions for different values of PDE parameters without any need for retraining the network. The training is carried out in a self-supervised manner, similar to pretraining approaches applied in language and vision tasks. We hypothesize that the model is in effect learning a family of operators (for multiple parameters) mapping the initial condition to the solution of the PDE at any future time step t. We compare this approach with the Fourier Neural Operator (FNO), and demonstrate that it can generalize over the space of PDE parameters, despite having a higher prediction error for individual parameter values compared to the FNO. We show that performance on a specific parameter can be improved by finetuning the model with very small amounts of data. We also demonstrate that the model scales with data as well as model size.

The long-sought pair density wave (PDW) is an exotic phase of matter in which charge density wave (CDW) order is intertwined with the amplitude or phase of coexisting, superconducting order \cite{Berg2009,Berg2009b}. Originally predicted to exist in copper-oxides, circumstantial evidence for PDW order now exists in a variety of materials. Recently, scanning tunneling microscopy (STM) studies have reported evidence for a three-component charge density wave (CDW) at the surface of the heavy-fermion superconductor, UTe$_2$, persisting below its superconducting transition temperature. Here, we use hard x-ray diffraction measurements on crystals of UTe$_2$ at $T = 1.9$ K and $12$ K to search for a bulk signature of this CDW. Using STM measurements as a constraint, we calculate the expected locations of CDW superlattice peaks, and sweep a large volume of reciprocal space in search of a signature. We failed to find any evidence for a CDW near any of the expected superlattice positions in many Brillouin zones. We estimate an upper bound on the CDW lattice distortion of $u_{max} \lesssim 4 \times 10^{-3} \mathrm{\AA}$. Our results suggest that the CDW observed in STM is either purely electronic, somehow lacking a signature in the structural lattice, or is restricted to the material surface.

We consider spherically symmetric gravity coupled to a spherically symmetric scalar field with a specific coupling which depends on the Areal Radius. Classical collapse is described by the Vaidya solution. The semiclassical Einstein equations are well defined. Their solution describes an evaporating black hole. A balance law at future null infinity relates the rate of change of a back reaction-corrected Bondi mass to a manifestly positive flux. Its detailed form suggests a quantum extension of future null infinity along which all information including that in the collapsing matter is expected to emerge., Comment: 27 pages, 3 figures, typos corrected,references added

A near memory hardware accelerator, based on a novel direct path computational model, for real-time emulation of radio frequency systems is demonstrated. Our evaluation of hardware performance uses both application-specific integrated circuits (ASIC) and field programmable gate arrays (FPGA) methodologies: 1). The ASIC testchip implementation, using TSMC 28nm CMOS, leverages distributed autonomous control to extract concurrency in compute as well as low latency. It achieves a $518$ MHz per channel bandwidth in a prototype $4$-node system. The maximum emulation range supported in this paradigm is $9.5$ km with $0.24$ $\mu$s of per-sample emulation latency. 2). The FPGA-based implementation, evaluated on a Xilinx ZCU104 board, demonstrates a $9$-node test case (two Transmitters, one Receiver, and $6$ passive reflectors) with an emulation range of $1.13$ km to $27.3$ km at $215$ MHz bandwidth.

In this paper we consider the problem of developing a computational model for emulating an RF channel. The motivation for this is that an accurate and scalable emulator has the potential to minimize the need for field testing, which is expensive, slow, and difficult to replicate. Traditionally, emulators are built using a tapped delay line model where long filters modeling the physical interactions of objects are implemented directly. For an emulation scenario consisting of $M$ objects all interacting with one another, the tapped delay line model's computational requirements scale as $O(M^3)$ per sample: there are $O(M^2)$ channels, each with $O(M)$ complexity. In this paper, we develop a new ``direct path" model that, while remaining physically faithful, allows us to carefully factor the emulator operations, resulting in an $O(M^2)$ per sample scaling of the computational requirements. The impact of this is drastic, a $200$ object scenario sees about a $100\times$ reduction in the number of per sample computations. Furthermore, the direct path model gives us a natural way to distribute the computations for an emulation: each object is mapped to a computational node, and these nodes are networked in a fully connected communication graph. Alongside a discussion of the model and the physical phenomena it emulates, we show how to efficiently parameterize antenna responses and scattering profiles within this direct path framework. To verify the model and demonstrate its viability in hardware, we provide several numerical experiments produced using a cycle level C++ simulator of a hardware implementation of the model.

