🥇Top ML Papers of the Week
The Top ML Papers of the Week (April 1 - April 7)
1). Many-shot Jailbreaking - proposes a jailbreaking technique called many-shot jailbreaking to evade the safety guardrails of LLMs; this jailbreaking technique exploits the longer context window supported by many modern LLMs; it includes a very large number of faux dialogues (~256) preceding the final question which effectively steers the model to produce harmful responses. (paper | tweet)
2). SWE-Agent - a new open-source agentic system that can automatically solve GitHub issues with similar accuracy as Devin on the SWE-bench; the agent interacts with a specialized terminal and enables important processing of files and executable tests to achieve good performance; on SWE-bench, SWE-agent resolves 12.29% of issues, achieving the state-of-the-art performance on the full test set. (paper | tweet)
3). Mixture-of-Depths - demonstrates that transformer models can learn to efficiently and dynamically allocate FLOPs to specific positions in a sequence; this helps to optimize the allocation along the sequence for different layers across model depth; findings suggest that for a given FLOP budget models can be trained to perform faster and better than their baseline counterparts. (paper | tweet)
4). Local Context LLMs Struggle with Long In-Context Learning - finds that after evaluating 13 long-context LLMs on long in-context learning the LLMs perform relatively well under the token length of 20K. However, after the context window exceeds 20K, most LLMs except GPT-4 will dip dramatically. (paper | tweet)
5). Visualization-of-Thought - inspired by a human cognitive capacity to imagine unseen worlds, this new work proposes Visualization-of-Thought (VoT) prompting to elicit spatial reasoning in LLMs; VoT enables LLMs to "visualize" their reasoning traces, creating internal mental images, that help to guide subsequent reasoning steps; when tested on multi-hop spatial reasoning tasks like visual tiling and visual navigation, VoT outperforms existing multimodal LLMs. (paper | tweet)
Sponsor message
DAIR.AI presents a live cohort-based course, Prompt Engineering for LLMs, that teaches advanced prompting techniques, RAG, tool use in LLMs, and other approaches that improve the capabilities, performance, and reliability of LLMs. Use promo code MAVENAI20 for a 20% discount.
6). The Unreasonable Ineffectiveness of the Deeper Layers - finds that a simple layer-pruning strategy of popular open-weight pretraining LLMs shows minimal performance degradation until after a large fraction (up to half) of the layers are removed; using a layer similarity mechanism optimal blocks are identified and pruned followed by a small amount of fine-tuning to heal damage. (paper | tweet)
7). JetMoE - an 8B model trained with less than $ 0.1 million cost but outperforms LLaMA2-7B; shows that LLM training can be much cheaper than generally thought; JetMoE-8B has 24 blocks where each block has two MoE layers: Mixture of Attention heads (MoA) and Mixture of MLP Experts (MoE); each MoA and MoE layer has 8 experts, and 2 experts are activated for each input token with 2.2B active parameters. (paper | tweet)
8). Representation Finetuning for LMs - proposes a method for representation fine-tuning (ReFT) that operates on a frozen base model and learns task-specific interventions on hidden representations; in other words, by manipulating a small fraction of model representations it is possible to effectively steer model behavior to achieve better downstream performance at inference time; also proposes LoReFT as a drop-in replacement for PEFTs that is 10-50x more parameter efficient. (paper | tweet)
9). Advancing LLM Reasoning - proposes a suite of LLMs (Eurus) optimized for reasoning and achieving SoTA among open-source models on tasks such as mathematics and code generation; Eurus-70B outperforms GPT-3.5 Turbo in reasoning largely due to a newly curated, high-quality alignment dataset designed for complex reasoning tasks; the data includes instructions with preference tree consisting of reasoning chains, multi-turn interactions and pairwise data for preference learning. (paper | tweet)
10). Training LLMs over Neurally Compressed Text - explores training LLMs with neural text compressors; the proposed compression technique segments text into blocks that each compress to the same bit length; the approach improves at scale and outperforms byte-level baselines on both perplexity and inference speed benchmarks; latency is reduced to the shorter sequence length. (paper | tweet)
Reach out to hello@dair.ai if you would like to partner and advertise with us. Our newsletter is read by over 50K AI Researchers, Engineers, and Developers.