Almost Surely Safe Alignment of Large Language Models at Inference-Time
Paper
•
2502.01208
•
Published
•
11
Haitham Bou Ammar
hba123
·
AI & ML interests
LLMs, VLMs, Robotics, Reinforcement Learning, Bayesian Optimisation
Recent Activity
authored
a paper
7 days ago
Almost Surely Safe Alignment of Large Language Models at Inference-Time
reacted
to
their
post
with 😎
8 days ago
We developed a method that ensures almost-sure safety (i.e., safety with probability approaching 1). We proved this result. We then, present a practical implementation which we call InferenceGuard. InferenceGuard has impressive practical results: 91.04% on Alpaca-7B and 100% safety results on Beaver 7B-v3.
Now, it is easy to get high safety results like those if we want a dumb model, e.g., just don't answer or answer with EOS and so on. However, our goal is not to only have safe results, but also to make sure that the rewards are high - we want a good trade-off between safety and rewards! That's exactly, what we show. InferenceGuard achieves that!
Check it out: https://huggingface.co/papers/2502.01208
reacted
to
their
post
with 🔥
8 days ago
We developed a method that ensures almost-sure safety (i.e., safety with probability approaching 1). We proved this result. We then, present a practical implementation which we call InferenceGuard. InferenceGuard has impressive practical results: 91.04% on Alpaca-7B and 100% safety results on Beaver 7B-v3.
Now, it is easy to get high safety results like those if we want a dumb model, e.g., just don't answer or answer with EOS and so on. However, our goal is not to only have safe results, but also to make sure that the rewards are high - we want a good trade-off between safety and rewards! That's exactly, what we show. InferenceGuard achieves that!
Check it out: https://huggingface.co/papers/2502.01208
Organizations
None yet
hba123's activity
authored
a
paper
7 days ago
reacted to
their
post with 😎🔥
8 days ago
Post
1708
We developed a method that ensures almost-sure safety (i.e., safety with probability approaching 1). We proved this result. We then, present a practical implementation which we call InferenceGuard. InferenceGuard has impressive practical results: 91.04% on Alpaca-7B and 100% safety results on Beaver 7B-v3.
Now, it is easy to get high safety results like those if we want a dumb model, e.g., just don't answer or answer with EOS and so on. However, our goal is not to only have safe results, but also to make sure that the rewards are high - we want a good trade-off between safety and rewards! That's exactly, what we show. InferenceGuard achieves that!
Check it out: Almost Surely Safe Alignment of Large Language Models at Inference-Time (2502.01208)
Now, it is easy to get high safety results like those if we want a dumb model, e.g., just don't answer or answer with EOS and so on. However, our goal is not to only have safe results, but also to make sure that the rewards are high - we want a good trade-off between safety and rewards! That's exactly, what we show. InferenceGuard achieves that!
Check it out: Almost Surely Safe Alignment of Large Language Models at Inference-Time (2502.01208)
posted
an
update
8 days ago
Post
1708
We developed a method that ensures almost-sure safety (i.e., safety with probability approaching 1). We proved this result. We then, present a practical implementation which we call InferenceGuard. InferenceGuard has impressive practical results: 91.04% on Alpaca-7B and 100% safety results on Beaver 7B-v3.
Now, it is easy to get high safety results like those if we want a dumb model, e.g., just don't answer or answer with EOS and so on. However, our goal is not to only have safe results, but also to make sure that the rewards are high - we want a good trade-off between safety and rewards! That's exactly, what we show. InferenceGuard achieves that!
Check it out: Almost Surely Safe Alignment of Large Language Models at Inference-Time (2502.01208)
Now, it is easy to get high safety results like those if we want a dumb model, e.g., just don't answer or answer with EOS and so on. However, our goal is not to only have safe results, but also to make sure that the rewards are high - we want a good trade-off between safety and rewards! That's exactly, what we show. InferenceGuard achieves that!
Check it out: Almost Surely Safe Alignment of Large Language Models at Inference-Time (2502.01208)
upvoted
a
paper
8 days ago
commented
a
paper
8 days ago
reacted to
their
post with 🚀
about 1 month ago
Post
1830
I have some New Year presents for you, #MachineLearning and #AI community! We just opened our code for new state-of-the-art results that beat EAGLE-2 and Medusa #LLM inference.
We also shared the model check pt on @huggingface ! @MatthieuZ
Check the blog out: https://huggingface.co/blog/hba123/sotaspeculativedecoding
We also shared the model check pt on @huggingface ! @MatthieuZ
Check the blog out: https://huggingface.co/blog/hba123/sotaspeculativedecoding
upvoted
an
article
about 1 month ago
Article
Accelerating Language Model Inference with Mixture of Attentions
By
and 1 other
•
•
24
posted
an
update
about 1 month ago
Post
1830
I have some New Year presents for you, #MachineLearning and #AI community! We just opened our code for new state-of-the-art results that beat EAGLE-2 and Medusa #LLM inference.
We also shared the model check pt on @huggingface ! @MatthieuZ
Check the blog out: https://huggingface.co/blog/hba123/sotaspeculativedecoding
We also shared the model check pt on @huggingface ! @MatthieuZ
Check the blog out: https://huggingface.co/blog/hba123/sotaspeculativedecoding
published
an
article
about 1 month ago
Article
Accelerating Language Model Inference with Mixture of Attentions
By
and 1 other
•
•
24
liked
a
model
about 1 month ago
authored
2
papers
about 1 month ago
reacted to
their
post with 🚀
about 2 months ago
Post
1816
Blindly applying algorithms without understanding the math behind them is not a good idea frmpv. So, I am on a quest to fix this!
I wrote my first hugging face article on how you would derive closed-form solutions for KL-regularised reinforcement learning problems - what is used for DPO.
Check it out: https://huggingface.co/blog/hba123/derivingdpo
I wrote my first hugging face article on how you would derive closed-form solutions for KL-regularised reinforcement learning problems - what is used for DPO.
Check it out: https://huggingface.co/blog/hba123/derivingdpo
posted
an
update
about 2 months ago
Post
1816
Blindly applying algorithms without understanding the math behind them is not a good idea frmpv. So, I am on a quest to fix this!
I wrote my first hugging face article on how you would derive closed-form solutions for KL-regularised reinforcement learning problems - what is used for DPO.
Check it out: https://huggingface.co/blog/hba123/derivingdpo
I wrote my first hugging face article on how you would derive closed-form solutions for KL-regularised reinforcement learning problems - what is used for DPO.
Check it out: https://huggingface.co/blog/hba123/derivingdpo
commented
a
paper
about 2 months ago
upvoted
an
article
about 2 months ago
Article
Deriving DPO's Loss
By
•
•
26
commented
a
paper
2 months ago
upvoted
a
paper
3 months ago
commented
a
paper
3 months ago