The integration of language and 3D perception is crucial for developing embodied agents and robots that comprehend and interact with the physical world. While large language models (LLMs) have demonstrated impressive language understanding and generation capabilities, their adaptation to 3D environments (3D-LLMs) remains in its early stages. A primary challenge is the absence of large-scale datasets that provide dense grounding between language and 3D scenes. In this paper, we introduce 3D-GRAND, a pioneering large-scale dataset comprising 40,087 household scenes paired with 6.2 million densely-grounded scene-language instructions. Our results show that instruction tuning with 3D-GRAND significantly enhances grounding capabilities and reduces hallucinations in 3D-LLMs. As part of our contributions, we propose a comprehensive benchmark 3D-POPE to systematically evaluate hallucination in 3D-LLMs, enabling fair comparisons among future models. Our experiments highlight a scaling effect between dataset size and 3D-LLM performance, emphasizing the critical role of large-scale 3D-text datasets in advancing embodied AI research. Notably, our results demonstrate early signals for effective sim-to-real transfer, indicating that models trained on large synthetic data can perform well on real-world 3D scans. Through 3D-GRAND and 3D-POPE, we aim to equip the embodied AI community with essential resources and insights, setting the stage for more reliable and better-grounded 3D-LLMs. Project website: https://3d-grand.github.io, Comment: Project website: https://3d-grand.github.io

We study a variant of causal contextual bandits where the context is chosen based on an initial intervention chosen by the learner. At the beginning of each round, the learner selects an initial action, depending on which a stochastic context is revealed by the environment. Following this, the learner then selects a final action and receives a reward. Given $T$ rounds of interactions with the environment, the objective of the learner is to learn a policy (of selecting the initial and the final action) with maximum expected reward. In this paper we study the specific situation where every action corresponds to intervening on a node in some known causal graph. We extend prior work from the deterministic context setting to obtain simple regret minimization guarantees. This is achieved through an instance-dependent causal parameter, $\lambda$, which characterizes our upper bound. Furthermore, we prove that our simple regret is essentially tight for a large class of instances. A key feature of our work is that we use convex optimization to address the bandit exploration problem. We also conduct experiments to validate our theoretical results, and release our code at our project GitHub repository: https://github.com/adaptiveContextualCausalBandits/aCCB., Comment: Reinforcement Learning Conference (RLC) 2024, 10 pages (31 pages including appendix), 8 plots. arXiv admin note: text overlap with arXiv:2111.00886

Manipulating solids using the time-periodic drive of a laser pulse is a promising route to generate new phases of matter. Whether such `Floquet-Bloch' manipulation can be achieved in topological magnetic systems with disorder has so far been unclear. In this work, we realize Floquet-Bloch manipulation of the Dirac surface-state mass of the topological antiferromagnet (AFM) MnBi$_2$Te$_4$. Using time- and angle-resolved photoemission spectroscopy (tr-ARPES), we show that opposite helicities of mid-infrared circularly polarized light result in substantially different Dirac mass gaps in the AFM phase, despite the equilibrium Dirac cone being massless. We explain our findings in terms of a Dirac fermion with a random mass. Our results underscore Floquet-Bloch manipulation as a powerful tool for controlling topology even in the presence of disorder, and for uncovering properties of materials that may elude conventional probes.

Artificial neural networks have advanced due to scaling dimensions, but conventional computing faces inefficiency due to the von Neumann bottleneck. In-memory computation architectures, like memristors, offer promise but face challenges due to hardware non-idealities. This work proposes and experimentally demonstrates layer ensemble averaging, a technique to map pre-trained neural network solutions from software to defective hardware crossbars of emerging memory devices and reliably attain near-software performance on inference. The approach is investigated using a custom 20,000-device hardware prototyping platform on a continual learning problem where a network must learn new tasks without catastrophically forgetting previously learned information. Results demonstrate that by trading off the number of devices required for layer mapping, layer ensemble averaging can reliably boost defective memristive network performance up to the software baseline. For the investigated problem, the average multi-task classification accuracy improves from 61 % to 72 % (< 1 % of software baseline) using the proposed approach.

