arXiv subjects

Neel Rajani

Publications and source records attributed to Neel Rajani.

3 recordsLinked to original sources

arXiv2026

Chain-of-Thought Faithfulness of Reasoning Models Varies with Where and How Preference Cues Are Delivered

Chain-of-thought (CoT) monitoring assumes that reasoning traces faithfully record the information that shapes a model's answer. Existing faithfulness tests often place explicit bias cues in the user message, while agents may encounter preferences through tool returns or raw artifacts. We introduce FACE-Eval (Faithful Attribution of Cue Effects Evaluation), a 5,100-sample evaluation that varies cue location (user message or tool return) and explicitness (direct summary or raw artifact). We measure verbalized commitment among cue-following answers and unverbalized adoption among all cued samples. We evaluate 15 open-weight models from eight families, with total parameters ranging from 4B to 1.60T. Every model has lower verbalized commitment for tool-return than user-message cues and for implicit than explicit cues. Unverbalized adoption is higher for tool-return cues on all 15 models and for implicit cues in 28 of 30 model-channel comparisons. A source-attribution prompt narrows the channel gap on seven models, sometimes by increasing user-channel unverbalized adoption, while telling models that their reasoning will be monitored does not reliably close the gap. We also use two transcript monitors (GPT-5.6-Luna and GPT-4o-mini) to detect preference adoption in the largest model of each family. Across 32 model-channel-explicitness cells, higher unverbalized adoption is associated with lower detection ability for both monitors (Pearson r=-0.54 and r=-0.78, respectively). These results suggest that CoT monitoring may be less reliable when preference information arrives through tools or must be inferred from raw artifacts, within the single-call, prefilled-tool setting tested here.

cs.CL↗

arXiv2025

Scalpel vs. Hammer: GRPO Amplifies Existing Capabilities, SFT Replaces Them

Training large language models (LLMs) for reasoning via maths and code datasets has become a major new focus in LLM post-training. Two particularly popular approaches are reinforcement learning (RL) and supervised fine-tuning (SFT), but their training dynamics are poorly understood. We present a comparative analysis of RL and SFT on the same maths problems with the same model and similar hyperparameters. We find that RL yields minor in-domain gains on maths and slight degradation on knowledge-intensive benchmarks like MMLU, while both trends are more pronounced in SFT. We also analyse model parameters across checkpoints, observing that both algorithms modify query and key weights the most. Meanwhile, SFT exhibits greater updates and also affects mid-layer MLPs more, leading us to hypothesise that this may have caused the out-of-domain degradation. We therefore investigate whether freezing parts of the model during training can mitigate the reduced performance on knowledge-intensive benchmarks. However, our results are inconclusive, with benefits on GPQA:Diamond and degradation on other benchmarks. Taken together, our observations provide a preliminary indication for why RL amplifies existing capabilities, while SFT replaces old skills with new ones.

cs.LG↗

arXiv2024

KodeXv0.1: A Family of State-of-the-Art Financial Large Language Models

Although powerful, current cutting-edge LLMs may not fulfil the needs of highly specialised sectors. We introduce KodeXv0.1, a family of large language models that outclass GPT-4 in financial question answering. We utilise the base variants of Llama 3.1 8B and 70B and adapt them to the financial domain through a custom training regime. To this end, we collect and process a large number of publicly available financial documents such as earnings calls and business reports. These are used to generate a high-quality, synthetic dataset consisting of Context-Question-Answer triplets which closely mirror real-world financial tasks. Using the train split of this dataset, we perform RAG-aware 4bit LoRA instruction tuning runs of Llama 3.1 base variants to produce KodeX-8Bv0.1 and KodeX-70Bv0.1. We then complete extensive model evaluations using FinanceBench, FinQABench and the withheld test split of our dataset. Our results show that KodeX-8Bv0.1 is more reliable in financial contexts than cutting-edge instruct models in the same parameter regime, surpassing them by up to 9.24%. In addition, it is even capable of outperforming state-of-the-art proprietary models such as GPT-4 by up to 7.07%. KodeX-70Bv0.1 represents a further improvement upon this, exceeding GPT-4's performance on every tested benchmark.

cs.CL↗