Thomas Davidson
@thomasdavidson.bsky.social
Sociologist at Rutgers. Studies far-right politics, populism, and hate speech. Computational social science.
https://www.thomasrdavidson.com/
https://www.thomasrdavidson.com/
Ha thanks, Rohan. It’s a short one but it scratches the surface of what we can do with LRMs. And I haven’t seen any evaluations like this in the CS literature. Let me know if you have any feedback
September 9, 2025 at 10:40 PM
Ha thanks, Rohan. It’s a short one but it scratches the surface of what we can do with LRMs. And I haven’t seen any evaluations like this in the CS literature. Let me know if you have any feedback
Substantively, the results show how reasoning effort and traces could help tasks like content moderation
There are, of course, caveats: LRMs do not replicate human cognition, the models have limited capabilities, and reasoning traces are not always faithful
Preprint: arxiv.org/pdf/2508.20262
There are, of course, caveats: LRMs do not replicate human cognition, the models have limited capabilities, and reasoning traces are not always faithful
Preprint: arxiv.org/pdf/2508.20262
arxiv.org
September 9, 2025 at 4:00 PM
Substantively, the results show how reasoning effort and traces could help tasks like content moderation
There are, of course, caveats: LRMs do not replicate human cognition, the models have limited capabilities, and reasoning traces are not always faithful
Preprint: arxiv.org/pdf/2508.20262
There are, of course, caveats: LRMs do not replicate human cognition, the models have limited capabilities, and reasoning traces are not always faithful
Preprint: arxiv.org/pdf/2508.20262
Analysis of the reasoning traces for Gemini 2.5 shows that the model identifies second-order factors when faced with these decisions, helping to address common false positives like flagging reclaimed slurs as hate speech (Warning: offensive language in example)
September 9, 2025 at 4:00 PM
Analysis of the reasoning traces for Gemini 2.5 shows that the model identifies second-order factors when faced with these decisions, helping to address common false positives like flagging reclaimed slurs as hate speech (Warning: offensive language in example)
On a content moderation task, humans take longer and LRMs use more tokens when offensiveness is identical or fixed.
This suggests that LRM behavior is consistent with dual process theories of cognition, as the models expend more reasoning effort when simple heuristics are insufficient
This suggests that LRM behavior is consistent with dual process theories of cognition, as the models expend more reasoning effort when simple heuristics are insufficient
September 9, 2025 at 4:00 PM
On a content moderation task, humans take longer and LRMs use more tokens when offensiveness is identical or fixed.
This suggests that LRM behavior is consistent with dual process theories of cognition, as the models expend more reasoning effort when simple heuristics are insufficient
This suggests that LRM behavior is consistent with dual process theories of cognition, as the models expend more reasoning effort when simple heuristics are insufficient
The results are consistent across three frontier LRMs: o3, Gemini 2.5 Pro, and Grok 4
September 9, 2025 at 4:00 PM
The results are consistent across three frontier LRMs: o3, Gemini 2.5 Pro, and Grok 4
Contributors include: @sxz.bsky.social @ajalvero.bsky.social @eollion.bsky.social @lpargyle.bsky.social @davidbroska.bsky.social @austin-van-loon.bsky.social @bstewart.bsky.social @hwaight.bsky.social @solmg.bsky.social @tinalaw.bsky.social @lauraknelson.bsky.social @oms279.bsky.social
August 1, 2025 at 3:13 PM