The classical dances from India utilize a set of hand gestures known as Mudras, serving as the foundational elements of its posture vocabulary. Identifying these mudras represents a primary task in digitizing the dance performances. With Kathakali, a dance-drama, as the focus, this work addresses mudra recognition by framing it as a 24-class classification problem and proposes a novel vector-similarity-based approach leveraging pose estimation techniques. This method obviates the need for extensive training or fine-tuning, thus mitigating the issue of limited data availability common in similar AI applications. Achieving an accuracy rate of 92%, our approach demonstrates comparable or superior performance to existing model-training-based methodologies in this domain. Notably, it remains effective even with small datasets comprising just 1 or 5 samples, albeit with a slightly diminished performance. Furthermore, our system supports processing images, videos, and real-time streams, accommodating both hand-cropped and full-body images. As part of this research, we have curated and released a publicly accessible Hasta Mudra dataset, which applies to multiple South Indian art forms including Kathakali. The implementation of the proposed method is also made available as a web application.

Superparamagnetic tunnel junctions are important devices for a range of emerging technologies, but most existing compact models capture only their mean switching rates. Capturing qualitatively accurate analog dynamics of these devices will be important as the technology scales up. Here we present results using a one-dimensional overdamped Langevin equation that captures statistical properties of measured time traces, including voltage histograms, drift and diffusion characteristics as measured with Kramers-Moyal coefficients, and dwell times distributions. While common macrospin models are more physically-motivated magnetic models than the Langevin model, we show that for the device measured here, they capture even fewer of the measured experimental behaviors., Comment: 22 pages (13 pages of main text), 21 figures

We demonstrate the miniaturization of perpendicularly magnetized superparamagnetic tunnel junctions (SMTJs) down to 50 nm in diameter. We experimentally show stochastic reversals in those junctions, with tunable mean dwell times down to a few nanoseconds through applied magnetic field and voltage. The mean dwell times measured at negligible bias voltage agree with our simulations based on Langer's theory. We shed light on an Arrhenius prefactor $\tau_0$ of a few femtoseconds, implying that the rates of thermally-activated magnetic transitions exceed the GHz-to-THz limitation of macrospin models, whereby $\tau_0\sim1$ ns. We explain the small prefactor values by a Meyer-Neldel compensation phenomenon, where the prefactor exhibits a large entropic contribution with an exponential dependence on the activation energy. These findings pave the way towards the development of ultrafast, low-power unconventional computing schemes operating by leveraging thermal noise in perpendicular SMTJs, which are scalable below 20 nm., Comment: 7 pages, 3 figures

Researchers, government bodies, and organizations have been repeatedly calling for a shift in the responsible AI community from general principles to tangible and operationalizable practices in mitigating the potential sociotechnical harms of AI. Frameworks like the NIST AI RMF embody an emerging consensus on recommended practices in operationalizing sociotechnical harm mitigation. However, private sector organizations currently lag far behind this emerging consensus. Implementation is sporadic and selective at best. At worst, it is ineffective and can risk serving as a misleading veneer of trustworthy processes, providing an appearance of legitimacy to substantively harmful practices. In this paper, we provide a foundation for a framework for evaluating where organizations sit relative to the emerging consensus on sociotechnical harm mitigation best practices: a flexible maturity model based on the NIST AI RMF.

Superparamagnetic tunnel junctions (SMTJs) are promising sources of randomness for compact and energy efficient implementations of probabilistic computing techniques. Augmenting an SMTJ with electronic circuits, to convert the random telegraph fluctuations of its resistance state to stochastic digital signals, gives a basic building block known as a probabilistic bit or $p$-bit. Though scalable probabilistic computing methods connecting $p$-bits have been proposed, practical implementations are limited by either minimal tunability or energy inefficient microprocessors-in-the-loop. In this work, we experimentally demonstrate the functionality of a scalable analog unit cell, namely a pair of $p$-bits with programmable electrical coupling. This tunable coupling is implemented with operational amplifier circuits that have a time constant of approximately 1us, which is faster than the mean dwell times of the SMTJs over most of the operating range. Programmability enables flexibility, allowing both positive and negative couplings, as well as coupling devices with widely varying device properties. These tunable coupling circuits can achieve the whole range of correlations from $-1$ to $1$, for both devices with similar timescales, and devices whose time scales vary by an order of magnitude. This range of correlation allows such circuits to be used for scalable implementations of simulated annealing with probabilistic computing.