Both of these studies involve a rodent inflammation model where complete freund's adjuvant is injected to incite inflammation, and then rats are given an alcohol-based coriander seed extract or other injections.
These are pretty low-quality papers in low-quality journals. I'll start with the one you link second:
Deepa, B., Acharya, S., & Holla, R. (2020). Evaluation of antiarthritic activity of Coriander seed essential oil in Wistar albino rats. Research Journal of Pharmacy and Technology, 13(2), 761-766.
This study, which declares "The TNF α levels were statistically reduced by CSEO group compared to Indomethacin group." makes this declaration based on "p<0.05" where the p-value they actually report is 0.055. 0.055 is not less than 0.05, and based on their statistical threshold they should not have rejected the null hypothesis of no effect.
Importantly, they also completely failed their Stats 101 class and instead of comparing the interaction between Time (i.e., baseline vs 21 days) and Treatment, they just compare Baseline to Day 21 in individual groups. This is wrong, and not meaningful. At worst, they should have compared Day 21 among treatment groups, but they did not do this. Their statistical tests do not measure what they say they do.
You can throw this study in the bin, it's worthless as presented, and it's embarrassing to the journal, the researchers, and their institution that it was published.
Nair, V., Singh, S., & Gupta, Y. K. (2012). Evaluation of disease modifying activity of Coriandrum sativum in experimental models. The Indian journal of medical research, 135(2), 240.
The other paper is not quite as bad, but still lacking in statistical methodology. They don't really provide enough information to critique their analysis, which to me is sufficient to consider it junk. You don't make your case better by leaving out details that can be used to critique, we have to assume that those details are hiding something when they are missing. They present standard errors that are miniscule compared to what one would expect in these sorts of experiments, so I have to assume that they may have included pseudoreplication by testing multiple samples from the same animals and not accounting for these multiple comparisons in the statistical analysis.
They've also used a vehicle (control) that is sufficiently different from their treatment that it is just as likely that the non-active ingredients cause the effects they see.
Their IL-6 results are purely qualitative, and even the qualitative approach they use is seriously lacking. Basically "look at these example pictures and trust us".
In summary, these studies do not actually conflict because they do not actually show anything robust - on this they are actually quite in agreement.